You are on page 1of 103

Singapore Management University

Institutional Knowledge at Singapore Management University
Dissertations and Teses Collection (Open Access) Dissertations and Teses
2011
Analysis of Singapore's Foreign Exchange Market
Microstructure
Chee Wai Wan
Singapore Management University, cheewai.wan.2007@smu.edu.sg
Tis Master Tesis is brought to you for free and open access by the Dissertations and Teses at Institutional Knowledge at Singapore Management
University. It has been accepted for inclusion in Dissertations and Teses Collection (Open Access) by an authorized administrator of Institutional
Knowledge at Singapore Management University. For more information, please email libIR@smu.edu.sg.
Recommended Citation
Wan, Chee Wai, "Analysis of Singapore's Foreign Exchange Market Microstructure" (2011). Dissertations and Teses Collection (Open
Access). Paper 75.
htp://ink.library.smu.edu.sg/etd_coll/75




ANALYSIS OF
SINGAPORE’S FOREIGN EXCHANGE
MARKET MICROSTRUCTURE





WAN CHEE WAI






SINGAPORE MANAGEMENT UNIVERSITY
2011
Copyright (2011) Wan Chee Wai

Analysis of
Singapore’s Foreign Exchange
Market Microstructure


by
Wan Chee Wai



Submitted to the School of Economics in partial fulfillment of the
requirements for the Degree of Master of Science in Economics



Thesis Committee:


Tse Yiu Kuen (Supervisor/Chair)
Professor of Economics
Singapore Management University

Hoon Hian Teck
Professor of Economics
Singapore Management University

Anthony Tay
Associate Professor of Economics
Singapore Management University



Singapore Management University
2011


Analysis of
Singapore’s Foreign Exchange
Market Microstructure

Wan Chee Wai

Abstract

This paper analyses the Singapore foreign exchange market from a
microstructure approach. Specifically, by applying and modifying the empirical
methodology designed by Bollerslev and Melvin (1994), we examine the
relationship between bid-ask spreads and the underlying volatility of the
USD/SGD. Our data set comprises high-frequency USD/SGD tick data of three
separate years (April-June 1989, April-May 2006, April-May 2009). We found
that for the USD/SGD: i) the size of bid-ask spreads are positively related to the
underlying exchange rate volatility; ii) the magnitude of the dependence on
underlying volatility increases as tick volume increases; and iii) the size of the
bid-ask spreads may also be positively related to the directional movement of
exchange rates.
i
Table of Contents
Table of Contents ......................................................................................................i
Acknowledgement ..................................................................................................iv
1 Introduction....................................................................................................... 1
1.1 The Singapore Foreign Exchange Market ........................................................................ 2
1.2 The Microstructure Approach to Exchange Rate Economics ........................................... 5
1.3 Organization of Thesis...................................................................................................... 9
2 Bid-Ask Spreads in Exchange Rates............................................................... 10
2.1 Bid-Ask Spreads and Asymmetric Information.............................................................. 11
2.2 Bid-Ask Spreads and Volatility....................................................................................... 18
2.3 Bollerslev & Melvin’s Model of Volatility and Bid-Ask Spread.................................... 21
3 Empirical Analysis.......................................................................................... 27
3.1 Empirical Methodology.................................................................................................. 28
3.2 Description of the Data................................................................................................... 34
3.3 GARCH Analysis ........................................................................................................... 44
3.4 Ordered Response Analysis ............................................................................................ 47
3.5 Incorporating Returns ..................................................................................................... 50
3.6 Summary......................................................................................................................... 53
4 Conclusion ...................................................................................................... 55
References.............................................................................................................. 57
Appendix 1 – EViews Results................................................................................ 60
A.1.1 Dataset 1: USD-DEM April 1989 to June 1989: ........................................................ 60
A.1.1.1 MA(1)-GARCH(1,1) Results: .................................................................................... 60
A.1.1.2 Ordered Probit Results: 4 Ordered Values.................................................................. 61
A.1.1.3 Ordered Probit (with Returns) Results: 4 Ordered Values.......................................... 62
A.1.1.4 Ordered Probit Results: 10 Ordered Values................................................................ 63
A.1.1.5 Ordered Probit (with Returns) Results: 10 Ordered Values........................................ 64
A.1.1.6 Ordered Logit Results: 4 Ordered Values................................................................... 65
A.1.1.7 Ordered Logit (with Returns) Results: 4 Ordered Values........................................... 66
A.1.1.8 Ordered Logit Results: 10 Ordered Values ................................................................. 67
A.1.1.9 Ordered Logit (with Returns) Results: 10 Ordered Values......................................... 68
A.1.2 Dataset 2: USD-USD April 1989 to June 1989: ......................................................... 69
A.1.2.1 MA(1)-GARCH(1,1) Results: .................................................................................... 69
A.1.2.2 Ordered Probit Results: 4 Ordered Values.................................................................. 70
A.1.2.3 Ordered Probit (with Returns) Results: 4 Ordered Values.......................................... 71
A.1.2.4 Ordered Probit Results: 10 Ordered Values................................................................ 72
A.1.2.5 Ordered Probit (with Returns) Results: 10 Ordered Values........................................ 73
A.1.2.6 Ordered Logit Results: 4 Ordered Values................................................................... 74
A.1.2.7 Ordered Logit (with Returns) Results: 4 Ordered Values........................................... 75
ii
A.1.2.8 Ordered Logit Results: 10 Ordered Values ................................................................. 76
A.1.2.9 Ordered Logit (with Returns) Results: 10 Ordered Values......................................... 77
A.1.3 Dataset 3: USD-USD April 2006 to May 2006: ......................................................... 78
A.1.3.1 MA(1)-GARCH(1,1) Results: .................................................................................... 78
A.1.3.2 Ordered Probit Results: 4 Ordered Values.................................................................. 79
A.1.3.3 Ordered Probit (with Returns) Results: 4 Ordered Values.......................................... 80
A.1.3.4 Ordered Probit Results: 10 Ordered Values................................................................ 81
A.1.3.5 Ordered Probit (with Returns) Results: 10 Ordered Values........................................ 82
A.1.3.6 Ordered Logit Results: 4 Ordered Values................................................................... 83
A.1.3.7 Ordered Logit (with Returns) Results: 4 Ordered Values........................................... 84
A.1.3.8 Ordered Logit Results: 10 Ordered Values ................................................................. 85
A.1.3.9 Ordered Logit (with Returns) Results: 10 Ordered Values......................................... 86
A.1.4 Dataset 4: USD-USD April 2009 to May 2009: ......................................................... 87
A.1.4.1 MA(1)-GARCH(1,1) Results: .................................................................................... 87
A.1.4.2 Ordered Probit Results: 4 Ordered Values.................................................................. 88
A.1.4.3 Ordered Probit (with Returns) Results: 4 Ordered Values.......................................... 89
A.1.4.4 Ordered Probit Results: 10 Ordered Values................................................................ 90
A.1.4.5 Ordered Probit (with Returns) Results: 10 Ordered Values........................................ 91
A.1.4.6 Ordered Logit Results: 4 Ordered Values................................................................... 92
A.1.4.7 Ordered Logit (with Returns) Results: 4 Ordered Values........................................... 93
A.1.4.8 Ordered Logit Results: 10 Ordered Values ................................................................. 94
A.1.4.9 Ordered Logit (with Returns) Results: 10 Ordered Values......................................... 95

iii
List of Figures & Tables

Figure 1: USD/SGD vs GDP & Quotes .....................................................................................3

Table 1: USD/SGD April-May 2006 Spread Behavior ..............................................................8
Table 2: Distribution of Spreads of USD/DEM April-June 1989.............................................32
Table 3: Quote Volume differences between B&M and Purchased Data USD/DEM 1989 .....35
Table 4: Frequency of No Spread Change reported by B&M USD/DEM 1989 ......................36
Table 5: Frequency of No Spread Change with Olsen USD/DEM 1989 .................................36
Table 6: Frequency Distribution of Spreads USD/DEM 1989.................................................37
Table 7: Volume of Quotes of USD/SGD April-June 1989......................................................37
Table 8: Frequency of No Spread Change for USD/SGD April-June 1989 .............................38
Table 9: Frequency Distribution of Spreads USD/SGD April-June 1989................................39
Table 10: Volume of Quotes of USD-SGD April-May 2006....................................................39
Table 11: Frequency of No Spread Change for USD-SGD April-May 2006 ...........................40
Table 12: Frequency Distribution of Spreads USD-SGD April-June 2006..............................41
Table 13: Volume of Quotes of USD/SGD April-May 2009....................................................41
Table 14: Frequency of No Spread Change for USD-SGD April-May 2009...........................42
Table 15: Frequency Distribution of Spreads USD-SGD April-May 2009..............................43
Table 16: GARCH Estimates for all Datasets ..........................................................................45
Table 17: Ordered Probit Estimates for all Datasets (4 ordered indicator values) ...................47
Table 18: Ordered Logit Estimates for all Datasets (4 ordered indicator values).....................47
Table 19: Ordered Probit Estimates for all Datasets (10 ordered indicator values) .................49
Table 20: Ordered Logit Estimates for all Datasets (10 ordered indicator values)...................50
Table 21: Ordered Probit Estimates (with Returns) for all Datasets (4 ordered values)...........51
Table 22: Ordered Logit Estimates (with Returns) for all Datasets (4 ordered values)............51
Table 23: Ordered Probit Estimates (with Returns) for all Datasets (10 ordered values).........52
Table 24: Ordered Logit Estimates (with Returns) for all Datasets (10 ordered values)..........52



iv
Acknowledgement

I would like to thank Professor Tse Yiu Kuen, Professor Hoon Hian Teck and
Professor Anthony Tay for their patience, guidance and advice to me with
regards to the completion of this thesis. Due to challenges in my professional
life (overseas posting!) and priorities in my personal life (wedding!), my
academic life (the completion of this thesis!) had unfortunately taken a backseat.
Without the generous assistance of these kind professors, this paper would never
have seen the light of day.

I also like to thank Ms. Lilian Seah for her tireless efforts with regards to the
administrative aspects of the thesis, and for tolerating my nonsense.

Lastly, I would like to thank Erfei for her love, patience, and steadfast
encouragement to me during this time when I spent more time on this thesis than
with her. I promise this will be my last part-time academic pursuit.

1
1 Introduction

In “Bid-Ask Spreads and Volatility in the Foreign Exchange Market – An
Empirical Analysis”, an early microstructure paper in 1994, Tim Bollerslev, and
Michael Melvin (henceforth, B&M), performed an empirical analysis on the
USD/DEM, one of the most highly-traded currency pair in 1989, and showed
that the size of the bid-ask spread of is positively related to its underlying
exchange rate uncertainty. Their dataset consist of more than 300,000
continuously recorded USD/DEM quotes over a 3 month period from April to
June in 1989.

In this paper, we want to examine whether B&M’s result is applicable to a much
lesser-traded currency belonging to a much smaller developing economy with a
government-managed floating exchange rate regime – the USD/SGD. We begin
with a dataset of the USD/SGD in the same 3 month period as per B&M. The
volume of USD/SGD quotes from April to June in 1989 is slightly over 8,000.
We also want to examine how relationship between the size of the bid-ask
spreads and exchange rate volatility changes as the USD/SGD grows in volume
and significance, and as Singapore evolves into a developed country. Hence, we
fast-forward 17 and 20 years into the future from 1989, and repeat the analysis
on over 600,000 USD/SGD quotes from April to May 2006, and on over 1
million USD/SGD quotes from the same months in 2009.
2
This paper examines the Singapore foreign exchange market from a
microstructure approach, specifically focusing on the bid-ask spreads of the
USD/SGD. We also present a review of some microstructure literature in this
area. But first, we provide more background to these two underlying themes in
the next two sections:

1.1 The Singapore Foreign Exchange Market
The Monetary Authority of Singapore operates a float regime for the Singapore
dollar that is managed against a basket of currencies of the country’s major
trading partners and competitors. The various currencies are given different
weights depending on the extent of trade dependence with that particular country.
The composition of the basket is revised periodically to take into account
changes in Singapore’ trade patterns.

The trade-weighted exchange rate is allowed to fluctuate within an undisclosed
policy band, which provides flexibility for the system to accommodate
short-term fluctuations in the foreign exchange markets as well as some buffer
in the estimation of Singapore’s equilibrium exchange rate.

On a trade-weighted basis, the SGD has appreciated against the exchange rates
of its major trading partners and competitors since 1981, reflected rapid
economic development, high productivity growth, and high savings rate.
3
The following figure shows the USD/SGD, Singapore GDP, and volume of
quotes over the months of April and May from 1989 to 2009.
USD/SGD vs GDP & Quotes
1.20
1.30
1.40
1.50
1.60
1.70
1.80
1.90
2.00
1988 1993 1998 2003 2008
Year
U
S
D
/
S
G
D
0
200
400
600
800
1000
1200
1400
1600
1800
G
D
P

S
$
B

/

Q
u
o
t
e
s

(
'
0
0
0
s
)
USD/SGD GDP Quotes

Figure 1: USD/SGD vs GDP & Quotes
1

From 1989 to 2009, as the Singapore GDP experienced a gradual growth from
SGD 61 billion in 1989 to SGD 266 billion in 2009, the USD/SGD fluctuated
between 1.40 and 1.90. Interestingly though, from 2002 onwards the growth of
USD/SGD quote volume for April and May experienced a sharp increase from
only 33,000 quotes in 2001 to 180,000 quotes in 2002. This growth would then
accelerate to 1,630,000 quotes in 2009. It is as if all of a sudden the Singapore
dollar started to become more widely traded than ever before. With increasing
volume, the underlying exchange rate volatility would also be expected to
increase. How much effect would this have on the bid-ask spreads of the
USD/SGD. We examine this using data from 1989, 2006, and 2009.

1
USD/SGD and GDP figures obtained from Singapore Department of Statistics (singstat.gov.sg); Quotes
volume obtained from Olsen Financial Technologies GmbH.
4
In 2006 the global economy had recorded a fourth consecutive year of strong
growth despite the drag from crude oil prices, a buildup in global electronics
inventories and adjustment in the US housing market. The MAS Annual Report
for 2005/2006 reported that “despite higher oil prices, rising interest rates and
natural disasters, the global economy expanded at a robust pace in 2005. This
growth momentum continued unabated in the first quarter of 2006. The strength
of the US economy was a major factor underpinning the continued growth of the
world economy last year. The US economy displayed remarkable resilience
against the backdrop of hurricane Katrina and 11 successive increases in the Fed
funds rate from 2.25% at the beginning of 2005 to 5% in May 2006. In the first
quarter of 2006, growth picked up strongly, led by a rebound in consumer
spending and business investment spending on equipment and software.”

The Singapore economy was also in a state of stability, as reported by the MAS
Annual Report 2005/2006: “In the early months of 2006, some signs of easing in
the domestic economy emerged with growth momentum slowing to 6.8% in Q1.
However, this is not indicative of a broad-based slowdown, but rather a
retraction to a more sustainable pace of growth.”

In 2009 the world found itself in the midst of the worst global financial crisis
ever since the Great Depression. The MAS Annual Report 2008/2009 reported
that “2008 was a tumultuous year for the global economy. While the surge in
5
commodity prices led to strong inflationary pressures in the first half of the year,
the onset of the global financial crisis caused world growth to fall sharply in the
later part of 2008 and into early 2009. The emergence of the Influenza A (H1N1)
virus in recent months has added a new dimension of risk to the fragile global
economy.”

The global financial crisis, which saw the collapse of Lehman Brothers in
September 2008, caused massive economic fallout worldwide. Amidst an
erosion of confidence, global trade and industrial production collapsed in the
first half of 2009, resulting in a 2.4% year-on-year contraction in world GDP
over the same period. During this time, the quote volume for USD/SGD (over
April and May) grew to over 1.6 million quotes, suggesting the increase in
volatility in the exchange rate.

1.2 The Microstructure Approach to Exchange Rate Economics
Exchange rate economics, the branch of international economics and finance
which attempt to explain the foreign exchange market, is an intriguing area of
research.

There are many theories of exchange rate determination, from the open economy
IS-LM models that are mandatory fare for any undergraduate economics course,
to more advanced models such as the Mundell-Fleming model, the
6
sticky/flexible-price monetary models, and the portfolio balance model. More
recently, new open-economy macroeconomists attempt for formalize exchange
rate in the context of dynamic general equilibrium models with explicit
microfoundations, nominal rigidities and imperfect competition.

When tested against empirical evidence, these theories have various degrees of
success in forecasting long-run exchange rates. All of them however are not able
to convincingly explain short-run exchange rate fluctuations.

From a common-sense perspective, this is hardly surprising. If the long-run
exchange rate between two countries is expected to change due to some shifting
fundamental value (say productivity level, for example), how would any macro
model (even one with microfoundations) designed to determine the “new”
equilibrium exchange rate be able to take into account all the possible paths
taken to transit from the “old” equilibrium rate to the “new” one? Some foreign
exchange transactions between the two countries might be related to shifting
fundamentals (e.g. import/export transactions), but others transactions may not
(e.g. tourism, speculation in each others’ asset markets, etc).

Macro foreign exchange models often assume that foreign exchange rates will
move when fundamentals move. But foreign exchange rates can only move
through a trading process in a foreign exchange market. Foreign exchange
7
microstructurists study this trading process.

In his book, “The Microstructure Approach to Exchange Rates”, Richard Lyons
(2001) defined “Order Flow” and “Bid-Ask Spreads”, two variables that are
absent from the macro approach, as the two hallmarks of the microstructure
approach. These are analogous to “Quantity” and “Price” in the dimension of
exchange rates.

Order flow is essentially transaction volume which is “signed”, meaning it
includes information if the transaction is a sale or a purchase. Such data is
usually hard to come by and are proprietary to banks and other high-level
participants in the foreign exchange market.

Well-connected researchers, such as Martin D. D. Evans and Richard K. Lyons
managed to obtain proprietary data on all end-user EUR/USD trades received at
Citibank over 6.5 years. In their 2007 paper “Exchange Rate Fundamentals and
Order Flow", they tested and established four empirical results: (1) transaction
flows forecast future macro variables such as output growth, money growth, and
inflation, (2) transaction flows forecast these macro variables significantly better
than the exchange rate does, (3) transaction flows (proprietary) forecast future
exchange rates, and (4) the forecasted part of fundamentals is better at
explaining exchange rates than standard measured fundamentals.
8
Data for the other hallmark, Bid-Ask Spreads, on the other hand, is much easier
to obtain. In fact, the four datasets for this paper were purchased from Olsen
Financial Technologies GmbH, while B&M obtained theirs by collecting every
USD/DEM quote posted on the Reuters screen for the interbank foreign
exchange market for three months in 1989. But other than being easily
obtainable, Lyons highlighted that one reason spreads receive so much attention
is because, being a core element of most data sets, they are a ready target for
testable hypotheses. This is in contrast to other features in the trading that are
not so easily measurable, such as information flow, belief dispersion, etc.

The behavior of bid-ask spreads in the foreign exchange markets offers many
opportunities for research. For example, Table 1 below shows the distribution of
over 600,000 USD/SGD quotes in April and May 2006, divided into nine
categories of price and spread movements.
SPREAD SPREAD SPREAD
UP SAME DOWN
PRICE UP 32.46% 6.35% 3.53%
PRICE SAME 4.05% 7.52% 3.42%
PRICE DOWN 3.21% 6.60% 32.85%
Total 39.73% 20.47% 39.81%

Table 1: USD/SGD April-May 2006 Spread Behavior

Out of over 600,000 quotes over two months, the spread remained unchanged
6.35% of the time when the price moved up, and 6.60% of time when the price
moved down. Though we may somewhat expect for spreads to move when
prices move, what is intriguing is that when prices move up, spreads tended to
9
widen 32.46% of the time, but yet narrowed only 3.53% of the time, Conversely,
when prices move down, spreads tended to narrow 32.85% of the time but yet
widened only 3.21% of the time.

Spread behavior like in the previous example may differ across different
currencies within the same time period, and may also differ across different time
periods within the same currency. The empirical objective of this paper is thus to
examine the relationship between the bid-ask spread and the underlying
exchange rate volatility from such two angles. The main currency for analysis is
the USD/SGD. For comparison against a different currency within the same time
period, we use B&M’s results for the USD/DEM. For comparison against
different time periods within the same currency, we perform this analysis for the
USD/SGD from the months of April and May in 1989, 2006, and 2009. In
addition, we also perform an empirical analysis on the phenomenon described
above in Table 1.

1.3 Organization of Thesis
The rest of this paper shall be organized as follows: Chapter 2 begins with a
review of microstructure papers which concerns bid-ask spreads, and ends with
a description of B&M’s model that relates volatility to bid-ask spreads. Chapter
3 describes B&M’s empirical methodology and presents the empirical analyses
for each dataset. Chapter 4 concludes.
10
2 Bid-Ask Spreads in Exchange Rates

As previously mentioned in Chapter 1, Lyons (2001) cites bid-ask spreads as
one of the two hallmarks of the microstructure approach. Besides being
relatively obtainable, he explained that spreads receive so much attention
because they form a core element of most data sets and are a ready target for
testable hypotheses. He also gave two more reasons for the heavy attention and
resources focused on bid-ask spreads.

The second reason was because practitioners are “intensely concerned with
managing trading costs”. The third reason had to do with the history of the field
of market microstructure which in its early days sought to distinguish itself from
the literature on trading under rational expectations. Rational expectations
models generally omit trading mechanisms when characterizing the relationship
between fundamentals and price. Contrastingly market microstructure focused
on how trading mechanisms affected prices, and this had led to “a focus on the
determination of real-world transaction prices – spreads.”

For the first section of this Chapter, guided by Sarno & Taylor (2002), we
present a short survey on some early microstructure literature that concerns the
determination of bid-ask spreads, and focus on “adverse selection” as a popular
theme. We then present the views of a more contemporary paper which refutes
11
“adverse selection” as a determinant of bid-ask spreads and proposes that
asymmetric information might be a more plausible candidate.

From the second section of this Chapter we return to our main topic of interest –
bid-ask spreads and exchange rate volatility. Similarly, following Sarno &
Taylor (2002), we first present a short survey on the literature which analyses
the proportional relationship between spreads and exchange rate volatility.

We then close this Chapter by presenting B&M’s simple asymmetric
information model which forms the framework for our empirical investigation.

2.1 Bid-Ask Spreads and Asymmetric Information
Sarno & Taylor (2002) identifies three main determinants of the bid-ask spread:
the cost of dealer services, inventory holding costs, and the cost of adverse
selection.

Cost of Dealer Services
The cost of dealer services is formally analyzed by Demsetz (1968) who
assumes the existence of some fixed costs of providing “predictable immediacy”
as the service for which compensation is required by market makers. While
Demsetz focused on the New York Stock Exchange, his definition for cost of
dealer services could also be applicable to the foreign exchange market.
12
According to Demsetz, “predictable immediacy is a rarity in human actions, and
to approximate it requires that costs be borne by persons who specialize in
standing ready and waiting to trade with the incoming orders of those who
demand immediate servicing of their orders. The ask-bid spread is the markup
that is paid for predictable immediacy of exchange in organized markets; in
other markets, it is the inventory markup of retailer or wholesaler”.

Inventory Holding Costs
The original argument of inventory costs as a crucial determinant of bid-ask
spreads was first propositioned by Barnear & Logue (1975), who tested a
modified theory of market-maker behavior first espoused by Bagehot (1971).
Barnear & Logue modified the theory of the market-maker spread by
distinguishing between the two major components of inventory risk. The second
component, which they termed “marketability risk”, relates to the market
maker's ability to make inventory adjustments when the market for an issue is
"thin." They showed that volume has a negative effect on the bid-ask spread for
two reasons: 1) high volume implies more competition if it implies more
competition among alternative market makers; and 2) high volume implies less
marketability risk, and, therefore, lower positioning costs.

Amihud & Mendelson (1980) considers the problem of a price-setting
monopolistic market-maker in a Garman (1976) dealership market where the
13
stochastic demand and supply are depicted by independent Poisson processes.
The focus of their analysis is the dependence of the bid-ask prices on the
market-maker’s inventory position. They derived the optimal policy the results
are shown to be consistent with some conjectures and observed phenomena, like
the existence of a ‘preferred’ inventory position and the downward monotonicity
of the bid-ask prices.

Ho & Stoll (1981) considers the stochastic dynamic programming problem of
solving for the optimal behavior of a single dealer of a single stock who is faced
with a stochastic demand for his services and return risk on his stock and on the
rest of his portfolio. They show that as time unfolds and transactions occur, the
dealer is able to set his bid price and ask price relative to his opinion of the
"true" price of the stock so as to maximize the expected utility of terminal
wealth. The bid-ask spread is given by a risk neutral spread that maximizes
expected profits for the given stochastic demand functions plus a risk premium
that depends on transaction size, the return variance of the stock and the dealer's
attitude toward risk. The bid-ask spread does not depend on the dealer's
inventory position, but the dealer’s price adjustment does. When inventory
increases both bid price and ask price decline, and the converse is true when
inventory decreases.


14
Generally, inventory costs models assume that market-markers optimize their
inventory holding, and generally imply that market-makers shift the spread
downwards and increase the width of the spread when a positive inventory is
accumulated.

Cost of Adverse Selection
Adverse selection is a common argument to explain the existence of bid-ask
spreads. The origin of this argument could be traced back to Bagehot (1971),
whose model includes two types of market participants – those are willing to
pay the price of the spread to the market-maker in exchange for predictable
immediacy and those who can speculate at the expense of the market-maker
using some private insider information. An adverse selection arises because
market-makers are not able to distinguish between the two types of participants
and resort to widening the spreads for both types. The bid-ask spread then
becomes the market-maker’s defense against adverse selection in “in the form of
exploitation of arbitrage opportunities”. Since Bagehot, numerous
microstructure papers have drawn on adverse selection as their primary
interpretive framework.

Copeland & Galai (1983) analyses the determination of bid-ask spreads in
organized financial markets, where the trading is done through economic agents
who specialize in market-making for a limited set of securities. The commitment
15
made by dealers to buy or sell at the bid and ask prices, respectively, is analyzed
as a combination of put and call options. Given the behavior of liquidity traders
and informed traders, the dealer is assumed to offer an out-of-the-money
straddle option for a fixed number of shares during a fixed time interval. The
exercise prices of the straddle determine the bid-ask spread. The dealer
establishes his profit maximizing spread by balancing the expected total
revenues from liquidity trading against the expected total losses from informed
trading. They showed that a monopolistic dealer will establish a wider bid-ask
spread than will perfectly competitive dealers, and that the bid-ask spread
increases with greater price volatility in the asset being traded, with a higher
asset price level, and with lower volume.

Glosten & Milgrom (1985) analyzed a model of a securities market in which the
arrival of traders over time is accommodated by a market-maker. They showed
that adverse selection, by itself, could account for the existence of a bid-ask
spread, and the average magnitude of the spread depends on many parameters,
including the exogenous arrival patterns of insiders and liquidity traders, the
elasticity of supply and demand among liquidity traders, and the quality of the
information held by insiders. They also showed that, because transaction prices
are informative, bid-ask spreads tend to decline with trade.


16
Lyons (1995) was likely one of the first who departed from early microstructure
work that focused almost entirely on stock markets and applied such theory to
foreign exchange markets. He presented a model which incorporated a number
of institutional features relevant to the FX market, such as the facts that major
currencies are traded in decentralized dealership markets; that over 80% of the
trading volume is between market-makers; that market net volume is only
partially observable; and that customer order flow is an important source of
private information. Lyons showed that trade size and the bid-ask spreads of a
particular dealer were positively related.

Payne (2003) estimates a VAR decomposition of interdealer trades and quotes
and interprets the results through the lens of adverse selection. Specifically, he
used one trading week’s worth of USD/DEM data derived from an electronic
foreign exchange brokerage and employed the framework contained in
Hasbrouck & Sofianos (1993) to test for the existence of private information
effects of trading on prices. His basic results confirm the existence of private
information on FX markets, indicating that adverse selection costs account for
around 60% of the half-spread.

Osler, Mende & Menkhoff (2006) however shows evidence that the behavior of
bid-ask spreads is inconsistent with adverse selection. They outline three factors
that seem likely to be important. The first factor, fixed operating costs, can
17
explain the negative relation between trade size and bid-ask spreads if some
costs are fixed, but cannot explain the cross-sectional variation across customer
types. To explain why bid-ask spreads are larger for commercial than financial
customers they suggest that asymmetric information – in the broad sense of
information that is held by some but not all market participants – may influence
spreads through two channels distinct from adverse selection, one involving
market power and a second involving strategic dealing.

The market power hypothesis suggests that firms, even in a market with
hundreds of competitors like foreign exchange, gain market power from holding
information. It can be costly for customer firms to search out the best available
quotes in the foreign exchange market, so each individual dealer can exert a
certain amount of market power despite the competition. As suggested in Green
et al. (2005), dealers may quote the widest spreads when their market power is
greatest, and market power in quote-driven markets depends on knowledge of
current market conditions. In foreign exchange, commercial customers typically
know far less about market conditions than financial customers so they might be
expected to pay wider spreads, as they do.

The second channel through which asymmetric information might affect bid-ask
spreads in foreign exchange involves strategic dealing. Building on abundant
evidence that customer order flow carries information (e.g., Evans and Lyons
18
(2007), Daníelsson et al. (2002)), Osler et al argue that rational foreign exchange
dealers might strategically vary spreads across customers, subsidizing spreads to
informed customers in order to gain information which they can then exploit in
upcoming interbank trades. In standard adverse-selection models, by contrast,
dealers passively accept the information content of order flow. The idea that
dealers strategically vary spreads to gather information was originally explored
in Leach and Madhavan (1992, 1993). When applied to two-tier markets in Naik
et al. (1999) it implies that bid-ask spreads will be narrower for trades with
information, consistent with the pattern in foreign exchange.

2.2 Bid-Ask Spreads and Volatility
The directly proportional relationship between bid-ask spreads and exchange
rate volatility now represents a fairly stylized fact in the microstructure literature.
Early studies modeled the spread as a function of transaction costs, the bank’s
profit from providing liquidity services, and the market-maker’s payoff for
facing the exchange rate risk when assuming an open position. The main
conclusions of these early studies are that exchange rate spreads are wider under
floating exchange rate than under fixed-exchange rate regimes (e.g. Aliber,
1975), and that measures of exchange rate volatility are followed closely by
exchange rate spreads (e.g. Fieleke, 1975; Overturf, 1982).


19
Glassman (1987) provides a significant contribution to this literature in that she
builds a model where variables representing transactions frequency are included
explicitly and the non-normality of the distribution of exchange rates is taken
into account. The model not only provides additional evidence on the
proportional relationship between exchange rate volatility and bid-ask spreads in
the foreign exchange market, but also suggests that market-makers consider
moments of the exchange rate higher than the second moment in order to
evaluate the probability of large exchange rate changes.

Admanti & Pfleiderer (1988) provides another fundamental theoretical
contribution to this area. In their model, there are three types of agents: informed
traders, who have relatively superior information and only trade on terms
favorable to them; discretionary liquidity traders, who must trade during a day
but can choose when to trade during the day in order to minimize costs; and
non-discretionary liquidity traders, who must trade at a precise time during the
day regardless of the cost. In this model, trading volume is explained by the
concentration of trade of informed traders and discretionary liquidity traders at
certain points in time: the concentrations occur because it is profitable for
informed traders to trade when there are many liquidity traders who do not have
the same information as themselves and because discretionary liquidity traders
are attracted because the larger the number of traders lowers the cost of trading.

20
Bollerslev & Domowitz (1993) used intradaily data to investigate the behavior
of quote arrivals and bid-ask spreads. They recorded quote arrivals and bid-ask
spreads over the trading day, across geographical locations as well as across
market participants. They found that trading activity and the bid-ask spreads for
traders whose activity is restricted to regional markets can be described by a
U-shaped distribution, which is consistent with Admanti & Pfleiderer’s (1989)
model. The patterns of trading activity and spreads during the day also strongly
suggest some degree of traders’ risk aversion, given which, the more trading
activity is executed by informed traders, the higher the cost of trading.

Goodhart & Figliuoli (1991) reported a study of minute bid-ask quotes on three
days in 1987 at a Reuters screen and found evidence that leptokurtosis and
heteroscedasticity are time-varying, and are less pronounced at the
minute-by-minute frequency than at lower frequencies. They also found that
trading volume is time-varying, being higher at the European and North
American openings and lower at the European lunch hour. The series was also
found to exhibit first-order negative serial correlation, which is especially
pronounced after immediately after jumps in the exchange rate. Multivariate
analysis suggested significant relationships between lagged exchange rates and
the current spot rate.

21
2.3 Bollerslev & Melvin’s Model of Volatility and Bid-Ask Spread
B&M is the main inspiration for this thesis, providing most importantly a
methodology to analyze the relationship between bid-ask spreads of exchange
rates and its underlying volatility.

In the early 1990’s as B&M were writing their paper, the bid-ask spread
component of transactions costs in the foreign exchange market had not received
much attention in the literature. Earlier studies on the subject, such as Glassman
(1987) and Boothe (1988), concentrated on the own statistical properties of the
spread. Researchers, such as Goodhart (1990), Bossaerts and Hillion (1991),
Black (1991), Melvin and Tan (1996) and Bollerslev and Domowitz (1993), had
attempted to offer empirical and theoretical analyses of the determinants of
foreign exchange market spreads, but no one had performed any explicit
analysis of the relationship between the magnitude of foreign exchange market
spreads and the underlying exchange rate volatility. Hence, B&M is likely to be
the first paper to touch on this subject.

B&M opined that while unambiguous 'good' or 'bad' news regarding the
fundamentals of the exchange rate should have no systematic effect on the
spread, as both the bid and the ask prices should adjust in the same direction in
response to the traders receiving buy or sell orders that reflect the particular
news event, however greater uncertainty regarding the future spot rate, as
22
associated with greater volatility of the spot rate, is likely to result in a widening
of the spread.

We now outline B&M’s simple theoretical framework that illustrates this role of
volatility in determining the spread.

The formal setup for B&M’s stylized market microstructure model is based on
the analysis in Glosten and Milgrom (1985), Admati and Pfleiderer (1989), and
Andersen (1993).

The model assumes that the foreign exchange market comprises two kinds of
traders: liquidity traders and information-based traders. Liquidity traders
participate in foreign exchange transactions only due to the needs of their
normal business activity which require international trade of goods, services and
financial assets. They are also not speculators. Information-based traders profit
by intermediating the demands and supplies of foreign exchange for the liquidity
traders. These traders also take positions in the foreign exchange market based
on information advantages received through their dealings with the liquidity
traders or, more generally, information asymmetries regarding fundamentals
underlying the determination of the spot exchange rate.


23
The liquidity traders constitute the proportion (1—ì) of the total market
participants. The liquidity traders receive a signal to either buy or sell foreign
currency regardless of the actual value of the currency in comparison with the
bid or ask prices prevailing at the time. Informed traders constitute the
remaining ì proportion of the market. This group of traders receives some
information about the true underlying fundamental value of the exchange rate s
t
.
This fundamental value is assumed to evolve over time according to a
martingale model,

(1)

where E
t-1
(c
t
)=0, E
t-1
(c
t
2
)= o
t
2
, and E
t-1
(•)

denotes the conditional expectation
based on the information set generated by the past values of s
t
. B&M further
assumes that the standardized innovations, c
t
o
t
-1
, are independent and
symmetrically, but not necessarily identically, distributed through time.

At time t-1, one of the many market-making traders will set bid and ask quotes,
B
t
and A
t
, good for trading at time t. The bid-ask spread is assumed to be set
symmetrically around the known fundamental price prevailing at the time of
quote formation, i.e. A
t
= s
t-1
+ k
t,t-1
, and B
t
= s
t-1
– k
t,t-1
. Thus, the quoted spread
for trades at time t, K
t
= A
t
, - B
t
= 2k
t,t-1
depends on time t-1 information only.

24
Trades at existing quotes will generate losses, on average, for the market-maker
when the opposite party an information-based trader. Information-based traders,
who received the signal c
t
buy currency if A
t
< s
t
and sell currency if s
t
, < B
t
. For
values of B
t
≤ s
t
≤ A
t
, the information-based traders cannot profit from
knowing the true fundamental value revealed by c
t.
The liquidity traders only
know s
t-1
and expect c
t
to equal zero.

Trader positions are limited by the convention that existing quotes are only good
for up to some maximum quantity of currency. Assuming that the market-makers
limit trading to one unit of currency at existing quotes, the loss for the quoting
trader relative to the true value s
t
arising from informed trading is therefore

(2)

Let P
t-1
(•) denote the probability conditional on the time t-1 information. Since
the standardized innovations, Z
t
= c
t
o
t
-1
, are assumed to be independent and
symmetrically distributed through time, the expected loss from informed trading
may be expressed as
(3)

25
Assuming an equal probability of a buy or a sell order from the liquidity traders,
it follows that the expected profit for the quoting trader conditional on an
uninformed trade equals:

(4)

Combining the expected trading loss in Eq. (3) with the gain in Eq. (4) yields the
expected profit for the market-maker conditional on time t-1 information:



In equilibrium, competition from other banks or market-makers will drive this
expected profit to zero. Expressing this zero profit condition in terms of the total
spread, K
t
=2
kt,t-1
yields

(6)

Since the conditional expectation and probabilities on the right-hand side of Eq.
(6) only depend on the time t-1 information set through o
t-1
k
t,t-1
, it follows that in
26
equilibrium the spread must move proportional to the conditional standard
deviation of the true fundamental value of the exchange rate.

B&M noted that while this simple proportionality condition would no longer
hold true in a more general model with endogenous information acquisition, the
result that an increase in o
t
2
leads to an increase in K
t
would still remain
generally valid.

Based on this relationship between exchange rate volatility and the bid-ask
spread, B&M designed the empirical methodology that we will describe in the
next chapter.


27
3 Empirical Analysis

B&M performed an empirical analysis on the USD/DEM, one of the most
highly-traded currency pair between two of the largest economies in the world in
1989 and showed that the size of the bid-ask spread of the USD/DEM is
positively related to its underlying exchange rate uncertainty. The USD/DEM is
a free-floating exchange rate. Their dataset consist of more than 300,000
continuously recorded USD/DEM quotes over a 3 month period from April to
June in 1989.

We are curious if 1) B&M’s result would hold for the USD/SGD, a semi-floating
currency from a much smaller economy, where over the same period from April
to May 1989, there were only slightly over 8,000 quotes; 2) B&M’s result would
hold 17 years later for that same currency, as that country becomes a significant
regional economic power in South-east Asia, and when the volume of quotes
increased to over 600,000 over April and May 2006.; and 3) B&M’s result
would hold as that economy enters into a period of worldwide financial crisis.

For the above purposes we purchased four sets of data from Olsen Financial
Technologies GmbH. Before we discuss the data and empirical results, we first
describe B&M’s empirical methodology and how we adapted it for this paper.

28
3.1 Empirical Methodology
B&M’s empirical methodology comprises two major steps.

Step 1: Creating a Proxy for Exchange Rate Volatility
The first step involves using a GARCH model as an explicit proxy for the
time-varying volatility of the spot rate, and as noted by Bollerslev et al. (1992),
such representations have been documented by numerous studies. B&M
employed a two-stage estimation procedure in which the conditional variance
for the spot exchange rate is first estimated as a GARCH process. These
estimates for the conditional variance are then used as the proxy for exchange
rate volatility in the second-stage model for the temporal behavior of the spread.

B&M use the ask price for estimation purposes; the bid and ask prices have
virtually identical higher order moments and differ only very slightly in their
conditional means. They found that the MA(1)-GARCH(1,1) model of the form
below seemed to fit their dataset well:

(7)

where I
t-1
denotes the time t-1 information set, and µ, u, e, o, and | are the
parameters to be estimated. The time t subscript refers to the place in the order
29
of the series of quotes, so that o
A,t
2
provides an estimate of the price volatility
between quotes.

The particular specification for the conditional variance in Eq. (7) may be
justified by the theoretical arguments in Nelson (1990, 1992). Intuitively, if the
sample path for the true unobservable volatility process is continuous, it follows
that on interpreting the GARCH(1,1) model as a non-parametric estimator, or a
one-sided filter, the resulting estimates for the conditional variance will
generally be consistent as the length of the sampling interval goes to zero.

The primary purpose of the GARCH estimation was to create proxies for the
conditional variance of the exchange rate to be used in the investigation of the
determinants of the spread.

But it was 1993 then, and B&M noted that estimating a GARCH model with
more than 300,000 observations in practice was not feasible in practice. Hence,
they divided up their dataset into 12 weeks of data and estimated each set for the
above GARCH parameters. They then saved all the estimates for the conditional
variances from each of the 12 models and combined into a single time series of
volatility estimates for the full set of weekday quotes. Today, we use modern
econometrics software EViews to perform GARCH analysis on our datasets and
to obtain the conditional variance time series.
30
Step 2: Estimate Relationship between Spreads and Proxy for Volatility
The second major step of the methodology involved using ordered response
models to estimate the relationship between the time series of volatility
estimates obtained in the first step and the bid-ask spreads.

Specifically, B&M used an Ordered Probit model with multiplicative
heteroskedasticity for this purpose.

The observed spread, K
t
, is assumed to take on only a fixed number of discrete
values, a
1
, a
2
, … a
J
. The unobservable continuous random variable, K*, is
defined by

K
*
t
= X
i
’| + c
K,t
(8)

The vector X
t
denotes a set of predetermined variables that affect the conditional
mean of K*
t
and c
K,t
, is conditionally normally distributed with mean zero and
variance, c
2
K,t


(9)

B&M allowed for multiplicative heteroskedasticity in the spread by
parameterizing the logarithm of o
2
K,t
as a linear function of the same
31
explanatory variables that enter the conditional mean of K*
t
.

We depart from this path in our analysis, firstly by omitting multiplicative
heteroskedasticity (more due to the limitations of EViews than by choice), and
secondly by estimating the relationships with both Ordered Probit models and
Ordered Logit models (to allow more flexibility in the behaviour of the error
term, since we omitted multiplicative heteroskedasticity).

In our analyses, the observed spread, K
t
, is similarly assumed to take on only a
fixed number of discrete values, a
1
, a
2
, … a
J
, but the unobservable continuous
random variable, K*, is defined by

K
*
t
= X
i
’| + c

(10)

where c is i.i.d. standard normal for the Ordered Probit model, and takes the
form of the logistic distribution for the Ordered Logit model.

The ordered response models relate the observed spreads to K* via

(11)

where the A
j
’s form an ordered partition of the real line into J disjoint intervals.
32
The probability that the spread takes on the value a
j
is equal to the probability
that K* falls into the appropriate partition, A
j
.

For tractability reasons, B&M based the empirical analysis on a classification of
the spread into only four different categories. From the distribution of spreads of
the USD/DEM in 1989, the four most commonly observed spreads account for
97.0 percent of the total quotes.

USD/DEM
Frequency Distribution fof Spreads
Spread All Quotes
0 < . < 5 2,304 (0.8%)
5 77,856 (25.6%)
5 < . < 7 607 (0.2%)
7 34,878 (11.5%)
7 < . < 10 2,977 (1.0%)
10 170,892 (56.1%)
10 < . < 15 329 (0.1%)
15 11,534 (3.8%)
15 < . < 20 39 (0.0%)
20 2,616 (0.9%)
20 < . 572 (0.2%)
Note: Spreads converted into basis points

Table 2: Distribution of Spreads of USD/DEM April-June 1989

We also performed the ordered response analyses with four ordered indicator
values. In addition, as we will see later, because the spreads distribution for the
USD/SGD 2006 and 2009 data have more converging points (certain spread
sizes where there are more quotes than others) than both the USD/DEM and the
USD/SGD in 1989, we used up to 10 ordered indicator values in our own
33
ordered response analyses.

In the case of B&M where by four ordered indicator values of a
j
’s were used, the
corresponding intervals for the unobservable latent variable K* are defined by:

(12)

The partition parameters, µ
I
, are estimated jointly with the other parameters of
the model.

The ordered response model defined above allows us to estimate the probability
of a particular spread being observed as a function of the predetermined
variables, X
t
. In order to test the hypothesis that the spread is partly determined
by the volatility of the spot rate, the GARCH estimate of the conditional
variance for the ask prices is included as one of the elements in X
t
.

B&M noted that Bollerslev and Domowitz (1993) indicated a distinct intra-day
pattern in the spread distribution and tt is possible that any significant effect of
the conditional variance in isolation may merely reflect this dependence rather
than provide an independent influence on the spread process. In order to take
34
account of this own temporal dependence, K
t-1
was included as an element of the
X
t
vector in the estimation of the ordered response functions for K*
t
:

K
t
* = o
1
o
A,t
2
+ o
2
K
t-1
+ c (13)

Given the partition boundaries determined by the data, if a higher conditional
mean o'X, is caused by a larger conditional variance of the spot rate, and this
raises the probability of observing a higher spread, we will infer that the
hypothesized theoretical link is supported by the empirical analysis.

We now describe each of the datasets in more details in the following section:

3.2 Description of the Data
We purchased four sets of data from Olsen Financial Technologies GmbH:
a. Dataset 1 – USD/DEM quotes from April 1989 to June 1989
b. Dataset 2 – USD/SGD quotes from April 1989 to June 1989
c. Dataset 3 – USD/SGD quotes from April 2006 to May 2006
d. Dataset 4 – USD/SGD quotes from April 2009 to May 2009

Dataset 1: USD/DEM from April 1989 to June 1989
For the first dataset we purchased the same dataset used by B&M – USD-DEM
quotes from April 1989 to June 1989. We wanted to repeat the empirical analysis
35
on the dataset again to allow for an apples-to-apples comparison between the
USD/DEM 1989 and the USD/SGD 1998 results.

First, B&M’s results were obtained in 1993 using probably not so
technologically advanced means. We wanted to repeat the estimation processes
again using EViews. Furthermore, as mentioned in the previous Chapter, we
departed from B&M’s ordered probit procedure by omitting multiplicative
heteroskedasticity. Second, B&M’s data were obtained from Reuters. From their
paper we were unable to ascertain the accuracy of this data, or how they
presented 12 workweeks of data from an actual 13 weeks in April to June 1989.
We obtained ours by purchasing from Olsen Financial Technologies GmbH and
upon comparison, we found some minor differences. The differences between
our purchased data and the actual data used by B&M are tabulated as follows:
a. Volume of Quotes
ORIGINAL BOLLERSLEV & MELVIN PURCHASED FROM OLSENDATA
DAY TICKS DAY TICKS
Sun 887 0.29% Sun 1,571 0.51%
Mon 60,095 19.66% Mon 56,600 18.25%
Tue 66,109 21.63% Tue 67,634 21.80%
Wed 63,812 20.88% Wed 66,325 21.38%
Thu 61,521 20.13% Thu 61,377 19.79%
Fri 53,082 17.37% Fri 56,646 18.26%
Sat 98 0.03% Sat 33 0.01%
Grand Total 305,604 100.00% Grand Total 310,186 100.00%

Table 3: Quote Volume differences between B&M and Purchased Data USD/DEM 1989
B&M collected 305,604 ticks of USD/DEM quotes while the dataset purchased
from Olsen Financial Technologies GmbH contains 310,186 ticks. The
distribution of the quotes over each workday is similar in terms of percentage in
both datasets.
36
b. Frequency of No Spread Change
USD/DEM
No. of Quotes, & Frequency of No Spread Change
No Change in Spread
Number Bid-Ask Rise Bid-Ask Fall
All quotes 304,619 8.00% 8.30%

Table 4: Frequency of No Spread Change reported by B&M USD/DEM 1989
B&M reported that 8% of all the quotes observed no change in spread when the
bid-ask price rose and that 8.3% of all the quotes observed no change in spread
when the bid-ask price fell. It was not clear, however, whether these percentages
where calculated based on Number of Ticks with No Spread Change divided by
Number of Ticks Moved, or on Number of Ticks with No Spread Change
divided by Total Number of Ticks. We make this distinction clearer with the
purchased data.
SPREAD SPREAD SPREAD
UP SAME DOWN
PRICE UP 64384 32787 23767
PRICE SAME 13083 39422 13523
PRICE DOWN 24045 33407 64164
Grand Total 101512 105616 101454
SPREAD SPREAD SPREAD
UP SAME DOWN
PRICE UP 53.24% 27.11% 19.65%
PRICE SAME 19.81% 59.70% 20.48%
PRICE DOWN 19.77% 27.47% 52.76%
GRAND TOTAL 32.90% 34.23% 32.88%
SPREAD SPREAD SPREAD
UP SAME DOWN
PRICE UP 20.86% 10.63% 7.70%
PRICE SAME 4.24% 12.78% 4.38%
PRICE DOWN 7.79% 10.83% 20.79%
GRAND TOTAL 32.90% 34.23% 32.88%
NO. OF QUOTES
DIVIDED BY NO OF TICKS MOVED
DIVIDED BY TOTAL NO. OF TICKS

Table 5: Frequency of No Spread Change with Olsen USD/DEM 1989
Among all ticks, we counted that 10.63% and 10.83% of all ticks observed no
change in spreads when the price rose and fell respectively.
37
c. Frequency Distribution of Spreads
Frequency Distribution of Spreads
Spread
0 < . < 5 2,304 (0.8%) 2,562 (0.8%)
5 77,856 (25.6%) 81,368 (26.4%)
5 < . < 7 607 (0.2%) 660 (0.2%)
7 34,878 (11.5%) 35,682 (11.6%)
7 < . < 10 2,977 (1.0%) 3,192 (1.0%)
10 170,892 (56.1%) 171,264 (55.5%)
10 < . < 15 329 (0.1%) 327 (0.1%)
15 11,534 (3.8%) 10,389 (3.4%)
15 < . < 20 39 (0.0%) 38 (0.0%)
20 2,616 (0.9%) 2,547 (0.8%)
20 < . 572 (0.2%) 553 (0.2%)
ORIGINAL OLSENDATA
Note: Spreads converted into basis points

Table 6: Frequency Distribution of Spreads USD/DEM 1989
It appears that the frequency distribution of spreads is quite similar between the
dataset reported by B&M and the dataset purchased from Olsen Financial
Technologies GmbH. The most common bid-ask spread is 10 basis points,
followed by 5 basis points.
Dataset 2: USD/SGD from April 1989 to June 1989
The second dataset are USD/SGD quotes from April 1989 to June 1989.
a. Volume of Quotes
DAY USD/SGD
Sun 6 0.07%
Mon 1,307 15.43%
Tue 2,024 23.89%
Wed 1,876 22.14%
Thu 1,652 19.50%
Fri 1,586 18.72%
Sat 21 0.25%
Grand Total 8,472 100.00%

Table 7: Volume of Quotes of USD/SGD April-June 1989
Compared against a major currency in 1989 over the same period from April to
June, there are only 8,472 USD/SGD quotes compared to over 310,000 for the
USD/DEM. Distribution of the quotes over the week are however quite similar,
38
with volume peaking during midweek.

b. Frequency of No Spread Change
Compared against the USD/DEM in 1989 over the same period from April to
June, we do not observe as much spread changes in the USD/SGD when prices
moves. When prices move upwards, the spread remained unchanged 89.21% of
the time. When prices move downwards, the spread remained unchanged
90.62% of the time.

SPREAD SPREAD SPREAD
UP SAME DOWN
PRICE UP 87 3108 289
PRICE SAME 62 1340 63
PRICE DOWN 256 3168 72
Grand Total 405 7616 424
SPREAD SPREAD SPREAD
UP SAME DOWN
PRICE UP 2.50% 89.21% 8.30%
PRICE SAME 4.23% 91.47% 4.30%
PRICE DOWN 7.32% 90.62% 2.06%
GRAND TOTAL 4.80% 90.18% 5.02%
SPREAD SPREAD SPREAD
UP SAME DOWN
PRICE UP 1.03% 36.80% 3.42%
PRICE SAME 0.73% 15.87% 0.75%
PRICE DOWN 3.03% 37.51% 0.85%
GRAND TOTAL 4.80% 90.18% 5.02%
NO. OF QUOTES
DIVIDED BY NO OF TICKS MOVED
DIVIDED BY TOTAL NO. OF TICKS

Table 8: Frequency of No Spread Change for USD/SGD April-June 1989




39
c. Frequency Distribution of Spreads
Frequency Distribution of Spreads
Spread
0 < . < 5 6 (0.1%)
5 137 (1.6%)
5 < . < 7 22 (0.3%)
7 122 (1.4%)
7 < . < 10 43 (0.5%)
10 7,854 (93.0%)
10 < . < 15 6 (0.1%)
15 76 (0.9%)
15 < . < 20 1 (0.0%)
20 164 (1.9%)
20 < . 14 (0.2%)
USD/SGD
Note: Spreads converted into basis points

Table 9: Frequency Distribution of Spreads USD/SGD April-June 1989

While spread changes in the USD/SGD are uncommon when prices moves, we
note similar characteristics to the USD/DEM in that the most common bid-ask
spread is 10 basis points. When spreads do change, they are most likely to be 5,
7 or 20 basis points.

Dataset 3: USD/SGD from April 2006 to May 2006
The third dataset are USD/SGD quotes from April 2006 to May 2006.
a. Volume of Quotes
DAY USD/SGD
Sun 3,832 0.63%
Mon 104,327 17.24%
Tue 126,812 20.96%
Wed 147,084 24.31%
Thu 135,037 22.32%
Fri 87,332 14.44%
Sat 555 0.09%
Grand Total 604,979 100.00%

Table 10: Volume of Quotes of USD-SGD April-May 2006
15 years later, the USD/SGD has grown to become a major currency in
Southeast Asia. Compared against itself in 1989 over the period from April to
May, the volume of quotes of the USD/SGD has grown from over 8,400 ticks in
40
3 months to over 604,000 ticks in 2 months. The distribution of quotations over
the week remains consistent with volume peaking during midweek.

b. Frequency of No Spread Change
In 2006, we observed that the tendency for spreads to remain unchanged when
prices move is reduced dramatically. Now, when prices move upwards, the
spread remained unchanged only 14.99% of the time compared to 89.21% of the
time in 1989. When prices move downwards, the spread remained unchanged
only 15.46% of the time compared to 90.62% of the time in 1989.

SPREAD SPREAD SPREAD
UP SAME DOWN
PRICE UP 194955 38121 21197
PRICE SAME 24340 45185 20568
PRICE DOWN 19302 39611 197313
Grand Total 238597 122917 239078
SPREAD SPREAD SPREAD
UP SAME DOWN
PRICE UP 76.67% 14.99% 8.34%
PRICE SAME 27.02% 50.15% 22.83%
PRICE DOWN 7.53% 15.46% 77.01%
GRAND TOTAL 39.73% 20.47% 39.81%
SPREAD SPREAD SPREAD
UP SAME DOWN
PRICE UP 32.46% 6.35% 3.53%
PRICE SAME 4.05% 7.52% 3.42%
PRICE DOWN 3.21% 6.60% 32.85%
GRAND TOTAL 39.73% 20.47% 39.81%
NO. OF QUOTES
DIVIDED BY NO OF TICKS MOVED
DIVIDED BY TOTAL NO. OF TICKS

Table 11: Frequency of No Spread Change for USD-SGD April-May 2006



41
c. Frequency Distribution of Spreads
Frequency Distribution of Spreads
Spread
0 < . < 5 51,090 (8.5%)
5 224,763 (37.2%)
5 < . < 7 51,163 (8.5%)
7 68,377 (11.3%)
7 < . < 10 53,582 (8.9%)
10 155,466 (25.7%)
10 < . < 15 25 (0.0%)
15 17 (0.0%)
15 < . < 20 15 (0.0%)
20 1 (0.0%)
20 < . 0 (0.0%)
USD/SGD
Note: Spreads converted into basis points

Table 12: Frequency Distribution of Spreads USD-SGD April-June 2006
The characteristics of the frequency distribution of spreads have also changed
almost completely over 15 years. The most common spread in 2006 is 5 basis
points, followed by 10 basis points. Most of the spreads recorded are either 10
basis points or lower, while quotes with spreads of more than 10 basis points
make up less than 0.01% of the entire spectrum of quotes.
Dataset 4: USD/SGD from April 2009 to May 2009
The fourth dataset are USD/SGD quotes from April 2009 to May 2009.
a. Volume of Quotes
DAY USD/SGD
Sun 9,336 0.87%
Mon 183,055 17.02%
Tue 217,496 20.22%
Wed 237,508 22.08%
Thu 237,604 22.09%
Fri 190,603 17.72%
Sat 32 0.00%
Grand Total 1,075,634 100.00%

Table 13: Volume of Quotes of USD/SGD April-May 2009
In the year of the 2009 economic crisis, we observed a dramatic increase in
volume of USD-SGD quotes over the period of April to May 2009, as compared
to the same period in 2006, although we noted earlier that the volume growth
42
had been exponential since 2002. The volume of quotes of the USD-SGD has
grown from over 604,000 ticks, to exceeding 1 million quotes over a 2 month
period.

b. Frequency of No Spread Change
We now observed that in 2009, the tendency for spreads to remain unchanged
when prices moved is reduced again compared to 2006. Now, when prices move
upwards, the spread remained unchanged 10.17% of the time compared to
14.99% of the time in 1989. When prices move downwards, the spread remained
unchanged 9.76% of the time compared to 15.46% of the time in 2006.

SPREAD SPREAD SPREAD
UP SAME DOWN
PRICE UP 304279 48129 120968
PRICE SAME 59290 18050 53790
PRICE DOWN 120972 45055 295733
Grand Total 484541 111234 470491
SPREAD SPREAD SPREAD
UP SAME DOWN
PRICE UP 64.28% 10.17% 25.55%
PRICE SAME 45.21% 13.76% 41.02%
PRICE DOWN 26.20% 9.76% 64.04%
GRAND TOTAL 45.44% 10.43% 44.13%
SPREAD SPREAD SPREAD
UP SAME DOWN
PRICE UP 28.54% 4.51% 11.35%
PRICE SAME 5.56% 1.69% 5.04%
PRICE DOWN 11.35% 4.23% 27.74%
GRAND TOTAL 45.44% 10.43% 44.13%
DIVIDED BY NO OF TICKS MOVED
DIVIDED BY TOTAL NO. OF TICKS
NO. OF QUOTES

Table 14: Frequency of No Spread Change for USD-SGD April-May 2009


43
c. Frequency Distribution of Spreads
Frequency Distribution of Spreads
Spread
<2.00 13 (0.0%)
2 13,684 (1.3%)
2.00 < . < 3.00 8,422 (0.8%)
3 54,432 (5.1%)
3.2 20,407 (1.9%)
3.20 < . < 3.50 6,484 (0.6%)
3.5 114,907 (10.7%)
3.50 < . < 4.00 10,882 (1.0%)
4 88,441 (8.2%)
4.00 < . < 5.00 13,748 (1.3%)
5 59,572 (5.5%)
5.00 < . < 6.00 6,932 (0.6%)
6 54,354 (5.1%)
6.00 < . < 7.00 2,903 (0.3%)
7 165,828 (15.4%)
7.00 < . < 8.00 2,080 (0.2%)
8 268,577 (25.0%)
8.00 < . < 9.00 1,158 (0.1%)
9 39,789 (3.7%)
9.00 < . < 10.00 830 (0.1%)
10 91,613 (8.5%)
10.00 < . < 11.00 579 (0.1%)
11 11,304 (1.1%)
> 11.00 38,696 (3.6%)
USD/SGD
Note: Spreads converted into basis points

Table 15: Frequency Distribution of Spreads USD-SGD April-May 2009
The characteristics of the frequency distribution of spreads have changed again
over 3 years. The most common spread in 2006 was 5 basis points, followed by
10 basis points. In 2009, the most common spread had risen to 8 basis points,
followed by 7 basis points and 3.5 basis points respectively. Most of the spreads
recorded are still either 10 basis points or lower, but quotations with spreads of
more than 10 basis points have increased to slightly under 5% of the entire
spectrum of quotations.

Two observations are of noteworthy regarding the 2009 dataset. First, the most
common spreads are no longer “significant” numbers such as 5 or 10. This could
indicate advancement in the market’s ability to evaluate foreign exchange risk,
44
and hence being able to price bids and asks more accurately. Second, the volume
of quotes with spreads of more than 10 basis points increased from 0.01% in
2006 to slightly under 5% in 2009. This could be attributed to the increased
amount of uncertainty in the financial markets during that period of global
economic crisis. Both phenomena could warrant future research.

3.3 GARCH Analysis
The primary purpose of the GARCH estimation was to create proxies for the
conditional variance of the exchange rate to be used in the investigation of the
determinants of the spread. Recall from Section 3.1 that, following B&M, we
estimate for each dataset with the following MA(1)-GARCH(1,1) model:

(7)

where I
t-1
denotes the time t-1 information set, and µ, u, e, o, and | are the
parameters to be estimated. The time t subscript refers to the place in the order
of the series of quotes, so that o
A,t
2
provides an estimate of the price volatility
between quotes.

Following B&M, we first removed all weekend quotes. We then used EViews to
estimate the parameters for each dataset and the full results are attached in
45
Appendix 1. Here we present a summary of the results:
B&M
DEM1989 DEM1989 SGD1989 SGD2006 SGD2009
µ 0.0065 0.0069 0.0452 0.0025 0.0023
(0.0042) (0.0013) (0.0138) (0.0004) (0.0003)
u -0.5953 -0.5867 -0.3281 -0.7817 -0.7802
(0.0052) (0.0014) (0.0114) (0.0009) (0.0006)
e 0.1008 0.0500 0.6833 0.0578 0.0453
(0.0053) (0.0005) (0.0256) (0.0008) (0.0002)
o 0.0652 0.0561 0.2650 0.0632 0.0513
(0.0018) (0.0003) (0.0081) (0.0006) (0.0002)
| 0.9057 0.9327 0.6540 0.9025 0.9197
(0.0030) (0.0003) (0.0075) (0.0009) (0.0003)
o+| 0.9708 0.9888 0.9190 0.9657 0.9710
T 304,608 308,748 8,444 551,355 1,006,736
Asymptotic errors are reported in parenthesis

Table 16: GARCH Estimates for all Datasets

The first column contains the average of the weekly estimates from the original
B&M dataset (they only estimated these parameters per workweek, for 12
weeks). The second to fifth columns contain the EViews GARCH estimates for
the parameters for the purchased USD/DEM 1989, USD/SGD 1989, USD/SGD
2006, and USD/SGD 2009 data respectively.

First we observe that the EViews estimates for the purchased USD/DEM dataset
compares very well with B&M’s original results, though standard errors are
significantly lower. This provides confidence that the EViews results for the rest
of the datasets are appropriate for comparison.

46
The second observation is that all estimates for u are negative, which
corresponds to B&M’s results, and they noted that “the negative estimates for u
may be partly attributed to a non-synchronous quoting phenomenon; see Lo and
MacKinlay (1990) for a formal analysis.”

The GARCH effects for all datasets are all highly significant. Comparing
GARCH effects between the USD/DEM 1989 dataset and USD/SGD 1989
dataset, it is clear that the USD/DEM dataset shows much stronger effects, and
stronger o+| volatility persistence. This is expected, as the USD/DEM dataset
contains at least 36 times more observations and from Table 8, the spread did not
change 90.18% of the time.

As the Singapore economy grows and the USD/SGD becomes a significant
regional currency, the results show that the GARCH effects and persistence of
volatility is consistent with this phenomenon.

The primary purpose of the GARCH estimation was to create proxies for the
conditional variance of the exchange rate to be used in the investigation of the
determinants of the spread. To this end, we used EViews to obtain GARCH
variance series for each dataset.

47
3.4 Ordered Response Analysis
Recall from Section 3.1 that we estimate for each dataset with the following
ordered response model (probit and logit):

K
t
* = o
1
o
A,t
2
+ o
2
K
t-1
+ c (13)

We first use four ordered indicator values and B&M’s definitions for a
j
’s:

a
1
: <= 5 a
2
: 5 < . < 10 a
3
: = 10 a
4
: > 10

The results for the above parameters for all four datasets are attached in
Appendix 1, but a summary is presented below:
DEM1989 SGD1989 SGD2006 SGD2009
o1 0.0214 0.0091 0.0464 0.0988
(73.9604) (10.6719) (29.2412) (113.8863)
o2 0.3116 0.0422 -0.1306 0.2622
(135.3568) (22.3298) -(69.4622) (177.1729)
T 308,748 8,444 551,355 1,006,736
Z-statistics are reported in parenthesis

Table 17: Ordered Probit Estimates for all Datasets (4 ordered indicator values)

DEM1989 SGD1989 SGD2006 SGD2009
o1 0.1299 0.0231 0.0802 0.2146
(102.6436) (9.2342) (28.9306) (109.5740)
o2 0.4330 1.8160 -0.2040 0.4395
(108.2955) (23.1592) -(66.7912) (169.4684)
T 308,748 8,444 551,355 1,006,736
Z-statistics are reported in parenthesis

Table 18: Ordered Logit Estimates for all Datasets (4 ordered indicator values)
48
The positive o
1
coefficients for both ordered probit and logit analyses above
suggest that there is a significantly positive effect of exchange rate volatility on
the spread for all datasets. The conditional mean of K
t
* is an increasing function
of o
A,t
2
. This is consistent with the implications drawn from B&M’s theoretical
model. The estimates for o
2
are indicative of intra-day persistence in the spread
process.

The magnitude of o
1
for each dataset supports what we intuitively already know.
Comparing the USD/DEM and USD/SGD in 1989, although both o
1
values are
statistically significant, volatility appears to play a much larger role in
determining the size of the spread for the USD/DEM. After all, as noted above,
the spread remained the same 90.18% of the time for the USD/SGD in 1989. For
the same reason, the o
2
values show that the dependence on the previous spread
appears to be more significant for the USD/SGD than for the USD/DEM.

As the country grow in economic significance and the SGD becoming a major
regional currency over 17 years, the magnitude of the o
1
values almost
quadrupled, coincidentally matching the growth in quote volume. It is
interesting to note the negative o
2
values for the USD/SGD 2006 dataset, but it
might be due to seasonality.


49
In 2009, in midst of the worldwide financial crisis, we observed a higher level of
volatility in the USD/SGD. The much higher o
1
values for this dataset, together
with the larger distribution of spreads as shown in Table 15, supports the theory
that the bid-ask spreads of foreign exchange rates is very much positively related
to the underlying volatility.

We repeated the ordered response analyses using 10 ordered indicator values,
due to the larger distribution of spreads as seen in the 2006 and 2009 data. The
a
j
’s are:

a
1
: < 3 a
2
: 3 <= . < 4 a
3
: 4 <= . < 5
a
4
: 5 <= . < 6 a
5
: 6 <= . < 7 a
6
: 7 <= . < 8
a
7
: 8 <= . < 9 a
8
: 9 <= . < 10 a
9
: = 10
a
10
: > 10

Again, the results for the above parameters for all four datasets are attached in
Appendix 1, but a summary is presented below:
DEM1989 SGD1989 SGD2006 SGD2009
o1 0.0345 0.0106 0.0296 0.0964
(90.1741) (12.7664) (19.2017) (117.0911)
o2 0.1124 0.2933 -0.0404 0.0884
(122.5758) (16.0766) -(61.5938) (212.8967)
T 308,748 8,444 551,355 1,006,736
Z-statistics are reported in parenthesis

Table 19: Ordered Probit Estimates for all Datasets (10 ordered indicator values)
50
DEM1989 SGD1989 SGD2006 SGD2009
o1 0.1557 0.0305 0.0599 0.2047
(115.2591) (10.8283) (22.1246) (110.9948)
o2 0.1599 0.5637 -0.0755 0.1493
(100.3382) (16.7295) -(67.7863) (205.0571)
T 308,748 8,444 551,355 1,006,736
Z-statistics are reported in parenthesis

Table 20: Ordered Logit Estimates for all Datasets (10 ordered indicator values)

Generally, the conclusions from the previous ordered response analyses with 4
ordered indicator values still holds after we attempt to be more discerning with
the ordered indicator values.

3.5 Incorporating Returns
Recall that we noted (see Tables 5, 11 and 14) an interesting phenomenon
regarding the behavior of bid-ask spreads of foreign exchange rates. When
prices move up, and the bid-ask spreads change, they tend to widen. Conversely,
when prices move down, and the bid-ask spreads change, they tend to narrow.

Previously we were examining the relationship between bid-ask spreads and the
underlying volatility of exchange rates and found that positive relationships
generally exists between the two. We are now curious if there is any relationship
between bid-ask spreads and price movement. To examine this we repeat the
ordered response analyses by including the Returns variable from the
MA(1)-GARCH(1,1) model into the ordered response model:
51

K
t
* = o
1
o
A,t
2
+ o
2
K
t-1
+ o
3
10,000*AlogA
t
+ c (14)

We repeated the ordered probit and ordered logit analyses for all four datesets
using both 4 and 10 ordered indicator values. The results are attached in
Appendix 1, but we summarize the results below:
DEM1989 SGD1989 SGD2006 SGD2009
o1 0.0148 0.0093 0.0325 0.0406
(51.4649) (10.9454) (19.1288) (44.7718)
o2 0.0383 1.0090 0.5203 0.2606
(16.2662) (23.4338) (210.7702) (146.0215)
o3 0.0970 0.0827 0.6241 0.4158
(98.3419) (12.2743) (449.5486) (437.6876)
T 308,748 8,444 551,355 1,006,736
Z-statistics are reported in parenthesis

Table 21: Ordered Probit Estimates (with Returns) for all Datasets (4 ordered values)

DEM1989 SGD1989 SGD2006 SGD2009
o1 0.1319 0.0235 0.0485 0.2225
(101.6830) (8.7724) (14.8530) (96.1127)
o2 0.6732 1.9380 1.0532 1.9644
(154.6104) (24.1228) (229.6602) (494.6543)
o3 0.3205 0.1796 1.2152 1.3865
(174.0656) (12.2520) (409.7483) (585.1713)
T 308,748 8,444 551,355 1,006,736
Z-statistics are reported in parenthesis

Table 22: Ordered Logit Estimates (with Returns) for all Datasets (4 ordered values)
52
DEM1989 SGD1989 SGD2006 SGD2009
o1 0.0205 0.0108 -0.0350 -0.0032
(54.0252) (13.0826) -(22.6917) -(3.8439)
o2 0.0042 0.3135 0.0185 0.0949
(4.5016) (17.0041) (22.5621) (188.9157)
o3 0.0961 0.0745 0.0012 0.3996
(105.6836) (11.3375) (298.6273) (468.5108)
T 308,748 8,444 551,355 1,006,736
Z-statistics are reported in parenthesis

Table 23: Ordered Probit Estimates (with Returns) for all Datasets (10 ordered values)

DEM1989 SGD1989 SGD2006 SGD2009
o1 0.1592 0.0317 0.0673 0.2688
(113.7732) (10.6241) (21.1859) (124.5891)
o2 0.2532 0.6059 0.4030 0.5755
(146.3672) (17.7207) (251.9007) (574.3314)
o3 0.3182 0.1677 1.1891 1.2835
(173.2477) (11.3873) (434.4841) (654.4871)
T 308,748 8,444 551,355 1,006,736

Table 24: Ordered Logit Estimates (with Returns) for all Datasets (10 ordered values)

In general (except for Ordered Probit estimates for 10 ordered values), the
results for o
1
and o
2
values are consistent with the results in the previous section.
Even when taking Returns into account, B&M’s theory still holds. The results
for o
3
values show that a significant positive relationship also exists between the
size of the bid-ask spread and the direction of the exchange rate movement.


53
One interesting result to note are the negative o
1
and o
2
values for the ordered
probit analyses of 2006 and 2009 for 10 ordered indicator values (see Table 23).
Not only does this violate B&M’s prediction, the observation that the o
1
values
appear to decrease with increasing tick volume runs counter to the rest of the
ordered response results. One possible explanation for this is that multiplicative
heteroskedasticity, which our ordered response models omitted but has been
shown to be significant by B&M, is likely to be the cause of this results.

Also since we know that the error term of equations 13 and 14 are not i.d.d.
standard normal, the results for the ordered logit analyses are probably more
meaningful.

3.6 Summary
B&M provided the theoretical framework and showed empirical evidence that
the size of the bid-ask spread of exchange rates is positively related to the
exchange rate’s underlying volatility. Their empirical subject was over 300,000
USD/DEM quotes from April to June of 1989. The USD/DEM was obviously
one of the most traded currency pair during that time, and both the USD and
DEM belonged to floating exchange rates regimes of two of the top three largest
world economies in 1989.


54
We then wanted to see if both theory and empirical results would hold for a
much lesser traded subject currency, if the currency belonged to a small country
and from a semi-floating exchange rate regime – the USD/SGD.

We applied a modified version of B&M’s methodology on the USD/SGD from
the same months of April to June 1989 and found that the evidence (as
documented in the previous section) supported their theory. Furthermore, we
repeated the analysis on the USD/SGD 17 years later in 2006 when the world
economy was at a point of stability, and again in 2009 when the world economy
was somewhat in disarray. We found that B&M’s theory was still supported by
the evidence.

Along the way we also noted the interesting phenomenon regarding the behavior
of bid-ask spreads of foreign exchange rates. When prices move up, and the
bid-ask spreads change, they tend to widen. Conversely, when prices move
down, and the bid-ask spreads change, they tend to narrow. We repeated the
analyses on the datasets and included Returns as additional variable in the
ordered response models. We found that the relationship between bid-ask
spreads and price direction is positive and significant.



55
4 Conclusion

This paper has set out to empirically test B&M’s theory against a lesser-traded
currency from a developing country with a managed floating rate regime. The
results generally hold, and we found that the effects of the tested parameters
grew in strength as that currency grew in economic significance in its region.
We departed from B&M’s methodology by omitting multiplicative
heteroskedasticity, but we tested our data using both Ordered Probit and Ordered
Logit models. The generally “better-behaving” results from the Logit models
suggests that the disturbances are indeed not standard normal, and that future
work in this area using the same methodologies should include multiplicative
heteroskedasticity.

We also noted the interesting phenomenon regarding the behavior of bid-ask
spreads of foreign exchange rates. We included returns as additional variable in
the ordered response models and found that the relationship between bid-ask
spreads and price direction is positive and significant. Future work in this area
should attempt to build a model which links bid-ask spreads not only to
volatility and also returns.

We provided a simple survey on microstructure literature. We first touched on
those which analyses bid-ask spreads and asymmetric information, and also
56
presented snapshots of research involving bid-ask spreads and volatility. Bid-ask
spreads, one of two “hallmarks” of the microstructure approach, remain a
popular theme for research today.

57
References

Admanti, A., and Pfleiderer, P., (1988), “A Theory of Intraday Patterns: Volume and
Price Variability,” Review of Financial Studies, 1, pp. 3-40

Admanti, A.R. and Pfleiderer, P., (1989), “Divide and Conquer: A Theory of
Intra-day and Day of the Week Effects,” Review of Financial Studies, 2, pp.
189-223

Aliber, R.Z., (1975), “Monetary Independence under Floating Exchange Rates,”
Journal of Finance, 30, pp. 365-376

Amihud, Y. and Mendelson, H. (1980), “Dealership Market: Market-making with
Inventory”, Journal of Financial Economics, pp. 31-53

Bagehot, W. (1971), “The Only Game in Town,” Financial Analysts Journal, 22, pp.
12-14

Barnea, A. and Logue, D.E. (1975), “The Effect of Risk on the Market Maker’s
Spread”, Financial Analysts Journal, 31, pp. 45-49

Black, S.W., (1991), “Transactions Costs and Vehicle Currencies,” Journal of
International Money and Finance, 10, pp. 512-526

Bollerslev, T., Chou, R.Y., and Kroner, K.F., (1992), “ARCH Modeling in Finance:
A Review of the Theory and Empirical Evidence,” Journal of Econometrics, 52,
pp. 5-59

Bollerslev, T., and Domowitz, I., (1993), “Trading Patterns and Prices in the
Interbank Foreign Exchange Market,” Journal of Finance, 48, pp. 1421-1443

Bollerslev, T. and Melvin, M. (1994), “Bid-ask Spreads and Volatility in the Foreign
Exchange Market – an Empirical Analysis,” Journal of International Economics,
36, pp. 355-372

Boothe, P., (1988), “Exchange Rate Risk and the Bid-Ask Spread: A Seven Country
Comparison,” Economic Inquiry, 26, pp. 485-492

Bossaerts, P. and Hillion, P., (1991), “Market Microstructure Effects of Government
Intervention in the Foreign Exchange Market,” Review of Financial Studies, 4,
pp. 513-541

Copeland, T. and Galai, D. (1983), “Information Effects on the Bid-Ask Spread,”
58
Journal of Finance, 38, pp. 1457-1469

Demsetz, H (1968), “The Cost of Transacting”, Quarterly Journal of Economics, 82,
pp. 33-53

Evans, M. and Lyons, R. (2007), “Exchange Rate Fundamentals and Order Flow,”
NBER Working Papers 13151, National Bureau of Economic Research, Inc.

Fieleke, N.S., (1975), “Exchange Rate Flexibility and the Efficiency of Foreign
Exchange Markets,” Journal of Financial and Quantitative Analysis, 10, pp.
409-428

Glassman, D.A., (1987), “Exchange Rate Risks and Transaction Costs: Evidence
from Bid-Ask Spreads,” Journal of International Money and Finance, 6, pp.
479-490

Glosten, L.R. and Milgrom, P.R. (1985), “Bid, Ask and Transaction Prices in a
Specialist Market with Heterogeneously Informed Traders,” Journal of
Financial Economics, 14, pp. 71-100

Goodhart, C.A.E., and Figliuoli, L., (1991), “Every Minute Counts in Financial
Markets,” Journal of International Money and Finance, 10, pp. 23-52

Goodhart, C.A.E., (1990), “News and the Foreign Exchange market,” FMG
Discussion Papers dp71, Financial Markets Group

Green, R., Hollified, B. and Schurhoff, N. (2005), “Financial Intermediation And
The Costs Of Trading In An Opaque Market,” FAME Research Paper Series
rp130, International Center for Financial Asset Management and Engineering

Hasbrouck, J. and Sofianos, G.. (1993), “The Trades Of Market Makers: An
Empirical Analysis Of NYSE Specialists,” Journal of Finance, 48, pp.
1565-1593.

Ho, T. and Stoll, H. (1981), “Optimal Dealer Pricing under Transactions and Return
Uncertainty,” Journal of Financial Economics, 9, pp. 757-84

Leach, C.J., and Madhavan, A.N., (1992), “Intertemporal Price Discovery By
Market Makers: Active Versus Passive Learning,” Journal of Financial
Intermediation, 2, pp. 207-235

Leach, C.J., and Madhavan, A.N., (1993), “Price Experimentation And Security
Market Structure”, Review of Financial Studies, 6, pp. 375-404

59
Lo, A.W. and Mackinlay, A.C., (1990), “An Econometric Analysis of
Non-Synchronous Trading,” Journal of Econometrics, 45, pp. 181-210

Lo, A.W., and MacKinlay, A.C., (1992), “An Ordered Probit Analysis of
Transaction Stock Prices,” Journal of Financial Economics, 31, pp. 319-379

Lyons, R. (1995), “Tests of Microstructural Hypotheses in the Foreign Exchange
Market”, Journal of Financial Economics, 39, pp. 321-351

Lyons, R. (2001) The Microstructure Approach to Exchange Rates, The MIT Press,
Cambridge, MA

Melvin, M. and Tan, K.H., (1996), “Foreign Exchange Market Bid-Ask Spreads and
the Market Price of Social Unrest,” Oxford Economic Papers, Oxford
University Press, 48(2), pp. 329-41

Monetary Authority of Singapore Annual Report 2005/2006

Monetary Authority of Singapore Annual Report 2008/2009

Naik, N.Y., Neuberger, A. and Viswanathan, S., (1999), “Trade Disclosure
Regulation In Markets With Negotiated Trades,” Review of Financial Studies,
12, pp. 873-900

Nelson, D.B., (1990), “ARCH Models as Diffusion Approximations,” Journal of
Econometrics, 45, pp. 7-38

Nelson, D.B., (1992), “Filtering and Forecasting with Misspecified ARCH Models I:
Getting the Right Variance with the Wrong Model,” Journal of Econometrics,
52, pp. 61-90

Osler, C., Mende, A. and Menkhoff, L. (2006), “Price Discovery in Currency
Markets,” Diskussionspapiere der Wirtschaftswissenschaftlichen Fakultät der
Universität Hannover dp-351, Universität Hannover,
Wirtschaftswissenschaftliche Fakultät.

Overturf, S.F., (1982), “Risk, Transaction Charges, and the Market for Foreign
Exchange Services,” Economic Inquiry, 20, pp. 291-302

Payne, R. (2003), “Informed Trade In Spot Foreign Exchange Markets: An
Empirical Investigation,” Journal of International Economics, 61, pp. 307-329

Sarno, L. and Taylor M. (2002) The Exconomics of Exchange Rates Cambridge
University Press, Cambridge, UK
60
Appendix 1 – EViews Results

A.1.1 Dataset 1: USD-DEM April 1989 to June 1989:
A.1.1.1 MA(1)-GARCH(1,1) Results:

Dependent Variable: 10000*DLOG(ASK)
Method: ML - ARCH (Marquardt) - Normal distribution
Sample (adjusted): 2 304131
Included observations: 304130 after adjustments
Convergence achieved after 86 iterations
MA Backcast: 1
Presample variance: backcast (parameter = 0.7)
GARCH = C(3) + C(4)*RESID(-1)^2 + C(5)*GARCH(-1)
Variable Coefficient Std. Error z-Statistic Prob.
C 0.007732 0.001186 6.518723 0
MA(1) -0.586985 0.00129 -455.0911 0
Variance Equation
C 0.04555 0.000523 87.1623 0
RESID(-1)^2 0.055817 0.00029 192.773 0
GARCH(-1) 0.934608 0.000278 3358.51 0
R-squared 0.237055 Mean dependent var 0.00087
Adjusted R-squared 0.237053 S.D. dependent var 2.379553
S.E. of regression 2.078464 Akaike info criterion 4.102756
Sum squared resid 1313837 Schwarz criterion 4.102931
Log likelihood -623880.6 Hannan-Quinn criter. 4.102807
Durbin-Watson stat 1.942352
Inverted MA Roots 0.59


61
A.1.1.2 Ordered Probit Results: 4 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Probit (Quadratic hill climbing)
Sample (adjusted): 2 304131
Included observations: 304130 after adjustments
Number of ordered indicator values: 4
Convergence achieved after 4 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.021372 0.000289 73.96036 0
AJ(-1) 0.311626 0.002302 135.3568 0
Limit Points
LIMIT_2:C(3) 0.190472 0.005775 32.98297 0
LIMIT_3:C(4) 0.557843 0.005798 96.21363 0
LIMIT_4:C(5) 2.633627 0.007509 350.7163 0
Pseudo R-squared 0.041442 Akaike info criterion 2.077312
Schwarz criterion 2.077487 Log likelihood -315881.5
Hannan-Quinn criter. 2.077363 Restr. log likelihood -329538.1
LR statistic 27313.29 Avg. log likelihood -1.03864
Prob(LR statistic) 0


62
A.1.1.3 Ordered Probit (with Returns) Results: 4 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Probit (Quadratic hill climbing)
Sample (adjusted): 2 304131
Included observations: 304130 after adjustments
Number of ordered indicator values: 4
Failure to improve Likelihood after 12 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.014821 0.000288 51.46485 0
AJ(-1) 0.038314 0.002355 16.26616 0
10000*DLOG(ASK) 0.096973 0.000986 98.34188 0
Limit Points
LIMIT_2:C(4) -0.569596 0.005843 -97.47589 0
LIMIT_3:C(5) -0.204094 0.005933 -34.39776 0
LIMIT_4:C(6) 1.828335 0.007386 247.5462 0
Pseudo R-squared 0.041202 Akaike info criterion 2.077839
Schwarz criterion 2.078049 Log likelihood -315960.6
Hannan-Quinn criter. 2.0779 Restr. log likelihood -329538.1
LR statistic 27155.04 Avg. log likelihood -1.0389
Prob(LR statistic) 0


63
A.1.1.4 Ordered Probit Results: 10 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Probit (Quadratic hill climbing)
Sample (adjusted): 2 308749
Included observations: 307943 after adjustments
Number of ordered indicator values: 10
Convergence achieved after 7 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.034469 0.000382 90.17413 0
AJ(-1) 0.112367 0.000917 122.5758 0
Limit Points
LIMIT_2:C(3) -2.206482 0.01721 -128.2094 0
LIMIT_3:C(4) -1.756869 0.010998 -159.7475 0
LIMIT_4:C(5) -1.538779 0.00944 -162.9984 0
LIMIT_5:C(6) 0.322974 0.006894 46.85067 0
LIMIT_6:C(7) 0.329696 0.006894 47.8258 0
LIMIT_7:C(8) 0.666276 0.006924 96.2254 0
LIMIT_8:C(9) 0.693844 0.006929 100.1396 0
LIMIT_9:C(10) 0.694678 0.006929 100.2579 0
LIMIT_10:C(11) 2.788901 0.008505 327.928 0
Pseudo R-squared 0.037526 Akaike info criterion 2.258734
Schwarz criterion 2.259114 Log likelihood -347769.7
Hannan-Quinn criter. 2.258844 Restr. log likelihood -361329
LR statistic 27118.64 Avg. log likelihood -1.129331
Prob(LR statistic) 0


64
A.1.1.5 Ordered Probit (with Returns) Results: 10 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Probit (Quadratic hill climbing)
Sample (adjusted): 2 308749
Included observations: 307943 after adjustments
Number of ordered indicator values: 10
Failure to improve Likelihood after 9 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.020522 0.00038 54.02521 0
AJ(-1) 0.004221 0.000938 4.501637 0
10000*DLOG(ASK) 0.096148 0.00091 105.6836 0
Limit Points
LIMIT_2:C(4) -3.04669 0.017039 -178.812 0
LIMIT_3:C(5) -2.605993 0.011008 -236.7287 0
LIMIT_4:C(6) -2.393098 0.00953 -251.1147 0
LIMIT_5:C(7) -0.605194 0.006997 -86.49427 0
LIMIT_6:C(8) -0.598756 0.006998 -85.56097 0
LIMIT_7:C(9) -0.277872 0.007077 -39.26293 0
LIMIT_8:C(10) -0.251959 0.007086 -35.55753 0
LIMIT_9:C(11) -0.251127 0.007086 -35.43857 0
LIMIT_10:C(12) 1.820135 0.008402 216.6283 0
Pseudo R-squared 0.034829 Akaike info criterion 2.26507
Schwarz criterion 2.265485 Log likelihood -348744.3
Hannan-Quinn criter. 2.26519 Restr. log likelihood -361329
LR statistic 25169.41 Avg. log likelihood -1.132496
Prob(LR statistic) 0

65
A.1.1.6 Ordered Logit Results: 4 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Logit (Quadratic hill climbing)
Sample (adjusted): 2 304131
Included observations: 304130 after adjustments
Number of ordered indicator values: 4
Convergence achieved after 4 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.129908 0.001266 102.6436 0
AJ(-1) 0.432993 0.003998 108.2955 0
Limit Points
LIMIT_2:C(3) 0.483887 0.009847 49.14077 0
LIMIT_3:C(4) 1.100335 0.009948 110.6037 0
LIMIT_4:C(5) 4.89648 0.014927 328.0225 0
Pseudo R-squared 0.051821 Akaike info criterion 2.05482
Schwarz criterion 2.054994 Log likelihood -312461.1
Hannan-Quinn criter. 2.05487 Restr. log likelihood -329538.1
LR statistic 34153.96 Avg. log likelihood -1.027393
Prob(LR statistic) 0


66
A.1.1.7 Ordered Logit (with Returns) Results: 4 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Logit (Quadratic hill climbing)
Sample (adjusted): 3 304131
Included observations: 304128 after adjustments
Number of ordered indicator values: 4
Convergence achieved after 5 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.131878 0.001297 101.683 0
AJ(-1) 0.673158 0.004354 154.6104 0
10000*DLOG(ASK) 0.320523 0.001841 174.0656 0
Limit Points
LIMIT_2:C(4) 0.955897 0.010468 91.31565 0
LIMIT_3:C(5) 1.626336 0.010676 152.3301 0
LIMIT_4:C(6) 5.78533 0.016938 341.5674 0
Pseudo R-squared 0.102908 Akaike info criterion 1.944117
Schwarz criterion 1.944327 Log likelihood -295624.3
Hannan-Quinn criter. 1.944178 Restr. log likelihood -329536.2
LR statistic 67823.94 Avg. log likelihood -0.972039
Prob(LR statistic) 0

67
A.1.1.8 Ordered Logit Results: 10 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Logit (Quadratic hill climbing)
Sample (adjusted): 2 308749
Included observations: 307943 after adjustments
Number of ordered indicator values: 10
Convergence achieved after 7 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.155742 0.001351 115.2591 0
AJ(-1) 0.159926 0.001594 100.3382 0
Limit Points
LIMIT_2:C(3) -5.137085 0.053891 -95.32363 0
LIMIT_3:C(4) -3.762068 0.028652 -131.3013 0
LIMIT_4:C(5) -3.158783 0.022402 -141.0075 0
LIMIT_5:C(6) 0.717831 0.011771 60.98525 0
LIMIT_6:C(7) 0.729267 0.011771 61.95536 0
LIMIT_7:C(8) 1.29396 0.011879 108.9279 0
LIMIT_8:C(9) 1.339747 0.011894 112.6399 0
LIMIT_9:C(10) 1.341132 0.011895 112.7518 0
LIMIT_10:C(11) 5.194799 0.016561 313.6803 0
Pseudo R-squared 0.04758 Akaike info criterion 2.23514
Schwarz criterion 2.23552 Log likelihood -344136.9
Hannan-Quinn criter. 2.23525 Restr. log likelihood -361329
LR statistic 34384.2 Avg. log likelihood -1.117534
Prob(LR statistic) 0



68
A.1.1.9 Ordered Logit (with Returns) Results: 10 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Logit (Quadratic hill climbing)
Sample (adjusted): 2 308749
Included observations: 307943 after adjustments
Number of ordered indicator values: 10
Convergence achieved after 7 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.15917 0.001399 113.7732 0
AJ(-1) 0.253184 0.00173 146.3672 0
10000*DLOG(ASK) 0.318154 0.001836 173.2477 0
Limit Points
LIMIT_2:C(4) -4.663842 0.053994 -86.37753 0
LIMIT_3:C(5) -3.286633 0.028848 -113.9311 0
LIMIT_4:C(6) -2.681408 0.022654 -118.3644 0
LIMIT_5:C(7) 1.308867 0.012554 104.2614 0
LIMIT_6:C(8) 1.321246 0.012556 105.2298 0
LIMIT_7:C(9) 1.93451 0.012762 151.5792 0
LIMIT_8:C(10) 1.984385 0.012786 155.2056 0
LIMIT_9:C(11) 1.985893 0.012786 155.3148 0
LIMIT_10:C(12) 6.194035 0.018747 330.3986 0
Pseudo R-squared 0.093349 Akaike info criterion 2.127739
Schwarz criterion 2.128154 Log likelihood -327599.2
Hannan-Quinn criter. 2.127859 Restr. log likelihood -361329
LR statistic 67459.66 Avg. log likelihood -1.063831
Prob(LR statistic) 0

69
A.1.2 Dataset 2: USD-USD April 1989 to June 1989:
A.1.2.1 MA(1)-GARCH(1,1) Results:

Dependent Variable: 10000*DLOG(ASK)
Method: ML - ARCH
Sample (adjusted): 2 8445
Included observations: 8444 after adjustments
Convergence achieved after 37 iterations
MA Backcast: 1
Presample variance: backcast (parameter = 0.7)
GARCH = C(3) + C(4)*RESID(-1)^2 + C(5)*GARCH(-1)
Variable Coefficient Std. Error z-Statistic Prob.
C 0.045205 0.013771 3.282516 0.001
MA(1) -0.3281 0.011435 -28.693 0
Variance Equation
C 0.683349 0.025552 26.7437 0
RESID(-1)^2 0.265029 0.008055 32.9008 0
GARCH(-1) 0.653991 0.007513 87.05354 0
R-squared 0.046483 Mean dependent var 0.001809
Adjusted R-squared 0.04637 S.D. dependent var 2.611027
S.E. of regression 2.549771 Akaike info criterion 4.444882
Sum squared resid 54884.24 Schwarz criterion 4.449051
Log likelihood -18761.29 Hannan-Quinn criter. 4.446305
Durbin-Watson stat 1.819609
Inverted MA Roots 0.33


70
A.1.2.2 Ordered Probit Results: 4 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Probit (Quadratic hill climbing)
Sample (adjusted): 2 8445
Included observations: 8444 after adjustments
Number of ordered indicator values: 4
Convergence achieved after 5 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCH01 0.009127 0.000855 10.67187 0
AJ(-1) 0.942999 0.042231 22.3298 0
Limit Points
LIMIT_2:C(3) 0.61897 0.123303 5.019921 0
LIMIT_3:C(4) 1.002232 0.121611 8.241273 0
LIMIT_4:C(5) 4.898957 0.134626 36.38941 0
Pseudo R-squared 0.115395 Akaike info criterion 0.581462
Schwarz criterion 0.585632 Log likelihood -2449.933
Hannan-Quinn criter. 0.582885 Restr. log likelihood -2769.524
LR statistic 639.1806 Avg. log likelihood -0.290139
Prob(LR statistic) 0

71
A.1.2.3 Ordered Probit (with Returns) Results: 4 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Probit (Quadratic hill climbing)
Sample (adjusted): 2 8445
Included observations: 8444 after adjustments
Number of ordered indicator values: 4
Convergence achieved after 5 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.00933 0.000852 10.94538 0
AJ(-1) 1.008975 0.043056 23.43378 0
10000*DLOG(ASK) 0.08268 0.006736 12.27428 0
Limit Points
LIMIT_2:C(4) 0.775747 0.125149 6.198581 0
LIMIT_3:C(5) 1.166192 0.12345 9.446642 0
LIMIT_4:C(6) 5.148438 0.138288 37.22976 0
Pseudo R-squared 0.142475 Akaike info criterion 0.563936
Schwarz criterion 0.568939 Log likelihood -2374.936
Hannan-Quinn criter. 0.565644 Restr. log likelihood -2769.524
LR statistic 789.1746 Avg. log likelihood -0.281257
Prob(LR statistic) 0

72
A.1.2.4 Ordered Probit Results: 10 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Probit (Quadratic hill climbing)
Sample (adjusted): 2 8445
Included observations: 8444 after adjustments
Number of ordered indicator values: 10
Convergence achieved after 7 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.010588 0.000829 12.7664 0
AJ(-1) 0.293255 0.018241 16.07658 0
Limit Points
LIMIT_2:C(3) -1.046148 0.251299 -4.162967 0
LIMIT_3:C(4) -0.925167 0.225398 -4.104586 0
LIMIT_4:C(5) -0.71307 0.195539 -3.646689 0.0003
LIMIT_5:C(6) 0.463308 0.159031 2.91332 0.0036
LIMIT_6:C(7) 0.525466 0.158821 3.308537 0.0009
LIMIT_7:C(8) 0.77686 0.158252 4.908998 0
LIMIT_8:C(9) 0.837742 0.158161 5.296765 0
LIMIT_9:C(10) 0.843638 0.158153 5.334317 0
LIMIT_10:C(11) 4.644992 0.167892 27.66655 0
Pseudo R-squared 0.069425 Akaike info criterion 0.65859
Schwarz criterion 0.667763 Log likelihood -2769.568
Hannan-Quinn criter. 0.661721 Restr. log likelihood -2976.191
LR statistic 413.2452 Avg. log likelihood -0.327992
Prob(LR statistic) 0


73
A.1.2.5 Ordered Probit (with Returns) Results: 10 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Probit (Quadratic hill climbing)
Sample (adjusted): 2 8445
Included observations: 8444 after adjustments
Number of ordered indicator values: 10
Convergence achieved after 7 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.010826 0.000828 13.08258 0
AJ(-1) 0.313473 0.018435 17.00409 0
10000*DLOG(ASK) 0.074509 0.006572 11.33749 0
Limit Points
LIMIT_2:C(4) -0.920766 0.2527 -3.643716 0.0003
LIMIT_3:C(5) -0.799597 0.226892 -3.524124 0.0004
LIMIT_4:C(6) -0.586622 0.197098 -2.976292 0.0029
LIMIT_5:C(7) 0.609972 0.160333 3.804416 0.0001
LIMIT_6:C(8) 0.673407 0.160125 4.205514 0
LIMIT_7:C(9) 0.92919 0.159566 5.823252 0
LIMIT_8:C(10) 0.991005 0.159477 6.214104 0
LIMIT_9:C(11) 0.996988 0.159469 6.251929 0
LIMIT_10:C(12) 4.866967 0.170386 28.56432 0
Pseudo R-squared 0.09087 Akaike info criterion 0.64371
Schwarz criterion 0.653717 Log likelihood -2705.745
Hannan-Quinn criter. 0.647126 Restr. log likelihood -2976.191
LR statistic 540.8901 Avg. log likelihood -0.320434
Prob(LR statistic) 0


74
A.1.2.6 Ordered Logit Results: 4 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Logit (Quadratic hill climbing)
Sample (adjusted): 2 8445
Included observations: 8444 after adjustments
Number of ordered indicator values: 4
Convergence achieved after 7 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.023145 0.002506 9.234157 0
AJ(-1) 1.816022 0.078415 23.15919 0
Limit Points
LIMIT_2:C(3) 1.151811 0.221143 5.208446 0
LIMIT_3:C(4) 2.052707 0.215518 9.524505 0
LIMIT_4:C(5) 9.285628 0.252771 36.73541 0
Pseudo R-squared 0.11296 Akaike info criterion 0.583059
Schwarz criterion 0.587229 Log likelihood -2456.677
Hannan-Quinn criter. 0.584483 Restr. log likelihood -2769.524
LR statistic 625.6929 Avg. log likelihood -0.290938
Prob(LR statistic) 0


75
A.1.2.7 Ordered Logit (with Returns) Results: 4 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Logit (Quadratic hill climbing)
Sample (adjusted): 2 8445
Included observations: 8444 after adjustments
Number of ordered indicator values: 4
Convergence achieved after 8 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.023498 0.002679 8.772378 0
AJ(-1) 1.937969 0.080338 24.12276 0
10000*DLOG(ASK) 0.179594 0.014658 12.25201 0
Limit Points
LIMIT_2:C(4) 1.431605 0.224685 6.371611 0
LIMIT_3:C(5) 2.341427 0.219477 10.66822 0
LIMIT_4:C(6) 9.770045 0.261472 37.36562 0
Pseudo R-squared 0.139367 Akaike info criterion 0.565974
Schwarz criterion 0.570977 Log likelihood -2383.542
Hannan-Quinn criter. 0.567682 Restr. log likelihood -2769.524
LR statistic 771.9628 Avg. log likelihood -0.282276
Prob(LR statistic) 0


76
A.1.2.8 Ordered Logit Results: 10 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Logit (Quadratic hill climbing)
Sample (adjusted): 2 8445
Included observations: 8444 after adjustments
Number of ordered indicator values: 10
Convergence achieved after 6 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.030535 0.00282 10.82827 0
AJ(-1) 0.563651 0.033692 16.72951 0
Limit Points
LIMIT_2:C(3) -3.399721 0.756259 -4.495444 0
LIMIT_3:C(4) -2.993661 0.63663 -4.702359 0
LIMIT_4:C(5) -2.298438 0.488651 -4.703637 0
LIMIT_5:C(6) 0.926596 0.286528 3.233876 0.0012
LIMIT_6:C(7) 1.076884 0.285436 3.77277 0.0002
LIMIT_7:C(8) 1.66173 0.282811 5.875773 0
LIMIT_8:C(9) 1.797737 0.282485 6.364009 0
LIMIT_9:C(10) 1.810792 0.282458 6.41083 0
LIMIT_10:C(11) 8.845043 0.310773 28.46145 0
Pseudo R-squared 0.068507 Akaike info criterion 0.659237
Schwarz criterion 0.66841 Log likelihood -2772.3
Hannan-Quinn criter. 0.662368 Restr. log likelihood -2976.191
LR statistic 407.7821 Avg. log likelihood -0.328316
Prob(LR statistic) 0

77
A.1.2.9 Ordered Logit (with Returns) Results: 10 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Logit (Quadratic hill climbing)
Sample (adjusted): 2 8445
Included observations: 8444 after adjustments
Number of ordered indicator values: 10
Convergence achieved after 8 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.031698 0.002984 10.62409 0
AJ(-1) 0.605872 0.03419 17.72067 0
10000*DLOG(ASK) 0.167709 0.014728 11.38731 0
Limit Points
LIMIT_2:C(4) -3.108606 0.756454 -4.109446 0
LIMIT_3:C(5) -2.702375 0.63687 -4.243211 0
LIMIT_4:C(6) -2.004677 0.48917 -4.098116 0
LIMIT_5:C(7) 1.236004 0.289368 4.271391 0
LIMIT_6:C(8) 1.387414 0.288368 4.811259 0
LIMIT_7:C(9) 1.976024 0.286018 6.908731 0
LIMIT_8:C(10) 2.112879 0.285744 7.394317 0
LIMIT_9:C(11) 2.126018 0.285722 7.440867 0
LIMIT_10:C(12) 9.336144 0.318108 29.349 0
Pseudo R-squared 0.09058 Akaike info criterion 0.643914
Schwarz criterion 0.653921 Log likelihood -2706.607
Hannan-Quinn criter. 0.64733 Restr. log likelihood -2976.191
LR statistic 539.1676 Avg. log likelihood -0.320536
Prob(LR statistic) 0



78
A.1.3 Dataset 3: USD-USD April 2006 to May 2006:
A.1.3.1 MA(1)-GARCH(1,1) Results:

Dependent Variable: 10000*DLOG(ASK)
Method: ML - ARCH (Marquardt) - Normal distribution
Sample (adjusted): 2 551356
Included observations: 551355 after adjustments
Convergence achieved after 14 iterations
MA Backcast: 1
Presample variance: backcast (parameter = 0.7)
GARCH = C(3) + C(4)*RESID(-1)^2 + C(5)*GARCH(-1)
Variable Coefficient Std. Error z-Statistic Prob.
C 0.002535 0.000358 7.075531 0
MA(1) -0.781719 0.000858 -911.204 0
Variance Equation
C 0.057826 0.000813 71.15628 0
RESID(-1)^2 0.063219 0.000558 113.2785 0
GARCH(-1) 0.902484 0.000889 1015.04 0
R-squared 0.372225 Mean dependent var -0.000425
Adjusted R-squared 0.372224 S.D. dependent var 1.657286
S.E. of regression 1.313106 Akaike info criterion 3.302185
Sum squared resid 950669.3 Schwarz criterion 3.302287
Log likelihood -910333.2 Hannan-Quinn criter. 3.302214
Durbin-Watson stat 1.97678
Inverted MA Roots 0.78


79
A.1.3.2 Ordered Probit Results: 4 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Probit (Quadratic hill climbing)
Sample (adjusted): 2 551356
Included observations: 551355 after adjustments
Number of ordered indicator values: 4
Convergence achieved after 4 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.046421 0.001588 29.24124 0
AJ(-1) -0.130554 0.001879 -69.46221 0
Limit Points
LIMIT_2:C(3) -0.267391 0.004557 -58.67524 0
LIMIT_3:C(4) 0.497267 0.004577 108.652 0
LIMIT_4:C(5) 3.624999 0.036661 98.87791 0
Pseudo R-squared 0.00466 Akaike info criterion 2.122714
Schwarz criterion 2.122815 Log likelihood -585179.4
Hannan-Quinn criter. 2.122742 Restr. log likelihood -587919.3
LR statistic 5479.851 Avg. log likelihood -1.061348
Prob(LR statistic) 0


80
A.1.3.3 Ordered Probit (with Returns) Results: 4 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Probit (Quadratic hill climbing)
Sample (adjusted): 2 551356
Included observations: 551355 after adjustments
Number of ordered indicator values: 4
Convergence achieved after 6 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.032475 0.001698 19.12882 0
AJ(-1) 0.520271 0.002468 210.7702 0
10000*DLOG(ASK) 0.624067 0.001388 449.5486 0
Limit Points
LIMIT_2:C(4) 0.803072 0.005446 147.4683 0
LIMIT_3:C(5) 1.84724 0.00582 317.4066 0
LIMIT_4:C(6) 6.264271 0.051481 121.682 0
Pseudo R-squared 0.202812 Akaike info criterion 1.700133
Schwarz criterion 1.700255 Log likelihood -468682.4
Hannan-Quinn criter. 1.700167 Restr. log likelihood -587919.3
LR statistic 238473.8 Avg. log likelihood -0.850056
Prob(LR statistic) 0


81
A.1.3.4 Ordered Probit Results: 10 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Probit (Quadratic hill climbing)
Sample (adjusted): 2 551356
Included observations: 530382 after adjustments
Number of ordered indicator values: 10
Convergence achieved after 5 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.029623 0.001543 19.20169 0
AJ(-1) -0.040361 0.000655 -61.5938 0
Limit Points
LIMIT_2:C(3) -2.770052 0.008137 -340.4275 0
LIMIT_3:C(4) -2.123895 0.005776 -367.693 0
LIMIT_4:C(5) -1.541017 0.005154 -299.0033 0
LIMIT_5:C(6) -0.292156 0.004841 -60.34619 0
LIMIT_6:C(7) -0.076775 0.004831 -15.89255 0
LIMIT_7:C(8) 0.218854 0.004827 45.33503 0
LIMIT_8:C(9) 0.463249 0.00484 95.70278 0
LIMIT_9:C(10) 0.472672 0.004841 97.63469 0
LIMIT_10:C(11) 3.593189 0.037289 96.36043 0
Pseudo R-squared 0.00228 Akaike info criterion 3.351617
Schwarz criterion 3.351849 Log likelihood -888807.6
Hannan-Quinn criter. 3.351682 Restr. log likelihood -890838.9
LR statistic 4062.696 Avg. log likelihood -1.675788
Prob(LR statistic) 0


82
A.1.3.5 Ordered Probit (with Returns) Results: 10 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Probit (Quadratic hill climbing)
Sample (adjusted): 2 551356
Included observations: 530382 after adjustments
Number of ordered indicator values: 10
Failure to improve Likelihood after 17 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR -0.035043 0.001544 -22.69168 0
AJ(-1) 0.01852 0.000821 22.56212 0
10000*DLOG(ASK) 0.365238 0.001223 298.6273 0
Limit Points
LIMIT_2:C(4) -2.768453 0.008501 -325.6782 0
LIMIT_3:C(5) -2.089688 0.006173 -338.5115 0
LIMIT_4:C(6) -1.484632 0.00565 -262.7437 0
LIMIT_5:C(7) -0.148442 0.005549 -26.74997 0
LIMIT_6:C(8) 0.084178 0.005598 15.03649 0
LIMIT_7:C(9) 0.395925 0.005688 69.60886 0
LIMIT_8:C(10) 0.648621 0.005789 112.0419 0
LIMIT_9:C(11) 0.659757 0.005795 113.8488 0
LIMIT_10:C(12) 4.041679 0.025918 155.9393 0
Pseudo R-squared 0.101511 Akaike info criterion 3.018282
Schwarz criterion 3.018535 Log likelihood -800409.1
Hannan-Quinn criter. 3.018353 Restr. log likelihood -890838.9
LR statistic 180859.6 Avg. log likelihood -1.509118
Prob(LR statistic) 0

83
A.1.3.6 Ordered Logit Results: 4 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Logit (Quadratic hill climbing)
Sample (adjusted): 2 551356
Included observations: 551355 after adjustments
Number of ordered indicator values: 4
Convergence achieved after 4 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.080152 0.00277 28.93061 0
AJ(-1) -0.204018 0.003055 -66.79117 0
Limit Points
LIMIT_2:C(3) -0.41091 0.007581 -54.20336 0
LIMIT_3:C(4) 0.830244 0.007658 108.4198 0
LIMIT_4:C(5) 9.135262 0.144521 63.21045 0
Pseudo R-squared 0.004388 Akaike info criterion 2.123295
Schwarz criterion 2.123397 Log likelihood -585339.7
Hannan-Quinn criter. 2.123324 Restr. log likelihood -587919.3
LR statistic 5159.321 Avg. log likelihood -1.061638
Prob(LR statistic) 0


84
A.1.3.7 Ordered Logit (with Returns) Results: 4 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Logit (Quadratic hill climbing)
Sample (adjusted): 2 551356
Included observations: 551355 after adjustments
Number of ordered indicator values: 4
Convergence achieved after 6 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.048475 0.003264 14.85304 0
AJ(-1) 1.053246 0.004586 229.6602 0
10000*DLOG(ASK) 1.215211 0.002966 409.7483 0
Limit Points
LIMIT_2:C(4) 1.600267 0.009958 160.7016 0
LIMIT_3:C(5) 3.453473 0.011048 312.5789 0
LIMIT_4:C(6) 13.6294 0.167398 81.41909 0
Pseudo R-squared 0.216154 Akaike info criterion 1.671678
Schwarz criterion 1.671801 Log likelihood -460838.1
Hannan-Quinn criter. 1.671713 Restr. log likelihood -587919.3
LR statistic 254162.3 Avg. log likelihood -0.835828
Prob(LR statistic) 0



85
A.1.3.8 Ordered Logit Results: 10 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Logit (Quadratic hill climbing)
Sample (adjusted): 2 551356
Included observations: 530382 after adjustments
Number of ordered indicator values: 10
Convergence achieved after 5 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.059869 0.002706 22.12459 0
AJ(-1) -0.07549 0.001114 -67.78632 0
Limit Points
LIMIT_2:C(3) -5.681481 0.021378 -265.7677 0
LIMIT_3:C(4) -3.965107 0.011679 -339.4939 0
LIMIT_4:C(5) -2.690845 0.009268 -290.3316 0
LIMIT_5:C(6) -0.512979 0.008257 -62.12349 0
LIMIT_6:C(7) -0.168038 0.008229 -20.41918 0
LIMIT_7:C(8) 0.3095 0.008227 37.61812 0
LIMIT_8:C(9) 0.714603 0.008278 86.32978 0
LIMIT_9:C(10) 0.73051 0.008281 88.22038 0
LIMIT_10:C(11) 9.036028 0.147643 61.20197 0
Pseudo R-squared 0.0028 Akaike info criterion 3.34987
Schwarz criterion 3.350102 Log likelihood -888344.5
Hannan-Quinn criter. 3.349936 Restr. log likelihood -890838.9
LR statistic 4988.841 Avg. log likelihood -1.674915
Prob(LR statistic) 0

86
A.1.3.9 Ordered Logit (with Returns) Results: 10 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Logit (Quadratic hill climbing)
Sample (adjusted): 2 551356
Included observations: 530382 after adjustments
Number of ordered indicator values: 10
Convergence achieved after 6 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.067262 0.003175 21.18593 0
AJ(-1) 0.402995 0.0016 251.9007 0
10000*DLOG(ASK) 1.189108 0.002737 434.4841 0
Limit Points
LIMIT_2:C(4) -3.894364 0.022026 -176.8036 0
LIMIT_3:C(5) -2.114505 0.012808 -165.0968 0
LIMIT_4:C(6) -0.7165 0.010767 -66.54475 0
LIMIT_5:C(7) 2.04924 0.010738 190.8341 0
LIMIT_6:C(8) 2.528082 0.010914 231.6353 0
LIMIT_7:C(9) 3.225475 0.011271 286.1774 0
LIMIT_8:C(10) 3.846774 0.011692 329.0122 0
LIMIT_9:C(11) 3.872262 0.011712 330.6322 0
LIMIT_10:C(12) 13.94908 0.170499 81.81305 0
Pseudo R-squared 0.146499 Akaike info criterion 2.867156
Schwarz criterion 2.867409 Log likelihood -760331.9
Hannan-Quinn criter. 2.867227 Restr. log likelihood -890838.9
LR statistic 261014.1 Avg. log likelihood -1.433555
Prob(LR statistic) 0





87
A.1.4 Dataset 4: USD-USD April 2009 to May 2009:
A.1.4.1 MA(1)-GARCH(1,1) Results:

Dependent Variable: 10000*DLOG(ASK)
Method: ML - ARCH
Sample (adjusted): 2 1006737
Included observations: 1006736 after adjustments
Convergence achieved after 21 iterations
MA Backcast: 1
Presample variance: backcast (parameter = 0.7)
GARCH = C(3) + C(4)*RESID(-1)^2 + C(5)*GARCH(-1)
Variable Coefficient Std. Error z-Statistic Prob.
C 0.002329 0.000251 9.274136 0
MA(1) -0.78024 0.000611 -1276.001 0
Variance Equation
C 0.045284 0.000225 201.6421 0
RESID(-1)^2 0.051345 0.00021 244.0686 0
GARCH(-1) 0.919658 0.000263 3502.001 0
R-squared 0.36927 Mean dependent var -0.000386
Adjusted R-squared 0.369269 S.D. dependent var 1.602816
S.E. of regression 1.272934 Akaike info criterion 3.196957
Sum squared resid 1631271 Schwarz criterion 3.197016
Log likelihood -1609241 Hannan-Quinn criter. 3.196973
Durbin-Watson stat 1.970126
Inverted MA Roots 0.78


88
A.1.4.2 Ordered Probit Results: 4 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Probit (Quadratic hill climbing)
Sample (adjusted): 2 1006737
Included observations: 1006736 after adjustments
Number of ordered indicator values: 4
Convergence achieved after 3 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.098754 0.000867 113.8863 0
AJ(-1) 0.262245 0.00148 177.1729 0
Limit Points
LIMIT_2:C(3) 0.300383 0.003019 99.50314 0
LIMIT_3:C(4) 1.797373 0.003373 532.8891 0
LIMIT_4:C(5) 2.392075 0.003801 629.366 0
Pseudo R-squared 0.024794 Akaike info criterion 2.069774
Schwarz criterion 2.069833 Log likelihood -1041853
Hannan-Quinn criter. 2.069791 Restr. log likelihood -1068342
LR statistic 52977.45 Avg. log likelihood -1.034882
Prob(LR statistic) 0


89
A.1.4.3 Ordered Probit (with Returns) Results: 4 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Probit (Quadratic hill climbing)
Sample (adjusted): 2 1006737
Included observations: 1006736 after adjustments
Number of ordered indicator values: 4
Convergence achieved after 12 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.040619 0.000907 44.7718 0
AJ(-1) 0.260624 0.001785 146.0215 0
10000*DLOG(ASK) 0.415817 0.00095 437.6876 0
Limit Points
LIMIT_2:C(4) 0.072632 0.003441 21.10602 0
LIMIT_3:C(5) 1.741812 0.004059 429.1621 0
LIMIT_4:C(6) 2.420409 0.004675 517.7516 0
Pseudo R-squared 0.176356 Akaike info criterion 1.748103
Schwarz criterion 1.748173 Log likelihood -879933.1
Hannan-Quinn criter. 1.748122 Restr. log likelihood -1068342
LR statistic 376817.7 Avg. log likelihood -0.874046
Prob(LR statistic) 0


90
A.1.4.4 Ordered Probit Results: 10 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Probit (Quadratic hill climbing)
Sample (adjusted): 2 1006737
Included observations: 1006736 after adjustments
Number of ordered indicator values: 10
Convergence achieved after 4 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.096416 0.000823 117.0911 0
AJ(-1) 0.088372 0.000415 212.8967 0
Limit Points
LIMIT_2:C(3) -1.466867 0.003573 -410.4852 0
LIMIT_3:C(4) -0.169471 0.002717 -62.38151 0
LIMIT_4:C(5) 0.142757 0.00271 52.67088 0
LIMIT_5:C(6) 0.32063 0.002718 117.9753 0
LIMIT_6:C(7) 0.463947 0.002729 170.0266 0
LIMIT_7:C(8) 0.855945 0.002774 308.5626 0
LIMIT_8:C(9) 1.630826 0.002965 549.9868 0
LIMIT_9:C(10) 1.799457 0.003034 593.0574 0
LIMIT_10:C(11) 2.394888 0.003495 685.2099 0
Pseudo R-squared 0.016257 Akaike info criterion 4.058805
Schwarz criterion 4.058934 Log likelihood -2043062
Hannan-Quinn criter. 4.05884 Restr. log likelihood -2076824
LR statistic 67525.47 Avg. log likelihood -2.029392
Prob(LR statistic) 0

91
A.1.4.5 Ordered Probit (with Returns) Results: 10 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Probit (Quadratic hill climbing)
Sample (adjusted): 2 1006737
Included observations: 1006736 after adjustments
Number of ordered indicator values: 10
Convergence achieved after 9 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR -0.003187 0.000829 -3.843887 0.0001
AJ(-1) 0.094896 0.000502 188.9157 0
10000*DLOG(ASK) 0.399614 0.000853 468.5108 0
Limit Points
LIMIT_2:C(4) -1.889585 0.00377 -501.2568 0
LIMIT_3:C(5) -0.540899 0.002964 -182.4773 0
LIMIT_4:C(6) -0.211211 0.003013 -70.09902 0
LIMIT_5:C(7) -0.014555 0.003062 -4.753915 0
LIMIT_6:C(8) 0.153294 0.003113 49.24206 0
LIMIT_7:C(9) 0.615502 0.003255 189.0877 0
LIMIT_8:C(10) 1.481455 0.003579 413.8783 0
LIMIT_9:C(11) 1.676514 0.003697 453.4845 0
LIMIT_10:C(12) 2.40264 0.004378 548.8559 0
Pseudo R-squared 0.096012 Akaike info criterion 3.729748
Schwarz criterion 3.729889 Log likelihood -1877424
Hannan-Quinn criter. 3.729787 Restr. log likelihood -2076824
LR statistic 398801.1 Avg. log likelihood -1.864862
Prob(LR statistic) 0

92
A.1.4.6 Ordered Logit Results: 4 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Logit (Quadratic hill climbing)
Sample (adjusted): 2 1006737
Included observations: 1006736 after adjustments
Number of ordered indicator values: 4
Convergence achieved after 4 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.214626 0.001959 109.574 0
AJ(-1) 0.43948 0.002593 169.4684 0
Limit Points
LIMIT_2:C(3) 0.593662 0.00533 111.3912 0
LIMIT_3:C(4) 3.114943 0.006332 491.9238 0
LIMIT_4:C(5) 4.301442 0.007555 569.3761 0
Pseudo R-squared 0.024601 Akaike info criterion 2.070184
Schwarz criterion 2.070243 Log likelihood -1042059
Hannan-Quinn criter. 2.0702 Restr. log likelihood -1068342
LR statistic 52565.01 Avg. log likelihood -1.035087
Prob(LR statistic) 0

93
A.1.4.7 Ordered Logit (with Returns) Results: 4 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Logit (Quadratic hill climbing)
Sample (adjusted): 2 1006737
Included observations: 1006736 after adjustments
Number of ordered indicator values: 4
Convergence achieved after 5 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.222491 0.002315 96.11273 0
AJ(-1) 1.964362 0.003971 494.6543 0
10000*DLOG(ASK) 1.38648 0.002369 585.1713 0
Limit Points
LIMIT_2:C(4) 2.992082 0.007366 406.2264 0
LIMIT_3:C(5) 6.796605 0.010304 659.6204 0
LIMIT_4:C(6) 8.504201 0.012206 696.7327 0
Pseudo R-squared 0.27793 Akaike info criterion 1.532523
Schwarz criterion 1.532594 Log likelihood -771417.2
Hannan-Quinn criter. 1.532543 Restr. log likelihood -1068342
LR statistic 593849.5 Avg. log likelihood -0.766256
Prob(LR statistic) 0


94
A.1.4.8 Ordered Logit Results: 10 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Logit (Quadratic hill climbing)
Sample (adjusted): 2 1006737
Included observations: 1006736 after adjustments
Number of ordered indicator values: 10
Convergence achieved after 4 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.204722 0.001844 110.9948 0
AJ(-1) 0.149279 0.000728 205.0571 0
Limit Points
LIMIT_2:C(3) -2.795295 0.007967 -350.8518 0
LIMIT_3:C(4) -0.187268 0.004866 -38.48186 0
LIMIT_4:C(5) 0.339075 0.004839 70.07788 0
LIMIT_5:C(6) 0.630813 0.004859 129.8187 0
LIMIT_6:C(7) 0.863371 0.004891 176.5268 0
LIMIT_7:C(8) 1.496184 0.005024 297.8132 0
LIMIT_8:C(9) 2.804875 0.005538 506.4991 0
LIMIT_9:C(10) 3.112486 0.005723 543.812 0
LIMIT_10:C(11) 4.297395 0.007016 612.4985 0
Pseudo R-squared 0.01601 Akaike info criterion 4.059823
Schwarz criterion 4.059952 Log likelihood -2043574
Hannan-Quinn criter. 4.059859 Restr. log likelihood -2076824
LR statistic 66500.5 Avg. log likelihood -2.029901
Prob(LR statistic) 0


95
A.1.4.9 Ordered Logit (with Returns) Results: 10 Ordered Values

Dependent Variable: AJ
Method: ML - Ordered Logit (Quadratic hill climbing)
Sample (adjusted): 2 1006737
Included observations: 1006736 after adjustments
Number of ordered indicator values: 10
Convergence achieved after 5 iterations
Covariance matrix computed using second derivatives
Variable Coefficient Std. Error z-Statistic Prob.
GARCHVAR 0.268841 0.002158 124.5891 0
AJ(-1) 0.575528 0.001002 574.3314 0
10000*DLOG(ASK) 1.283457 0.001961 654.4871 0
Limit Points
LIMIT_2:C(4) -1.499334 0.008444 -177.5703 0
LIMIT_3:C(5) 1.549476 0.005795 267.3935 0
LIMIT_4:C(6) 2.288018 0.005979 382.6506 0
LIMIT_5:C(7) 2.734323 0.006163 443.6451 0
LIMIT_6:C(8) 3.100859 0.006343 488.8769 0
LIMIT_7:C(9) 4.034615 0.006789 594.2899 0
LIMIT_8:C(10) 5.85286 0.007889 741.9462 0
LIMIT_9:C(11) 6.285226 0.008209 765.6376 0
LIMIT_10:C(12) 7.959922 0.010167 782.9249 0
Pseudo R-squared 0.158021 Akaike info criterion 3.473909
Schwarz criterion 3.47405 Log likelihood -1748643
Hannan-Quinn criter. 3.473948 Restr. log likelihood -2076824
LR statistic 656363.1 Avg. log likelihood -1.736943
Prob(LR statistic) 0

ANALYSIS OF SINGAPORE’S FOREIGN EXCHANGE MARKET MICROSTRUCTURE

WAN CHEE WAI

SINGAPORE MANAGEMENT UNIVERSITY 2011

Analysis of Singapore’s Foreign Exchange Market Microstructure

by Wan Chee Wai

Submitted to the School of Economics in partial fulfillment of the requirements for the Degree of Master of Science in Economics

Thesis Committee:

Tse Yiu Kuen (Supervisor/Chair) Professor of Economics Singapore Management University Hoon Hian Teck Professor of Economics Singapore Management University Anthony Tay Associate Professor of Economics Singapore Management University

Singapore Management University 2011 Copyright (2011) Wan Chee Wai

by applying and modifying the empirical methodology designed by Bollerslev and Melvin (1994). We found that for the USD/SGD: i) the size of bid-ask spreads are positively related to the underlying exchange rate volatility. . we examine the relationship between bid-ask spreads and the underlying volatility of the USD/SGD. Specifically. April-May 2006. and iii) the size of the bid-ask spreads may also be positively related to the directional movement of exchange rates.Analysis of Singapore’s Foreign Exchange Market Microstructure Wan Chee Wai Abstract This paper analyses the Singapore foreign exchange market from a microstructure approach. ii) the magnitude of the dependence on underlying volatility increases as tick volume increases. April-May 2009). Our data set comprises high-frequency USD/SGD tick data of three separate years (April-June 1989.

...1........5 3...............................2 2............75 i .1.............1 A...........3 Bid-Ask Spreads and Asymmetric Information............1..69 MA(1)-GARCH(1...................................................................2...........................44 Ordered Response Analysis ..........................1.......................1..................................................................1 A.............1.......................73 Ordered Logit Results: 4 Ordered Values ........................18 Bollerslev & Melvin’s Model of Volatility and Bid-Ask Spread .............................60 Ordered Probit Results: 4 Ordered Values.......2..............................................................................10 2......iv 1 Introduction........................................................................................................................................3 A..............62 Ordered Probit Results: 10 Ordered Values......................1.........................................................................1.....................1...................................................................1...............1..............................1............2 A...........6 Empirical Methodology ...............................................................5 A...1) Results: .........53 4 Conclusion ...........60 A..............................................................................3 A......64 Ordered Logit Results: 4 Ordered Values ...........................1 2.................1........................1..........................1.....................................60 MA(1)-GARCH(1...................................................21 3 Empirical Analysis ...........................................................1......74 Ordered Logit (with Returns) Results: 4 Ordered Values ...............4 A......71 Ordered Probit Results: 10 Ordered Values........57 Appendix 1 – EViews Results.........................................8 A..9 2 Bid-Ask Spreads in Exchange Rates..............................2 A.............6 A.......1................................................65 Ordered Logit (with Returns) Results: 4 Ordered Values ...............1.................................Table of Contents Table of Contents ...............9 A.55 References.................2......................1.......................................................................1..............................................................6 A......5 A......................................... 11 Bid-Ask Spreads and Volatility................1....................................47 Incorporating Returns .....................................................................................................1..................................70 Ordered Probit (with Returns) Results: 4 Ordered Values.....1 A......27 3.............1..................................2 The Microstructure Approach to Exchange Rate Economics ................2 3.................................................1 1............................................72 Ordered Probit (with Returns) Results: 10 Ordered Values....................................................2..........................61 Ordered Probit (with Returns) Results: 4 Ordered Values..............69 Ordered Probit Results: 4 Ordered Values.........................................1....................................................................................28 Description of the Data...............1...........................1........................67 Ordered Logit (with Returns) Results: 10 Ordered Values ..........................4 3..................................................2 1................................................i Acknowledgement ........2 A.........2...2..................7 Dataset 1: USD-DEM April 1989 to June 1989: ..............................4 A..........3 The Singapore Foreign Exchange Market ....................1..................................5 Organization of Thesis.........2............50 Summary............3 3..................7 A............1) Results: .......63 Ordered Probit (with Returns) Results: 10 Ordered Values...............1 3......................................................34 GARCH Analysis ......................68 Dataset 2: USD-USD April 1989 to June 1989: .....................................66 Ordered Logit Results: 10 Ordered Values ...................1 1...

.............5 A........3..............9 A..87 MA(1)-GARCH(1......95 ii .1..................................1..........84 Ordered Logit Results: 10 Ordered Values ..................3.............3 A....9 Ordered Logit Results: 10 Ordered Values .........................................9 A..80 Ordered Probit Results: 10 Ordered Values..................................................1.....91 Ordered Logit Results: 4 Ordered Values ...8 A....3 A....8 A............................................................................78 Ordered Probit Results: 4 Ordered Values...................................4.......88 Ordered Probit (with Returns) Results: 4 Ordered Values............................82 Ordered Logit Results: 4 Ordered Values ...........3 A.........3..........1) Results: .3......3.....7 A................................93 Ordered Logit Results: 10 Ordered Values .....1......3...........................................1................................8 A....92 Ordered Logit (with Returns) Results: 4 Ordered Values ...................1....76 Ordered Logit (with Returns) Results: 10 Ordered Values ..7 A...4 A.....4.......................................1...........1..............................1......................................2 A.................1......................3.............................89 Ordered Probit Results: 10 Ordered Values....1) Results: .......4...................4...........................86 Dataset 4: USD-USD April 2009 to May 2009: .................1..................2..........77 Dataset 3: USD-USD April 2006 to May 2006: ...............................5 A.........3.....4.....1......87 Ordered Probit Results: 4 Ordered Values...........1..................85 Ordered Logit (with Returns) Results: 10 Ordered Values ...............1 A.................94 Ordered Logit (with Returns) Results: 10 Ordered Values ................4 A.........1................78 MA(1)-GARCH(1.3.........2 A..................1.........1.......................................1..........83 Ordered Logit (with Returns) Results: 4 Ordered Values .........79 Ordered Probit (with Returns) Results: 4 Ordered Values....6 A.......6 A..............1................................1..........1.........81 Ordered Probit (with Returns) Results: 10 Ordered Values.2.4...4.90 Ordered Probit (with Returns) Results: 10 Ordered Values....1........................A.........................1 A.....4...1.....4 A.............................................4..

....52 iii .......................................................................40 Table 12: Frequency Distribution of Spreads USD-SGD April-June 2006................................51 Table 22: Ordered Logit Estimates (with Returns) for all Datasets (4 ordered values).......................50 Table 21: Ordered Probit Estimates (with Returns) for all Datasets (4 ordered values).....................45 Table 17: Ordered Probit Estimates for all Datasets (4 ordered indicator values) .43 Table 16: GARCH Estimates for all Datasets .....................................................................................................................47 Table 18: Ordered Logit Estimates for all Datasets (4 ordered indicator values)..........38 Table 9: Frequency Distribution of Spreads USD/SGD April-June 1989 .................................................................................8 Table 2: Distribution of Spreads of USD/DEM April-June 1989.39 Table 11: Frequency of No Spread Change for USD-SGD April-May 2006 ..........................................42 Table 15: Frequency Distribution of Spreads USD-SGD April-May 2009..............................39 Table 10: Volume of Quotes of USD-SGD April-May 2006...........................37 Table 8: Frequency of No Spread Change for USD/SGD April-June 1989 .............................41 Table 14: Frequency of No Spread Change for USD-SGD April-May 2009 ....................47 Table 19: Ordered Probit Estimates for all Datasets (10 ordered indicator values) .....................................................36 Table 5: Frequency of No Spread Change with Olsen USD/DEM 1989 ........................................................................35 Table 4: Frequency of No Spread Change reported by B&M USD/DEM 1989 ...........................................................List of Figures & Tables Figure 1: USD/SGD vs GDP & Quotes ..49 Table 20: Ordered Logit Estimates for all Datasets (10 ordered indicator values)....32 Table 3: Quote Volume differences between B&M and Purchased Data USD/DEM 1989 ...51 Table 23: Ordered Probit Estimates (with Returns) for all Datasets (10 ordered values)...........36 Table 6: Frequency Distribution of Spreads USD/DEM 1989 ...41 Table 13: Volume of Quotes of USD/SGD April-May 2009.............................................3 Table 1: USD/SGD April-May 2006 Spread Behavior ...........................37 Table 7: Volume of Quotes of USD/SGD April-June 1989...........52 Table 24: Ordered Logit Estimates (with Returns) for all Datasets (10 ordered values)............................

and steadfast encouragement to me during this time when I spent more time on this thesis than with her. guidance and advice to me with regards to the completion of this thesis. Lilian Seah for her tireless efforts with regards to the administrative aspects of the thesis. this paper would never have seen the light of day. patience. my academic life (the completion of this thesis!) had unfortunately taken a backseat. I would like to thank Erfei for her love. Due to challenges in my professional life (overseas posting!) and priorities in my personal life (wedding!). Lastly. Without the generous assistance of these kind professors.Acknowledgement I would like to thank Professor Tse Yiu Kuen. and for tolerating my nonsense. I also like to thank Ms. iv . I promise this will be my last part-time academic pursuit. Professor Hoon Hian Teck and Professor Anthony Tay for their patience.

and repeat the analysis on over 600. We also want to examine how relationship between the size of the bid-ask spreads and exchange rate volatility changes as the USD/SGD grows in volume and significance. and on over 1 million USD/SGD quotes from the same months in 2009. In this paper. Tim Bollerslev. an early microstructure paper in 1994.000 continuously recorded USD/DEM quotes over a 3 month period from April to June in 1989. B&M). we want to examine whether B&M’s result is applicable to a much lesser-traded currency belonging to a much smaller developing economy with a government-managed floating exchange rate regime – the USD/SGD. 1 . The volume of USD/SGD quotes from April to June in 1989 is slightly over 8.000 USD/SGD quotes from April to May 2006. and as Singapore evolves into a developed country. Their dataset consist of more than 300. and showed that the size of the bid-ask spread of is positively related to its underlying exchange rate uncertainty. We begin with a dataset of the USD/SGD in the same 3 month period as per B&M.000. performed an empirical analysis on the USD/DEM. one of the most highly-traded currency pair in 1989. Hence.1 Introduction In “Bid-Ask Spreads and Volatility in the Foreign Exchange Market – An Empirical Analysis”. and Michael Melvin (henceforth. we fast-forward 17 and 20 years into the future from 1989.

On a trade-weighted basis. specifically focusing on the bid-ask spreads of the USD/SGD. which provides flexibility for the system to accommodate short-term fluctuations in the foreign exchange markets as well as some buffer in the estimation of Singapore’s equilibrium exchange rate. the SGD has appreciated against the exchange rates of its major trading partners and competitors since 1981. we provide more background to these two underlying themes in the next two sections: 1.1 The Singapore Foreign Exchange Market The Monetary Authority of Singapore operates a float regime for the Singapore dollar that is managed against a basket of currencies of the country’s major trading partners and competitors. The trade-weighted exchange rate is allowed to fluctuate within an undisclosed policy band. But first. high productivity growth. The composition of the basket is revised periodically to take into account changes in Singapore’ trade patterns.This paper examines the Singapore foreign exchange market from a microstructure approach. The various currencies are given different weights depending on the extent of trade dependence with that particular country. and high savings rate. 2 . We also present a review of some microstructure literature in this area. reflected rapid economic development.

The following figure shows the USD/SGD, Singapore GDP, and volume of quotes over the months of April and May from 1989 to 2009.
USD/SGD vs GDP & Quotes 2.00 1.90 1.80 USD/SGD 1.70 1.60 800 1.50 1.40 1.30 1.20 1988 600 400 200 0 1993 1998 Year USD/SGD GDP Quotes 2003 2008 1800 1600 1400 1200 1000 GDP S$B / Quotes ('000s)

Figure 1: USD/SGD vs GDP & Quotes1

From 1989 to 2009, as the Singapore GDP experienced a gradual growth from SGD 61 billion in 1989 to SGD 266 billion in 2009, the USD/SGD fluctuated between 1.40 and 1.90. Interestingly though, from 2002 onwards the growth of USD/SGD quote volume for April and May experienced a sharp increase from only 33,000 quotes in 2001 to 180,000 quotes in 2002. This growth would then accelerate to 1,630,000 quotes in 2009. It is as if all of a sudden the Singapore dollar started to become more widely traded than ever before. With increasing volume, the underlying exchange rate volatility would also be expected to increase. How much effect would this have on the bid-ask spreads of the USD/SGD. We examine this using data from 1989, 2006, and 2009.

1

USD/SGD and GDP figures obtained from Singapore Department of Statistics (singstat.gov.sg); Quotes volume obtained from Olsen Financial Technologies GmbH.

3

In 2006 the global economy had recorded a fourth consecutive year of strong growth despite the drag from crude oil prices, a buildup in global electronics inventories and adjustment in the US housing market. The MAS Annual Report for 2005/2006 reported that “despite higher oil prices, rising interest rates and natural disasters, the global economy expanded at a robust pace in 2005. This growth momentum continued unabated in the first quarter of 2006. The strength of the US economy was a major factor underpinning the continued growth of the world economy last year. The US economy displayed remarkable resilience against the backdrop of hurricane Katrina and 11 successive increases in the Fed funds rate from 2.25% at the beginning of 2005 to 5% in May 2006. In the first quarter of 2006, growth picked up strongly, led by a rebound in consumer spending and business investment spending on equipment and software.”

The Singapore economy was also in a state of stability, as reported by the MAS Annual Report 2005/2006: “In the early months of 2006, some signs of easing in the domestic economy emerged with growth momentum slowing to 6.8% in Q1. However, this is not indicative of a broad-based slowdown, but rather a retraction to a more sustainable pace of growth.”

In 2009 the world found itself in the midst of the worst global financial crisis ever since the Great Depression. The MAS Annual Report 2008/2009 reported that “2008 was a tumultuous year for the global economy. While the surge in

4

commodity prices led to strong inflationary pressures in the first half of the year, the onset of the global financial crisis caused world growth to fall sharply in the later part of 2008 and into early 2009. The emergence of the Influenza A (H1N1) virus in recent months has added a new dimension of risk to the fragile global economy.”

The global financial crisis, which saw the collapse of Lehman Brothers in September 2008, caused massive economic fallout worldwide. Amidst an erosion of confidence, global trade and industrial production collapsed in the first half of 2009, resulting in a 2.4% year-on-year contraction in world GDP over the same period. During this time, the quote volume for USD/SGD (over April and May) grew to over 1.6 million quotes, suggesting the increase in volatility in the exchange rate.

1.2

The Microstructure Approach to Exchange Rate Economics

Exchange rate economics, the branch of international economics and finance which attempt to explain the foreign exchange market, is an intriguing area of research.

There are many theories of exchange rate determination, from the open economy IS-LM models that are mandatory fare for any undergraduate economics course, to more advanced models such as the Mundell-Fleming model, the

5

But foreign exchange rates can only move through a trading process in a foreign exchange market.g. tourism. and the portfolio balance model. nominal rigidities and imperfect competition. for example). import/export transactions).sticky/flexible-price monetary models. how would any macro model (even one with microfoundations) designed to determine the “new” equilibrium exchange rate be able to take into account all the possible paths taken to transit from the “old” equilibrium rate to the “new” one? Some foreign exchange transactions between the two countries might be related to shifting fundamentals (e. but others transactions may not (e. If the long-run exchange rate between two countries is expected to change due to some shifting fundamental value (say productivity level. All of them however are not able to convincingly explain short-run exchange rate fluctuations. these theories have various degrees of success in forecasting long-run exchange rates. new open-economy macroeconomists attempt for formalize exchange rate in the context of dynamic general equilibrium models with explicit microfoundations. From a common-sense perspective. More recently. Macro foreign exchange models often assume that foreign exchange rates will move when fundamentals move. this is hardly surprising. When tested against empirical evidence.g. etc). speculation in each others’ asset markets. Foreign exchange 6 .

Such data is usually hard to come by and are proprietary to banks and other high-level participants in the foreign exchange market. they tested and established four empirical results: (1) transaction flows forecast future macro variables such as output growth.microstructurists study this trading process. such as Martin D. Richard Lyons (2001) defined “Order Flow” and “Bid-Ask Spreads”. as the two hallmarks of the microstructure approach. D. and (4) the forecasted part of fundamentals is better at explaining exchange rates than standard measured fundamentals. Evans and Richard K. “The Microstructure Approach to Exchange Rates”. In their 2007 paper “Exchange Rate Fundamentals and Order Flow". Well-connected researchers. These are analogous to “Quantity” and “Price” in the dimension of exchange rates. two variables that are absent from the macro approach. Order flow is essentially transaction volume which is “signed”. and inflation. (3) transaction flows (proprietary) forecast future exchange rates. meaning it includes information if the transaction is a sale or a purchase. In his book.5 years. (2) transaction flows forecast these macro variables significantly better than the exchange rate does. Lyons managed to obtain proprietary data on all end-user EUR/USD trades received at Citibank over 6. 7 . money growth.

53% 3. such as information flow. But other than being easily obtainable. they are a ready target for testable hypotheses. spreads tended to 8 .21% 39.000 quotes over two months.60% 20. divided into nine categories of price and spread movements. and 6. the spread remained unchanged 6. Lyons highlighted that one reason spreads receive so much attention is because.05% 3. This is in contrast to other features in the trading that are not so easily measurable.85% 39.47% SPREAD DOWN 3.35% of the time when the price moved up.35% 7. Bid-Ask Spreads.Data for the other hallmark. is much easier to obtain. while B&M obtained theirs by collecting every USD/DEM quote posted on the Reuters screen for the interbank foreign exchange market for three months in 1989.73% SPREAD SAME 6. For example. the four datasets for this paper were purchased from Olsen Financial Technologies GmbH.52% 6. In fact.46% 4.81% PRICE UP PRICE SAME PRICE DOWN Total Table 1: USD/SGD April-May 2006 Spread Behavior Out of over 600. what is intriguing is that when prices move up. being a core element of most data sets. Though we may somewhat expect for spreads to move when prices move.000 USD/SGD quotes in April and May 2006. belief dispersion.42% 32.60% of time when the price moved down. on the other hand. etc. SPREAD UP 32. The behavior of bid-ask spreads in the foreign exchange markets offers many opportunities for research. Table 1 below shows the distribution of over 600.

1. and ends with a description of B&M’s model that relates volatility to bid-ask spreads.21% of the time. The main currency for analysis is the USD/SGD. spreads tended to narrow 32. Chapter 4 concludes. Spread behavior like in the previous example may differ across different currencies within the same time period.widen 32.85% of the time but yet widened only 3. For comparison against a different currency within the same time period. we also perform an empirical analysis on the phenomenon described above in Table 1. we perform this analysis for the USD/SGD from the months of April and May in 1989. and 2009. 9 . and may also differ across different time periods within the same currency. Conversely. when prices move down. 2006. but yet narrowed only 3. The empirical objective of this paper is thus to examine the relationship between the bid-ask spread and the underlying exchange rate volatility from such two angles.53% of the time.3 Organization of Thesis The rest of this paper shall be organized as follows: Chapter 2 begins with a review of microstructure papers which concerns bid-ask spreads. In addition.46% of the time. For comparison against different time periods within the same currency. Chapter 3 describes B&M’s empirical methodology and presents the empirical analyses for each dataset. we use B&M’s results for the USD/DEM.

” For the first section of this Chapter. He also gave two more reasons for the heavy attention and resources focused on bid-ask spreads. Contrastingly market microstructure focused on how trading mechanisms affected prices. Besides being relatively obtainable. The second reason was because practitioners are “intensely concerned with managing trading costs”. We then present the views of a more contemporary paper which refutes 10 . The third reason had to do with the history of the field of market microstructure which in its early days sought to distinguish itself from the literature on trading under rational expectations. and focus on “adverse selection” as a popular theme. Lyons (2001) cites bid-ask spreads as one of the two hallmarks of the microstructure approach. Rational expectations models generally omit trading mechanisms when characterizing the relationship between fundamentals and price. we present a short survey on some early microstructure literature that concerns the determination of bid-ask spreads. he explained that spreads receive so much attention because they form a core element of most data sets and are a ready target for testable hypotheses. and this had led to “a focus on the determination of real-world transaction prices – spreads.2 Bid-Ask Spreads in Exchange Rates As previously mentioned in Chapter 1. guided by Sarno & Taylor (2002).

We then close this Chapter by presenting B&M’s simple asymmetric information model which forms the framework for our empirical investigation. inventory holding costs. From the second section of this Chapter we return to our main topic of interest – bid-ask spreads and exchange rate volatility. Cost of Dealer Services The cost of dealer services is formally analyzed by Demsetz (1968) who assumes the existence of some fixed costs of providing “predictable immediacy” as the service for which compensation is required by market makers.“adverse selection” as a determinant of bid-ask spreads and proposes that asymmetric information might be a more plausible candidate. following Sarno & Taylor (2002). Similarly. 2. 11 . his definition for cost of dealer services could also be applicable to the foreign exchange market. While Demsetz focused on the New York Stock Exchange. we first present a short survey on the literature which analyses the proportional relationship between spreads and exchange rate volatility.1 Bid-Ask Spreads and Asymmetric Information Sarno & Taylor (2002) identifies three main determinants of the bid-ask spread: the cost of dealer services. and the cost of adverse selection.

According to Demsetz. relates to the market maker's ability to make inventory adjustments when the market for an issue is "thin. which they termed “marketability risk”. and 2) high volume implies less marketability risk. “predictable immediacy is a rarity in human actions. Inventory Holding Costs The original argument of inventory costs as a crucial determinant of bid-ask spreads was first propositioned by Barnear & Logue (1975). it is the inventory markup of retailer or wholesaler”." They showed that volume has a negative effect on the bid-ask spread for two reasons: 1) high volume implies more competition if it implies more competition among alternative market makers. The ask-bid spread is the markup that is paid for predictable immediacy of exchange in organized markets. and to approximate it requires that costs be borne by persons who specialize in standing ready and waiting to trade with the incoming orders of those who demand immediate servicing of their orders. who tested a modified theory of market-maker behavior first espoused by Bagehot (1971). and. therefore. Barnear & Logue modified the theory of the market-maker spread by distinguishing between the two major components of inventory risk. in other markets. Amihud & Mendelson (1980) considers the problem of a price-setting monopolistic market-maker in a Garman (1976) dealership market where the 12 . lower positioning costs. The second component.

The bid-ask spread is given by a risk neutral spread that maximizes expected profits for the given stochastic demand functions plus a risk premium that depends on transaction size. They show that as time unfolds and transactions occur. and the converse is true when inventory decreases. 13 . like the existence of a ‘preferred’ inventory position and the downward monotonicity of the bid-ask prices. but the dealer’s price adjustment does.stochastic demand and supply are depicted by independent Poisson processes. The focus of their analysis is the dependence of the bid-ask prices on the market-maker’s inventory position. The bid-ask spread does not depend on the dealer's inventory position. When inventory increases both bid price and ask price decline. Ho & Stoll (1981) considers the stochastic dynamic programming problem of solving for the optimal behavior of a single dealer of a single stock who is faced with a stochastic demand for his services and return risk on his stock and on the rest of his portfolio. the return variance of the stock and the dealer's attitude toward risk. They derived the optimal policy the results are shown to be consistent with some conjectures and observed phenomena. the dealer is able to set his bid price and ask price relative to his opinion of the "true" price of the stock so as to maximize the expected utility of terminal wealth.

whose model includes two types of market participants – those are willing to pay the price of the spread to the market-maker in exchange for predictable immediacy and those who can speculate at the expense of the market-maker using some private insider information. The commitment 14 . Cost of Adverse Selection Adverse selection is a common argument to explain the existence of bid-ask spreads. Copeland & Galai (1983) analyses the determination of bid-ask spreads in organized financial markets. The origin of this argument could be traced back to Bagehot (1971). inventory costs models assume that market-markers optimize their inventory holding. numerous microstructure papers have drawn on adverse selection as their primary interpretive framework. An adverse selection arises because market-makers are not able to distinguish between the two types of participants and resort to widening the spreads for both types.Generally. Since Bagehot. where the trading is done through economic agents who specialize in market-making for a limited set of securities. and generally imply that market-makers shift the spread downwards and increase the width of the spread when a positive inventory is accumulated. The bid-ask spread then becomes the market-maker’s defense against adverse selection in “in the form of exploitation of arbitrage opportunities”.

made by dealers to buy or sell at the bid and ask prices. The dealer establishes his profit maximizing spread by balancing the expected total revenues from liquidity trading against the expected total losses from informed trading. is analyzed as a combination of put and call options. They also showed that. and that the bid-ask spread increases with greater price volatility in the asset being traded. The exercise prices of the straddle determine the bid-ask spread. Given the behavior of liquidity traders and informed traders. including the exogenous arrival patterns of insiders and liquidity traders. They showed that a monopolistic dealer will establish a wider bid-ask spread than will perfectly competitive dealers. and with lower volume. could account for the existence of a bid-ask spread. Glosten & Milgrom (1985) analyzed a model of a securities market in which the arrival of traders over time is accommodated by a market-maker. 15 . the dealer is assumed to offer an out-of-the-money straddle option for a fixed number of shares during a fixed time interval. and the average magnitude of the spread depends on many parameters. by itself. bid-ask spreads tend to decline with trade. with a higher asset price level. because transaction prices are informative. and the quality of the information held by insiders. the elasticity of supply and demand among liquidity traders. respectively. They showed that adverse selection.

and that customer order flow is an important source of private information. can 16 . that market net volume is only partially observable. Osler. Mende & Menkhoff (2006) however shows evidence that the behavior of bid-ask spreads is inconsistent with adverse selection. Payne (2003) estimates a VAR decomposition of interdealer trades and quotes and interprets the results through the lens of adverse selection. that over 80% of the trading volume is between market-makers. He presented a model which incorporated a number of institutional features relevant to the FX market. he used one trading week’s worth of USD/DEM data derived from an electronic foreign exchange brokerage and employed the framework contained in Hasbrouck & Sofianos (1993) to test for the existence of private information effects of trading on prices. indicating that adverse selection costs account for around 60% of the half-spread. They outline three factors that seem likely to be important.Lyons (1995) was likely one of the first who departed from early microstructure work that focused almost entirely on stock markets and applied such theory to foreign exchange markets. such as the facts that major currencies are traded in decentralized dealership markets. Specifically. The first factor. His basic results confirm the existence of private information on FX markets. fixed operating costs. Lyons showed that trade size and the bid-ask spreads of a particular dealer were positively related.

so each individual dealer can exert a certain amount of market power despite the competition.explain the negative relation between trade size and bid-ask spreads if some costs are fixed. but cannot explain the cross-sectional variation across customer types. In foreign exchange. The second channel through which asymmetric information might affect bid-ask spreads in foreign exchange involves strategic dealing. The market power hypothesis suggests that firms.. (2005). As suggested in Green et al. one involving market power and a second involving strategic dealing. To explain why bid-ask spreads are larger for commercial than financial customers they suggest that asymmetric information – in the broad sense of information that is held by some but not all market participants – may influence spreads through two channels distinct from adverse selection. commercial customers typically know far less about market conditions than financial customers so they might be expected to pay wider spreads. It can be costly for customer firms to search out the best available quotes in the foreign exchange market. even in a market with hundreds of competitors like foreign exchange. Evans and Lyons 17 . as they do. Building on abundant evidence that customer order flow carries information (e.g. and market power in quote-driven markets depends on knowledge of current market conditions. dealers may quote the widest spreads when their market power is greatest. gain market power from holding information.

Overturf. 1982).2 Bid-Ask Spreads and Volatility The directly proportional relationship between bid-ask spreads and exchange rate volatility now represents a fairly stylized fact in the microstructure literature. The idea that dealers strategically vary spreads to gather information was originally explored in Leach and Madhavan (1992.g. 1993). 1975). Early studies modeled the spread as a function of transaction costs.(2007). 2. Osler et al argue that rational foreign exchange dealers might strategically vary spreads across customers. Daníelsson et al. consistent with the pattern in foreign exchange. subsidizing spreads to informed customers in order to gain information which they can then exploit in upcoming interbank trades. by contrast. When applied to two-tier markets in Naik et al. (2002)). and the market-maker’s payoff for facing the exchange rate risk when assuming an open position. (1999) it implies that bid-ask spreads will be narrower for trades with information. 18 . The main conclusions of these early studies are that exchange rate spreads are wider under floating exchange rate than under fixed-exchange rate regimes (e. and that measures of exchange rate volatility are followed closely by exchange rate spreads (e. the bank’s profit from providing liquidity services. Aliber.g. 1975. In standard adverse-selection models. Fieleke. dealers passively accept the information content of order flow.

The model not only provides additional evidence on the proportional relationship between exchange rate volatility and bid-ask spreads in the foreign exchange market. 19 . and non-discretionary liquidity traders. who have relatively superior information and only trade on terms favorable to them. In this model. who must trade at a precise time during the day regardless of the cost. there are three types of agents: informed traders. discretionary liquidity traders. trading volume is explained by the concentration of trade of informed traders and discretionary liquidity traders at certain points in time: the concentrations occur because it is profitable for informed traders to trade when there are many liquidity traders who do not have the same information as themselves and because discretionary liquidity traders are attracted because the larger the number of traders lowers the cost of trading. who must trade during a day but can choose when to trade during the day in order to minimize costs. Admanti & Pfleiderer (1988) provides another fundamental theoretical contribution to this area. but also suggests that market-makers consider moments of the exchange rate higher than the second moment in order to evaluate the probability of large exchange rate changes. In their model.Glassman (1987) provides a significant contribution to this literature in that she builds a model where variables representing transactions frequency are included explicitly and the non-normality of the distribution of exchange rates is taken into account.

They recorded quote arrivals and bid-ask spreads over the trading day. and are less pronounced at the minute-by-minute frequency than at lower frequencies. The patterns of trading activity and spreads during the day also strongly suggest some degree of traders’ risk aversion. the more trading activity is executed by informed traders. across geographical locations as well as across market participants. given which. which is especially pronounced after immediately after jumps in the exchange rate.Bollerslev & Domowitz (1993) used intradaily data to investigate the behavior of quote arrivals and bid-ask spreads. They found that trading activity and the bid-ask spreads for traders whose activity is restricted to regional markets can be described by a U-shaped distribution. The series was also found to exhibit first-order negative serial correlation. which is consistent with Admanti & Pfleiderer’s (1989) model. 20 . Goodhart & Figliuoli (1991) reported a study of minute bid-ask quotes on three days in 1987 at a Reuters screen and found evidence that leptokurtosis and heteroscedasticity are time-varying. the higher the cost of trading. They also found that trading volume is time-varying. being higher at the European and North American openings and lower at the European lunch hour. Multivariate analysis suggested significant relationships between lagged exchange rates and the current spot rate.

but no one had performed any explicit analysis of the relationship between the magnitude of foreign exchange market spreads and the underlying exchange rate volatility. Researchers. providing most importantly a methodology to analyze the relationship between bid-ask spreads of exchange rates and its underlying volatility. as 21 . Hence. B&M opined that while unambiguous 'good' or 'bad' news regarding the fundamentals of the exchange rate should have no systematic effect on the spread.3 Bollerslev & Melvin’s Model of Volatility and Bid-Ask Spread B&M is the main inspiration for this thesis. B&M is likely to be the first paper to touch on this subject.2. In the early 1990’s as B&M were writing their paper. Earlier studies on the subject. such as Glassman (1987) and Boothe (1988). concentrated on the own statistical properties of the spread. as both the bid and the ask prices should adjust in the same direction in response to the traders receiving buy or sell orders that reflect the particular news event. however greater uncertainty regarding the future spot rate. Bossaerts and Hillion (1991). Black (1991). Melvin and Tan (1996) and Bollerslev and Domowitz (1993). the bid-ask spread component of transactions costs in the foreign exchange market had not received much attention in the literature. had attempted to offer empirical and theoretical analyses of the determinants of foreign exchange market spreads. such as Goodhart (1990).

They are also not speculators. The formal setup for B&M’s stylized market microstructure model is based on the analysis in Glosten and Milgrom (1985). information asymmetries regarding fundamentals underlying the determination of the spot exchange rate. Admati and Pfleiderer (1989). more generally.associated with greater volatility of the spot rate. services and financial assets. The model assumes that the foreign exchange market comprises two kinds of traders: liquidity traders and information-based traders. Information-based traders profit by intermediating the demands and supplies of foreign exchange for the liquidity traders. and Andersen (1993). We now outline B&M’s simple theoretical framework that illustrates this role of volatility in determining the spread. These traders also take positions in the foreign exchange market based on information advantages received through their dealings with the liquidity traders or. is likely to result in a widening of the spread. 22 . Liquidity traders participate in foreign exchange transactions only due to the needs of their normal business activity which require international trade of goods.

one of the many market-making traders will set bid and ask quotes. B&M further assumes that the standardized innovations. The bid-ask spread is assumed to be set symmetrically around the known fundamental price prevailing at the time of quote formation. At time t-1. At = st-1 + kt. Thus.The liquidity traders constitute the proportion (1—) of the total market participants. The liquidity traders receive a signal to either buy or sell foreign currency regardless of the actual value of the currency in comparison with the bid or ask prices prevailing at the time.t-1. distributed through time. tt-1. Informed traders constitute the remaining  proportion of the market. . This fundamental value is assumed to evolve over time according to a martingale model. and Bt = st-1 – kt. are independent and symmetrically. 23 . the quoted spread for trades at time t. i. This group of traders receives some information about the true underlying fundamental value of the exchange rate st. Bt and At. (1) where Et-1(t)=0.Bt = 2kt. good for trading at time t. Et-1(t2)= t2.t-1.e. but not necessarily identically.t-1 depends on time t-1 information only. and Et-1(•) denotes the conditional expectation based on the information set generated by the past values of st. Kt = At.

The liquidity traders only know st-1 and expect t to equal zero. for the market-maker when the opposite party an information-based trader. the loss for the quoting trader relative to the true value st arising from informed trading is therefore (2) Let Pt-1(•) denote the probability conditional on the time t-1 information. the information-based traders cannot profit from knowing the true fundamental value revealed by t. Since the standardized innovations. on average. Zt = tt-1. Trader positions are limited by the convention that existing quotes are only good for up to some maximum quantity of currency. < Bt. For values of Bt ≤ st ≤ At. who received the signal t buy currency if At < st and sell currency if st. are assumed to be independent and symmetrically distributed through time.Trades at existing quotes will generate losses. Assuming that the market-makers limit trading to one unit of currency at existing quotes. Information-based traders. the expected loss from informed trading may be expressed as (3) 24 .

(3) with the gain in Eq. (4) yields the expected profit for the market-maker conditional on time t-1 information: In equilibrium. it follows that the expected profit for the quoting trader conditional on an uninformed trade equals: (4) Combining the expected trading loss in Eq. Kt =2kt. competition from other banks or market-makers will drive this expected profit to zero.t-1. (6) only depend on the time t-1 information set through t-1kt.t-1 yields (6) Since the conditional expectation and probabilities on the right-hand side of Eq. it follows that in 25 .Assuming an equal probability of a buy or a sell order from the liquidity traders. Expressing this zero profit condition in terms of the total spread.

Based on this relationship between exchange rate volatility and the bid-ask spread. B&M noted that while this simple proportionality condition would no longer hold true in a more general model with endogenous information acquisition. 26 .equilibrium the spread must move proportional to the conditional standard deviation of the true fundamental value of the exchange rate. B&M designed the empirical methodology that we will describe in the next chapter. the result that an increase in t2 leads to an increase in Kt would still remain generally valid.

one of the most highly-traded currency pair between two of the largest economies in the world in 1989 and showed that the size of the bid-ask spread of the USD/DEM is positively related to its underlying exchange rate uncertainty. a semi-floating currency from a much smaller economy. We are curious if 1) B&M’s result would hold for the USD/SGD.3 Empirical Analysis B&M performed an empirical analysis on the USD/DEM.000 quotes. and when the volume of quotes increased to over 600. there were only slightly over 8. 27 . Their dataset consist of more than 300.000 continuously recorded USD/DEM quotes over a 3 month period from April to June in 1989. as that country becomes a significant regional economic power in South-east Asia.000 over April and May 2006. The USD/DEM is a free-floating exchange rate. For the above purposes we purchased four sets of data from Olsen Financial Technologies GmbH. and 3) B&M’s result would hold as that economy enters into a period of worldwide financial crisis. Before we discuss the data and empirical results. we first describe B&M’s empirical methodology and how we adapted it for this paper. 2) B&M’s result would hold 17 years later for that same currency. where over the same period from April to May 1989..

. B&M use the ask price for estimation purposes. .1 Empirical Methodology B&M’s empirical methodology comprises two major steps. the bid and ask prices have virtually identical higher order moments and differ only very slightly in their conditional means. and  are the parameters to be estimated.1) model of the form below seemed to fit their dataset well: (7) where It-1 denotes the time t-1 information set. and as noted by Bollerslev et al. These estimates for the conditional variance are then used as the proxy for exchange rate volatility in the second-stage model for the temporal behavior of the spread. such representations have been documented by numerous studies.3. and . B&M employed a two-stage estimation procedure in which the conditional variance for the spot exchange rate is first estimated as a GARCH process. (1992). They found that the MA(1)-GARCH(1. . Step 1: Creating a Proxy for Exchange Rate Volatility The first step involves using a GARCH model as an explicit proxy for the time-varying volatility of the spot rate. The time t subscript refers to the place in the order 28 .

The primary purpose of the GARCH estimation was to create proxies for the conditional variance of the exchange rate to be used in the investigation of the determinants of the spread. They then saved all the estimates for the conditional variances from each of the 12 models and combined into a single time series of volatility estimates for the full set of weekday quotes. Intuitively. we use modern econometrics software EViews to perform GARCH analysis on our datasets and to obtain the conditional variance time series.000 observations in practice was not feasible in practice.1) model as a non-parametric estimator. 1992). they divided up their dataset into 12 weeks of data and estimated each set for the above GARCH parameters.of the series of quotes.t2 provides an estimate of the price volatility between quotes. Hence. (7) may be justified by the theoretical arguments in Nelson (1990. or a one-sided filter. 29 . the resulting estimates for the conditional variance will generally be consistent as the length of the sampling interval goes to zero. it follows that on interpreting the GARCH(1. But it was 1993 then. so that A. if the sample path for the true unobservable volatility process is continuous. The particular specification for the conditional variance in Eq. Today. and B&M noted that estimating a GARCH model with more than 300.

is defined by K*t = Xi’ + K.t. … aJ. Specifically. Kt.t (9) B&M allowed for multiplicative heteroskedasticity in the spread by parameterizing the logarithm of 2K. The observed spread. K*. a2.t (8) The vector Xt denotes a set of predetermined variables that affect the conditional mean of K*t and K. B&M used an Ordered Probit model with multiplicative heteroskedasticity for this purpose. a1. is conditionally normally distributed with mean zero and variance. The unobservable continuous random variable. 2K.t as a linear function of the same 30 .Step 2: Estimate Relationship between Spreads and Proxy for Volatility The second major step of the methodology involved using ordered response models to estimate the relationship between the time series of volatility estimates obtained in the first step and the bid-ask spreads. is assumed to take on only a fixed number of discrete values.

i. K*. The ordered response models relate the observed spreads to K* via (11) where the Aj’s form an ordered partition of the real line into J disjoint intervals. is defined by K*t = Xi’ +  (10) where  is i. Kt. firstly by omitting multiplicative heteroskedasticity (more due to the limitations of EViews than by choice). 31 . since we omitted multiplicative heteroskedasticity). the observed spread. is similarly assumed to take on only a fixed number of discrete values. but the unobservable continuous random variable.explanatory variables that enter the conditional mean of K*t. a2. a1. and takes the form of the logistic distribution for the Ordered Logit model.d. We depart from this path in our analysis. In our analyses. standard normal for the Ordered Probit model. and secondly by estimating the relationships with both Ordered Probit models and Ordered Logit models (to allow more flexibility in the behaviour of the error term. … aJ.

8%) 5 77. because the spreads distribution for the USD/SGD 2006 and 2009 data have more converging points (certain spread sizes where there are more quotes than others) than both the USD/DEM and the USD/SGD in 1989.892 (56.<5 2.2%) 7 34. < 10 2.8%) 15 < .9%) 20 < .0%) 10 170. < 15 329 (0. as we will see later. the four most commonly observed spreads account for 97. In addition.878 (11.304 (0.2%) Note: Spreads converted into basis points Table 2: Distribution of Spreads of USD/DEM April-June 1989 We also performed the ordered response analyses with four ordered indicator values.5%) 7 < .The probability that the spread takes on the value aj is equal to the probability that K* falls into the appropriate partition.977 (1.0 percent of the total quotes. < 20 39 (0. Aj. From the distribution of spreads of the USD/DEM in 1989.616 (0. B&M based the empirical analysis on a classification of the spread into only four different categories.0%) 20 2. 572 (0. For tractability reasons. USD/DEM Frequency Distribution fof Spreads Spread All Quotes 0<.6%) 5<. we used up to 10 ordered indicator values in our own 32 .1%) 10 < .534 (3.856 (25.<7 607 (0.1%) 15 11.

I. In order to test the hypothesis that the spread is partly determined by the volatility of the spot rate. Xt. In order to take 33 . The ordered response model defined above allows us to estimate the probability of a particular spread being observed as a function of the predetermined variables. the GARCH estimate of the conditional variance for the ask prices is included as one of the elements in Xt. In the case of B&M where by four ordered indicator values of aj’s were used.ordered response analyses. the corresponding intervals for the unobservable latent variable K* are defined by: (12) The partition parameters. are estimated jointly with the other parameters of the model. B&M noted that Bollerslev and Domowitz (1993) indicated a distinct intra-day pattern in the spread distribution and tt is possible that any significant effect of the conditional variance in isolation may merely reflect this dependence rather than provide an independent influence on the spread process.

Dataset 1 – USD/DEM quotes from April 1989 to June 1989 b. we will infer that the hypothesized theoretical link is supported by the empirical analysis. is caused by a larger conditional variance of the spot rate. and this raises the probability of observing a higher spread. Dataset 2 – USD/SGD quotes from April 1989 to June 1989 c.account of this own temporal dependence. Dataset 4 – USD/SGD quotes from April 2009 to May 2009 Dataset 1: USD/DEM from April 1989 to June 1989 For the first dataset we purchased the same dataset used by B&M – USD-DEM quotes from April 1989 to June 1989. Dataset 3 – USD/SGD quotes from April 2006 to May 2006 d.2 Description of the Data We purchased four sets of data from Olsen Financial Technologies GmbH: a. We now describe each of the datasets in more details in the following section: 3. We wanted to repeat the empirical analysis 34 . Kt-1 was included as an element of the Xt vector in the estimation of the ordered response functions for K*t: Kt* = 1A.t2 + 2Kt-1 +  (13) Given the partition boundaries determined by the data. if a higher conditional mean 'X.

325 21.604 100. The distribution of the quotes over each workday is similar in terms of percentage in both datasets. First.51% Mon 56.80% Wed 66.26% Sat 33 0.00% PURCHASED FROM OLSENDATA DAY TICKS Sun 1.13% Fri 53.186 100. The differences between our purchased data and the actual data used by B&M are tabulated as follows: a. Furthermore.600 18. we departed from B&M’s ordered probit procedure by omitting multiplicative heteroskedasticity. We wanted to repeat the estimation processes again using EViews.377 19. B&M’s results were obtained in 1993 using probably not so technologically advanced means. or how they presented 12 workweeks of data from an actual 13 weeks in April to June 1989.634 21.646 18. B&M’s data were obtained from Reuters. we found some minor differences.571 0. Second.186 ticks.79% Fri 56.521 20.on the dataset again to allow for an apples-to-apples comparison between the USD/DEM 1989 and the USD/SGD 1998 results. as mentioned in the previous Chapter.812 20.03% Grand Total 305.082 17.109 21.63% Wed 63.604 ticks of USD/DEM quotes while the dataset purchased from Olsen Financial Technologies GmbH contains 310.25% Tue 67.29% Mon 60.095 19. 35 . From their paper we were unable to ascertain the accuracy of this data.66% Tue 66.01% Grand Total 310. We obtained ours by purchasing from Olsen Financial Technologies GmbH and upon comparison.00% Table 3: Quote Volume differences between B&M and Purchased Data USD/DEM 1989 B&M collected 305.37% Sat 98 0.88% Thu 61.38% Thu 61. Volume of Quotes ORIGINAL BOLLERSLEV & MELVIN DAY TICKS Sun 887 0.

SPREAD UP 64384 13083 24045 101512 NO.86% 10.70% 20. 36 .30% Table 4: Frequency of No Spread Change reported by B&M USD/DEM 1989 B&M reported that 8% of all the quotes observed no change in spread when the bid-ask price rose and that 8. OF QUOTES SPREAD SAME 32787 39422 33407 105616 SPREAD DOWN 23767 13523 64164 101454 PRICE UP PRICE SAME PRICE DOWN Grand Total PRICE UP PRICE SAME PRICE DOWN GRAND TOTAL DIVIDED BY NO OF TICKS MOVED SPREAD SPREAD SPREAD UP SAME DOWN 53.79% 32.48% 19.b.23% 32.3% of all the quotes observed no change in spread when the bid-ask price fell. or on Number of Ticks with No Spread Change divided by Total Number of Ticks. We make this distinction clearer with the purchased data. It was not clear.90% 34.23% 32.88% PRICE UP PRICE SAME PRICE DOWN GRAND TOTAL Table 5: Frequency of No Spread Change with Olsen USD/DEM 1989 Among all ticks.38% 7.78% 4.76% 32.90% 34. Frequency of No Spread Change USD/DEM No.619 8.77% 27.88% DIVIDED BY TOTAL NO. whether these percentages where calculated based on Number of Ticks with No Spread Change divided by Number of Ticks Moved.24% 12. however.63% 7.70% 4.47% 52.24% 27.83% of all ticks observed no change in spreads when the price rose and fell respectively.00% 8. of Quotes.11% 19.63% and 10. OF TICKS SPREAD SPREAD SPREAD UP SAME DOWN 20.81% 59.65% 19. & Frequency of No Spread Change No Change in Spread Number Bid-Ask Rise Bid-Ask Fall All quotes 304.79% 10.83% 20. we counted that 10.

1%) 15 11.14% 19. Distribution of the quotes over the week are however quite similar. 37 .1%) 10 < . Frequency Distribution of Spreads Frequency Distribution of Spreads ORIGINAL Spread 0<.1%) 10.5%) 7 < .0%) 171.616 (0. there are only 8.304 (0.682 (11.977 (1.892 (56.0%) 20 2.<5 2.50% 18.0%) 10 170.652 1. Volume of Quotes DAY Sun Mon Tue Wed Thu Fri Sat Grand Total USD/SGD 6 1. a.8%) 15 < . < 10 2.00% Table 7: Volume of Quotes of USD/SGD April-June 1989 Compared against a major currency in 1989 over the same period from April to June.8%) 5 77.534 (3.2%) 35.9%) 20 < .8%) 553 (0.5%) 327 (0. followed by 5 basis points. 572 (0.07% 15.2%) Note: Spreads converted into basis points OLSENDATA 2.307 2.25% 100.000 for the USD/DEM.43% 23.876 1.264 (55.586 21 8.<7 607 (0.72% 0.6%) 3.c.6%) 5<.8%) 81.192 (1.89% 22.389 (3.856 (25.472 USD/SGD quotes compared to over 310.4%) 38 (0.0%) 2.2%) 7 34. < 15 329 (0.368 (26.4%) 660 (0. Dataset 2: USD/SGD from April 1989 to June 1989 The second dataset are USD/SGD quotes from April 1989 to June 1989.2%) Table 6: Frequency Distribution of Spreads USD/DEM 1989 It appears that the frequency distribution of spreads is quite similar between the dataset reported by B&M and the dataset purchased from Olsen Financial Technologies GmbH.024 1. The most common bid-ask spread is 10 basis points.878 (11.547 (0. < 20 39 (0.472 0.562 (0.

62% 2.62% of the time.02% DIVIDED BY TOTAL NO.06% 4. Frequency of No Spread Change Compared against the USD/DEM in 1989 over the same period from April to June. OF TICKS SPREAD SPREAD SPREAD UP SAME DOWN 1.02% PRICE UP PRICE SAME PRICE DOWN GRAND TOTAL Table 8: Frequency of No Spread Change for USD/SGD April-June 1989 38 . OF QUOTES SPREAD SAME 3108 1340 3168 7616 SPREAD DOWN 289 63 72 424 PRICE UP PRICE SAME PRICE DOWN GRAND TOTAL DIVIDED BY NO OF TICKS MOVED SPREAD SPREAD SPREAD UP SAME DOWN 2.50% 89.87% 0.73% 15.18% 5.80% 3. we do not observe as much spread changes in the USD/SGD when prices moves.47% 4.30% 4.80% 90.30% 7.with volume peaking during midweek.80% 90.85% 4.32% 90. PRICE UP PRICE SAME PRICE DOWN Grand Total SPREAD UP 87 62 256 405 NO.18% 5.03% 37.23% 91. b. the spread remained unchanged 90.03% 36. When prices move downwards. When prices move upwards. the spread remained unchanged 89.21% of the time.21% 8.42% 0.51% 0.75% 3.

24% 20. 14 (0. < 15 6 (0.31% 22.979 0.854 (93.1%) 5 137 (1.1%) 15 76 (0. When spreads do change.0%) 10 < . < 10 43 (0.037 87.c. a. Frequency Distribution of Spreads Frequency Distribution of Spreads USD/SGD Spread 0<.32% 14.<7 22 (0.9%) 15 < .812 147. Volume of Quotes DAY Sun Mon Tue Wed Thu Fri Sat Grand Total USD/SGD 3.0%) 20 164 (1.084 135.09% 100.9%) 20 < . Dataset 3: USD/SGD from April 2006 to May 2006 The third dataset are USD/SGD quotes from April 2006 to May 2006.00% Table 10: Volume of Quotes of USD-SGD April-May 2006 15 years later.<5 6 (0.44% 0. we note similar characteristics to the USD/DEM in that the most common bid-ask spread is 10 basis points.327 126.332 555 604.400 ticks in 39 . < 20 1 (0.5%) 10 7.6%) 5<. Compared against itself in 1989 over the period from April to May.63% 17. the volume of quotes of the USD/SGD has grown from over 8.2%) Note: Spreads converted into basis points Table 9: Frequency Distribution of Spreads USD/SGD April-June 1989 While spread changes in the USD/SGD are uncommon when prices moves.4%) 7 < . the USD/SGD has grown to become a major currency in Southeast Asia.832 104. they are most likely to be 5.96% 24.3%) 7 122 (1. 7 or 20 basis points.

99% of the time compared to 89.000 ticks in 2 months.46% of the time compared to 90.21% 6. when prices move upwards. the spread remained unchanged only 14.01% 39.62% of the time in 1989.53% 4. Now.60% 32. Frequency of No Spread Change In 2006. PRICE UP PRICE SAME PRICE DOWN Grand Total SPREAD UP 194955 24340 19302 238597 NO.15% 22. When prices move downwards. OF TICKS SPREAD SPREAD SPREAD UP SAME DOWN 32.3 months to over 604.73% 20. The distribution of quotations over the week remains consistent with volume peaking during midweek.47% 39.67% 14.42% 3.53% 15.83% 7.81% PRICE UP PRICE SAME PRICE DOWN GRAND TOTAL Table 11: Frequency of No Spread Change for USD-SGD April-May 2006 40 .52% 3.35% 3.47% 39. we observed that the tendency for spreads to remain unchanged when prices move is reduced dramatically. OF QUOTES SPREAD SAME 38121 45185 39611 122917 SPREAD DOWN 21197 20568 197313 239078 PRICE UP PRICE SAME PRICE DOWN GRAND TOTAL DIVIDED BY NO OF TICKS MOVED SPREAD SPREAD SPREAD UP SAME DOWN 76.02% 50.05% 7.21% of the time in 1989. the spread remained unchanged only 15.85% 39.46% 6. b.46% 77.99% 8.73% 20.34% 27.81% DIVIDED BY TOTAL NO.

<7 51. Dataset 4: USD/SGD from April 2009 to May 2009 The fourth dataset are USD/SGD quotes from April 2009 to May 2009.00% Table 13: Volume of Quotes of USD/SGD April-May 2009 In the year of the 2009 economic crisis.<5 51. The most common spread in 2006 is 5 basis points. < 10 53. Frequency Distribution of Spreads Frequency Distribution of Spreads USD/SGD Spread 0<.336 183. < 15 25 (0.08% 22.0%) 20 1 (0.00% 100. while quotes with spreads of more than 10 basis points make up less than 0.5%) 7 68.508 237.163 (8.0%) Note: Spreads converted into basis points Table 12: Frequency Distribution of Spreads USD-SGD April-June 2006 The characteristics of the frequency distribution of spreads have also changed almost completely over 15 years.02% 20.7%) 10 < .01% of the entire spectrum of quotes.090 (8. we observed a dramatic increase in volume of USD-SGD quotes over the period of April to May 2009. a.604 190.87% 17.0%) 15 17 (0. although we noted earlier that the volume growth 41 .496 237.2%) 5<.582 (8.9%) 10 155.c.0%) 15 < .603 32 1. Most of the spreads recorded are either 10 basis points or lower. as compared to the same period in 2006. < 20 15 (0.22% 22.5%) 5 224.09% 17.466 (25.075.634 0. followed by 10 basis points.72% 0.3%) 7 < . 0 (0.055 217.763 (37.0%) 20 < . Volume of Quotes DAY Sun Mon Tue Wed Thu Fri Sat Grand Total USD/SGD 9.377 (11.

the spread remained unchanged 9.51% 11. When prices move downwards.13% PRICE UP PRICE SAME PRICE DOWN GRAND TOTAL Table 14: Frequency of No Spread Change for USD-SGD April-May 2009 42 . b. when prices move upwards. Now.23% 27.46% of the time in 2006. The volume of quotes of the USD-SGD has grown from over 604. OF TICKS SPREAD SPREAD SPREAD UP SAME DOWN 28.43% 44.13% DIVIDED BY TOTAL NO. PRICE UP PRICE SAME PRICE DOWN Grand Total SPREAD UP 304279 59290 120972 484541 NO.had been exponential since 2002.43% 44.44% 10.17% of the time compared to 14.17% 25.69% 5. OF QUOTES SPREAD SAME 48129 18050 45055 111234 SPREAD DOWN 120968 53790 295733 470491 PRICE UP PRICE SAME PRICE DOWN GRAND TOTAL DIVIDED BY NO OF TICKS MOVED SPREAD SPREAD SPREAD UP SAME DOWN 64.28% 10.76% 64.99% of the time in 1989.04% 11.44% 10.55% 45. to exceeding 1 million quotes over a 2 month period.35% 4.20% 9.21% 13.54% 4. Frequency of No Spread Change We now observed that in 2009. the spread remained unchanged 10.35% 5.000 ticks. the tendency for spreads to remain unchanged when prices moved is reduced again compared to 2006.74% 45.04% 45.76% of the time compared to 15.76% 41.02% 26.56% 1.

684 (1.00 38.00 < . 43 .00 < .484 (0.158 (0.00 2. < 11.354 (5.00 6. < 10.572 (5.00 13 (0. < 6. This could indicate advancement in the market’s ability to evaluate foreign exchange risk.0%) 2 13.422 (0.7%) 3.5 basis points respectively.6%) Note: Spreads converted into basis points Table 15: Frequency Distribution of Spreads USD-SGD April-May 2009 The characteristics of the frequency distribution of spreads have changed again over 3 years.932 (0.080 (0.6%) 3.828 (15.748 (1.00 < .00 13.1%) > 11.00 < . < 4.903 (0. Two observations are of noteworthy regarding the 2009 dataset.4%) 7.50 6.304 (1.00 < .50 < .577 (25.432 (5. < 3.696 (3.2%) 4. followed by 10 basis points.3%) 7 165. the most common spread had risen to 8 basis points.20 < . Frequency Distribution of Spreads Frequency Distribution of Spreads USD/SGD Spread <2.0%) 4 88.c.0%) 8. < 8.5%) 10.00 10.1%) 9 39. The most common spread in 2006 was 5 basis points.3%) 2.1%) 11 11.2 20. In 2009.00 8.1%) 10 91.5 114.1%) 6.407 (1.00 < . the most common spreads are no longer “significant” numbers such as 5 or 10.2%) 8 268.6%) 6 54.00 579 (0. Most of the spreads recorded are still either 10 basis points or lower.5%) 5.00 < .00 2.3%) 5 59. followed by 7 basis points and 3.441 (8. < 9.1%) 3. but quotations with spreads of more than 10 basis points have increased to slightly under 5% of the entire spectrum of quotations. First.00 830 (0.00 < .9%) 3.613 (8.907 (10.882 (1.8%) 3 54. < 3.7%) 9. < 7.00 1.789 (3. < 5.

We then used EViews to estimate the parameters for each dataset and the full results are attached in 44 .3 GARCH Analysis The primary purpose of the GARCH estimation was to create proxies for the conditional variance of the exchange rate to be used in the investigation of the determinants of the spread.and hence being able to price bids and asks more accurately.1) model: (7) where It-1 denotes the time t-1 information set. Both phenomena could warrant future research. This could be attributed to the increased amount of uncertainty in the financial markets during that period of global economic crisis. Recall from Section 3. so that A. . Following B&M. . Second. and  are the parameters to be estimated.t2 provides an estimate of the price volatility between quotes. we first removed all weekend quotes. and .01% in 2006 to slightly under 5% in 2009. 3.1 that. we estimate for each dataset with the following MA(1)-GARCH(1. The time t subscript refers to the place in the order of the series of quotes. following B&M. the volume of quotes with spreads of more than 10 basis points increased from 0. .

736 304.0052) -0.3281 (0.0003) 0. This provides confidence that the EViews results for the rest of the datasets are appropriate for comparison.5953 (0. The second to fifth columns contain the EViews GARCH estimates for the parameters for the purchased USD/DEM 1989.0006) 0.0006) 0. and USD/SGD 2009 data respectively.0030) 0.0025 (0.0013) SGD1989 0.6540 (0.9197 (0.9190 0.0256) 0.0002) 0. 45 .006.0453 (0.9708 0.0003) 0.0138) SGD2006 0.7802 (0.0004) SGD2009 0.0053) 0.0081) 0.7817 (0.9327 (0.355 0.0008) 0.0018) 0.9025 (0.0069 (0.748 8.0632 (0.0652 (0.0014) -0.9888 0.0003) -0.0500 (0.1008 (0.9710 1. USD/SGD 1989.0065 (0.6833 (0.0561 (0.0452 (0.0009) -0.9057 (0. for 12 weeks). Here we present a summary of the results: B&M DEM1989       0.5867 (0.0009) 0. though standard errors are significantly lower. First we observe that the EViews estimates for the purchased USD/DEM dataset compares very well with B&M’s original results.0002) 0.0042) DEM1989 0.0075) 0.0003) 0.0513 (0.2650 (0.0005) 0.0114) -0.9657 551.608 308.0023 (0.444  Asymptotic errors are reported in parenthesis Table 16: GARCH Estimates for all Datasets The first column contains the average of the weekly estimates from the original B&M dataset (they only estimated these parameters per workweek. USD/SGD 2006.Appendix 1.0578 (0.

the spread did not change 90. To this end.” The GARCH effects for all datasets are all highly significant. Comparing GARCH effects between the USD/DEM 1989 dataset and USD/SGD 1989 dataset. see Lo and MacKinlay (1990) for a formal analysis. we used EViews to obtain GARCH variance series for each dataset. This is expected. the results show that the GARCH effects and persistence of volatility is consistent with this phenomenon.18% of the time. 46 . it is clear that the USD/DEM dataset shows much stronger effects.The second observation is that all estimates for  are negative. and stronger + volatility persistence. As the Singapore economy grows and the USD/SGD becomes a significant regional currency. as the USD/DEM dataset contains at least 36 times more observations and from Table 8. which corresponds to B&M’s results. The primary purpose of the GARCH estimation was to create proxies for the conditional variance of the exchange rate to be used in the investigation of the determinants of the spread. and they noted that “the negative estimates for  may be partly attributed to a non-synchronous quoting phenomenon.

2955) 1.748 8.1299 (102.6436) SGD1989 0.1592) -0.0091 (10.444 551.2342) SGD2006 0.4684) 308.3298) -0.4395 (169.8160 (23.1306 -(69.4 Ordered Response Analysis Recall from Section 3.6719) SGD2006 0. but a summary is presented below: DEM1989   0.355  Z-statistics are reported in parenthesis 1.736 Table 18: Ordered Logit Estimates for all Datasets (4 ordered indicator values) 47 .3568) 0.4622) 0.0464 (29.3.0988 (113.4330 (108.2622 (177.8863) 0.748 8.2412) SGD2009 0.0214 (73.t2 + 2Kt-1 +  (13) We first use four ordered indicator values and B&M’s definitions for aj’s: a1: <= 5 a2: 5 < .2146 (109.2040 -(66.006.0802 (28.006.9306) SGD2009 0.1 that we estimate for each dataset with the following ordered response model (probit and logit): Kt* = 1A.3116 (135.0231 (9.0422 (22.1729) 308.355  Z-statistics are reported in parenthesis 1.444 551.7912) 0.9604) SGD1989 0.736 Table 17: Ordered Probit Estimates for all Datasets (4 ordered indicator values) DEM1989   0. < 10 a3: = 10 a4: > 10 The results for the above parameters for all four datasets are attached in Appendix 1.5740) 0.

48 . It is interesting to note the negative 2 values for the USD/SGD 2006 dataset.The positive 1 coefficients for both ordered probit and logit analyses above suggest that there is a significantly positive effect of exchange rate volatility on the spread for all datasets.18% of the time for the USD/SGD in 1989.t2. the 2 values show that the dependence on the previous spread appears to be more significant for the USD/SGD than for the USD/DEM. the spread remained the same 90. although both 1 values are statistically significant. This is consistent with the implications drawn from B&M’s theoretical model. Comparing the USD/DEM and USD/SGD in 1989. coincidentally matching the growth in quote volume. but it might be due to seasonality. The conditional mean of Kt* is an increasing function of A. For the same reason. as noted above. After all. the magnitude of the 1 values almost quadrupled. The estimates for 2 are indicative of intra-day persistence in the spread process. As the country grow in economic significance and the SGD becoming a major regional currency over 17 years. The magnitude of 1 for each dataset supports what we intuitively already know. volatility appears to play a much larger role in determining the size of the spread for the USD/DEM.

7664) SGD2006 0.748 8.0911) 0. < 4 a5: 6 <= . together with the larger distribution of spreads as shown in Table 15.0345 (90.1741) SGD1989 0.8967) 308. the results for the above parameters for all four datasets are attached in Appendix 1.736 Table 19: Ordered Probit Estimates for all Datasets (10 ordered indicator values) 49 .0296 (19.006. due to the larger distribution of spreads as seen in the 2006 and 2009 data. < 6 a7: 8 <= . We repeated the ordered response analyses using 10 ordered indicator values.5758) 0.5938) 0. < 7 a8: 9 <= .2017) SGD2009 0. < 10 a3: 4 <= . but a summary is presented below: DEM1989   0.2933 (16.0964 (117. The much higher 1 values for this dataset.444 551.0404 -(61.0106 (12. in midst of the worldwide financial crisis. < 8 a9: = 10 Again.0884 (212. < 5 a6: 7 <= .1124 (122. supports the theory that the bid-ask spreads of foreign exchange rates is very much positively related to the underlying volatility. we observed a higher level of volatility in the USD/SGD. < 9 a10: > 10 a2: 3 <= .355  Z-statistics are reported in parenthesis 1. The aj’s are: a1 : < 3 a4: 5 <= .In 2009.0766) -0.

When prices move up. Previously we were examining the relationship between bid-ask spreads and the underlying volatility of exchange rates and found that positive relationships generally exists between the two.444 551.5637 (16.1599 (100. they tend to narrow. To examine this we repeat the ordered response analyses by including the Returns variable from the MA(1)-GARCH(1.0305 (10.1) model into the ordered response model: 50 .748 8. 11 and 14) an interesting phenomenon regarding the behavior of bid-ask spreads of foreign exchange rates.5 Incorporating Returns Recall that we noted (see Tables 5. and the bid-ask spreads change. Conversely.006.8283) SGD2006 0.2591) SGD1989 0.1246) SGD2009 0.DEM1989   0. the conclusions from the previous ordered response analyses with 4 ordered indicator values still holds after we attempt to be more discerning with the ordered indicator values. and the bid-ask spreads change.3382) 0.0755 -(67.736 Table 20: Ordered Logit Estimates for all Datasets (10 ordered indicator values) Generally. when prices move down.355  Z-statistics are reported in parenthesis 1.0571) 308. We are now curious if there is any relationship between bid-ask spreads and price movement.1493 (205.1557 (115.9948) 0. they tend to widen.0599 (22.7863) 0.7295) -0. 3.2047 (110.

0215) 0.2662) 1. The results are attached in Appendix 1.4158 (437.7702) 0.2606 (146.1796 (12.1127) 0.444 551.6830) SGD1989 0.0970 (98.2152 (409.006.0827 (12.3205 (174.6602) 1.2225 (96.0325 (19.1288) SGD2009 0.6876) 308.4338) 0. but we summarize the results below: DEM1989    0.7724) SGD2006 0.9644 (494.3865 (585.6543) 0.748 8.1228) 1.355  Z-statistics are reported in parenthesis 1.9380 (24.t2 + 2Kt-1 + 310.748 8.006.8530) SGD2009 0.0656) 0.6241 (449.2743) 0.9454) SGD2006 0.6732 (154.736 Table 22: Ordered Logit Estimates (with Returns) for all Datasets (4 ordered values) 51 .6104) 1.355  Z-statistics are reported in parenthesis 1.0383 (16.2520) 1.0093 (10.1713) 308.0485 (14.000*logAt +  (14) We repeated the ordered probit and ordered logit analyses for all four datesets using both 4 and 10 ordered indicator values.0406 (44.0235 (8.5486) 0.444 551.1319 (101.Kt* = 1A.7483) 1.5203 (210.0532 (229.4649) SGD1989 0.736 Table 21: Ordered Probit Estimates (with Returns) for all Datasets (4 ordered values) DEM1989    0.0148 (51.3419) 0.7718) 0.0090 (23.

3996 (468.2532 (146.6059 (17.0317 (10.1592 (113.0012 (298.736 Table 23: Ordered Probit Estimates (with Returns) for all Datasets (10 ordered values) DEM1989     0.DEM1989    0.9157) 0. Even when taking Returns into account.3672) 0.0032 -(3.006.3873) 1.7732) SGD1989 0.748 8.5108) 308.0673 (21.6836) 0.736 Table 24: Ordered Logit Estimates (with Returns) for all Datasets (10 ordered values) In general (except for Ordered Probit estimates for 10 ordered values).1677 (11.0350 -(22.355 1.6917) SGD2009 -0.0949 (188.6273) 0. B&M’s theory still holds.3375) 0.4841) 1.0108 (13.0961 (105.748 8.8439) 0.3314) 0.5016) 0.0745 (11.0041) 0.4871) 308.0205 (54.5755 (574.2688 (124.5621) 0. The results for 3 values show that a significant positive relationship also exists between the size of the bid-ask spread and the direction of the exchange rate movement.1891 (434.006.0042 (4.4030 (251.3182 (173.7207) 0.5891) 0.6241) SGD2006 0.0826) SGD2006 -0. the results for 1 and 2 values are consistent with the results in the previous section.355  Z-statistics are reported in parenthesis 1.444 551.3135 (17.2477) 0.2835 (654.1859) SGD2009 0.0252) SGD1989 0.9007) 0.444 551.0185 (22. 52 .

the observation that the 1 values appear to decrease with increasing tick volume runs counter to the rest of the ordered response results. Their empirical subject was over 300. 53 . is likely to be the cause of this results. The USD/DEM was obviously one of the most traded currency pair during that time.d. 3.d. and both the USD and DEM belonged to floating exchange rates regimes of two of the top three largest world economies in 1989.6 Summary B&M provided the theoretical framework and showed empirical evidence that the size of the bid-ask spread of exchange rates is positively related to the exchange rate’s underlying volatility. One possible explanation for this is that multiplicative heteroskedasticity. standard normal.000 USD/DEM quotes from April to June of 1989. Not only does this violate B&M’s prediction. the results for the ordered logit analyses are probably more meaningful. which our ordered response models omitted but has been shown to be significant by B&M.One interesting result to note are the negative 1 and 2 values for the ordered probit analyses of 2006 and 2009 for 10 ordered indicator values (see Table 23). Also since we know that the error term of equations 13 and 14 are not i.

When prices move up. and again in 2009 when the world economy was somewhat in disarray. if the currency belonged to a small country and from a semi-floating exchange rate regime – the USD/SGD. We found that the relationship between bid-ask spreads and price direction is positive and significant.We then wanted to see if both theory and empirical results would hold for a much lesser traded subject currency. they tend to widen. Along the way we also noted the interesting phenomenon regarding the behavior of bid-ask spreads of foreign exchange rates. Conversely. We applied a modified version of B&M’s methodology on the USD/SGD from the same months of April to June 1989 and found that the evidence (as documented in the previous section) supported their theory. We repeated the analyses on the datasets and included Returns as additional variable in the ordered response models. Furthermore. We found that B&M’s theory was still supported by the evidence. we repeated the analysis on the USD/SGD 17 years later in 2006 when the world economy was at a point of stability. and the bid-ask spreads change. 54 . and the bid-ask spreads change. when prices move down. they tend to narrow.

and also 55 . and we found that the effects of the tested parameters grew in strength as that currency grew in economic significance in its region. We included returns as additional variable in the ordered response models and found that the relationship between bid-ask spreads and price direction is positive and significant. Future work in this area should attempt to build a model which links bid-ask spreads not only to volatility and also returns. The generally “better-behaving” results from the Logit models suggests that the disturbances are indeed not standard normal. The results generally hold. We first touched on those which analyses bid-ask spreads and asymmetric information. We also noted the interesting phenomenon regarding the behavior of bid-ask spreads of foreign exchange rates. We departed from B&M’s methodology by omitting multiplicative heteroskedasticity. We provided a simple survey on microstructure literature.4 Conclusion This paper has set out to empirically test B&M’s theory against a lesser-traded currency from a developing country with a managed floating rate regime. and that future work in this area using the same methodologies should include multiplicative heteroskedasticity. but we tested our data using both Ordered Probit and Ordered Logit models.

one of two “hallmarks” of the microstructure approach. 56 .presented snapshots of research involving bid-ask spreads and volatility. Bid-ask spreads. remain a popular theme for research today.

. (1989). pp.. “The Effect of Risk on the Market Maker’s Spread”. I. A. pp..” Journal of Finance.. 30. “Trading Patterns and Prices in the Interbank Foreign Exchange Market. 189-223 Aliber. Financial Analysts Journal. “Monetary Independence under Floating Exchange Rates.R. pp.” Journal of International Money and Finance. “Transactions Costs and Vehicle Currencies. and Logue.” 57 . (1983). M.. “Market Microstructure Effects of Government Intervention in the Foreign Exchange Market..” Journal of Econometrics. D. R. (1993)..W. P.. (1992). W. A. R. and Pfleiderer. (1994).E. A. and Kroner. D. pp. (1991)..” Review of Financial Studies.. 4. 3-40 Admanti.. S. “Dealership Market: Market-making with Inventory”. P. 485-492 Bossaerts. 48. and Domowitz. 45-49 Black. 12-14 Barnea. (1975). K.. pp. 31. 5-59 Bollerslev. and Mendelson. and Melvin. (1988). T. 52.” Review of Financial Studies. “ARCH Modeling in Finance: A Review of the Theory and Empirical Evidence. 2.References Admanti. T. pp. and Hillion. “Divide and Conquer: A Theory of Intra-day and Day of the Week Effects.” Review of Financial Studies. 1. 1421-1443 Bollerslev.Z. Chou. 513-541 Copeland. T. Journal of Financial Economics. (1991). 31-53 Bagehot. “Bid-ask Spreads and Volatility in the Foreign Exchange Market – an Empirical Analysis. 512-526 Bollerslev. 22. H. (1971). 26. pp. and Galai. (1975).” Journal of Finance. pp. 355-372 Boothe. 365-376 Amihud. P.” Financial Analysts Journal. Y. pp. 36.” Economic Inquiry. “Information Effects on the Bid-Ask Spread. (1988).F. P. and Pfleiderer. “A Theory of Intraday Patterns: Volume and Price Variability. T. pp. P. “The Only Game in Town. 10.” Journal of International Economics. “Exchange Rate Risk and the Bid-Ask Spread: A Seven Country Comparison. (1980). pp.Y. pp.

.. “Intertemporal Price Discovery By Market Makers: Active Versus Passive Learning. H. pp. Quarterly Journal of Economics. 6. (2007). Hollified. C. (1985). 33-53 Evans. 10. C. National Bureau of Economic Research..S. pp. and Madhavan. C. 23-52 Goodhart.R.J.” FMG Discussion Papers dp71. 82. “News and the Foreign Exchange market.E. and Sofianos. C. 48. pp. “Financial Intermediation And The Costs Of Trading In An Opaque Market.. (1993). 207-235 Leach. (1987). pp. 14. (1992). and Madhavan. 1565-1593. Financial Markets Group Green. 1457-1469 Demsetz. pp.” NBER Working Papers 13151..N. N. 38. “Every Minute Counts in Financial Markets..J. M. R. 409-428 Glassman. “The Cost of Transacting”. Inc. “The Trades Of Market Makers: An Empirical Analysis Of NYSE Specialists.R. “Exchange Rate Fundamentals and Order Flow.. and Stoll.Journal of Finance. “Optimal Dealer Pricing under Transactions and Return Uncertainty. 757-84 Leach. B.” Journal of Finance. and Figliuoli. pp. 479-490 Glosten. pp. Ho. 6.” Journal of International Money and Finance.A. “Exchange Rate Flexibility and the Efficiency of Foreign Exchange Markets. “Bid. (1975). International Center for Financial Asset Management and Engineering Hasbrouck.A. “Exchange Rate Risks and Transaction Costs: Evidence from Bid-Ask Spreads. L. 10. 9. T..N. “Price Experimentation And Security Market Structure”. (1993). and Lyons. J. pp. and Milgrom. Fieleke. (1991). pp. (1981).” Journal of Financial Intermediation. Ask and Transaction Prices in a Specialist Market with Heterogeneously Informed Traders.” Journal of Financial and Quantitative Analysis.. 375-404 58 . (1990). pp. N. D. A. R. 2. Review of Financial Studies.” Journal of International Money and Finance.E. 71-100 Goodhart. L.” Journal of Financial Economics..” FAME Research Paper Series rp130.” Journal of Financial Economics. A.A. (2005). P. and Schurhoff.. H (1968). G.

and Mackinlay. The MIT Press. A. UK 59 . 39. “Trade Disclosure Regulation In Markets With Negotiated Trades. 291-302 Payne.” Economic Inquiry. 319-379 Lyons. L. S. (2002) The Exconomics of Exchange Rates Cambridge University Press.F. “An Ordered Probit Analysis of Transaction Stock Prices. Journal of Financial Economics. (1995). Cambridge. R.Lo.” Journal of Econometrics. pp. pp. (1992). and Viswanathan. pp.C. 45. (1999). Universität Hannover. “Filtering and Forecasting with Misspecified ARCH Models I: Getting the Right Variance with the Wrong Model..W. R. MA Melvin. 61. A.” Oxford Economic Papers. R. “Foreign Exchange Market Bid-Ask Spreads and the Market Price of Social Unrest. 31.” Journal of International Economics.” Review of Financial Studies.. “ARCH Models as Diffusion Approximations.. A.B. pp. Neuberger. A. D. Transaction Charges. “Price Discovery in Currency Markets.C. 45. K. (1990). pp. and the Market for Foreign Exchange Services. A. 7-38 Nelson. Oxford University Press. and Menkhoff..Y. pp. “Informed Trade In Spot Foreign Exchange Markets: An Empirical Investigation. “Risk. 12. 52. 307-329 Sarno. (1996). 181-210 Lo.. 873-900 Nelson. S. Wirtschaftswissenschaftliche Fakultät.. 61-90 Osler.. “An Econometric Analysis of Non-Synchronous Trading.” Journal of Financial Economics. Mende. (2001) The Microstructure Approach to Exchange Rates.. A.W.. and Tan. N. 20.” Journal of Econometrics. (1992). 48(2). D. and MacKinlay. pp.B. “Tests of Microstructural Hypotheses in the Foreign Exchange Market”. L.” Journal of Econometrics. 321-351 Lyons.H. Overturf. (2003). and Taylor M.” Diskussionspapiere der Wirtschaftswissenschaftlichen Fakultät der Universität Hannover dp-351. (1990). M. pp. C.. Cambridge. (1982). 329-41 Monetary Authority of Singapore Annual Report 2005/2006 Monetary Authority of Singapore Annual Report 2008/2009 Naik. (2006). pp.

078464 1313837 -623880.00029 0.1 Dataset 1: USD-DEM April 1989 to June 1989: MA(1)-GARCH(1.1623 192.7) GARCH = C(3) + C(4)*RESID(-1)^2 + C(5)*GARCH(-1) Variable C MA(1) Coefficient 0.00087 2.001186 0.Normal distribution Sample (adjusted): 2 304131 Included observations: 304130 after adjustments Convergence achieved after 86 iterations MA Backcast: 1 Presample variance: backcast (parameter = 0.000523 0.1 A.055817 0.237055 0.Appendix 1 – EViews Results A.379553 4.ARCH (Marquardt) .102807 Std.942352 0. of regression Sum squared resid Log likelihood Durbin-Watson stat Inverted MA Roots 0.E. Error 0.586985 Variance Equation C RESID(-1)^2 GARCH(-1) R-squared Adjusted R-squared S.000278 87.0911 Prob.1.D.04555 0.934608 0.51 0 0 0 0.518723 -455. dependent var Akaike info criterion Schwarz criterion Hannan-Quinn criter.237053 2.102756 4.59 0.102931 4. 0 0 Mean dependent var S.00129 z-Statistic 6.6 1. 60 .773 3358.1) Results: Dependent Variable: 10000*DLOG(ASK) Method: ML .007732 -0.1.1.

7163 0 0 0 2.077487 2.007509 32.3568 Prob.077312 -315881.557843 2.633627 0.021372 0.005775 0.041442 2.Ordered Probit (Quadratic hill climbing) Sample (adjusted): 2 304131 Included observations: 304130 after adjustments Number of ordered indicator values: 4 Convergence achieved after 4 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) Coefficient 0.005798 0.311626 Limit Points LIMIT_2:C(3) LIMIT_3:C(4) LIMIT_4:C(5) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.1.002302 z-Statistic 73.98297 96.21363 350. log likelihood 61 .29 0 0.96036 135.A.1 -1.190472 0. log likelihood Avg.03864 Std.2 Ordered Probit Results: 4 Ordered Values Dependent Variable: AJ Method: ML . Error 0.5 -329538.1.077363 27313. LR statistic Prob(LR statistic) 0.000289 0. 0 0 Akaike info criterion Log likelihood Restr.

34188 Prob.041202 2.0779 27155.569596 -0. log likelihood 62 .096973 Limit Points LIMIT_2:C(4) LIMIT_3:C(5) LIMIT_4:C(6) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.077839 -315960. LR statistic Prob(LR statistic) -0.038314 0.46485 16.002355 0.204094 1.A.0389 Std.078049 2.39776 247.000986 z-Statistic 51.5462 0 0 0 2.Ordered Probit (Quadratic hill climbing) Sample (adjusted): 2 304131 Included observations: 304130 after adjustments Number of ordered indicator values: 4 Failure to improve Likelihood after 12 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) 10000*DLOG(ASK) Coefficient 0.007386 -97. Error 0.1.6 -329538. 0 0 0 Akaike info criterion Log likelihood Restr.005843 0.828335 0.1.1 -1.005933 0.47589 -34.04 0 0.014821 0.000288 0.26616 98.3 Ordered Probit (with Returns) Results: 4 Ordered Values Dependent Variable: AJ Method: ML . log likelihood Avg.

2094 -159.2254 100.2579 327.006929 0.000382 0.7 -361329 -1.756869 -1. log likelihood Avg.258734 -347769.008505 -128. log likelihood 63 .693844 0.006929 0.034469 0.1. 0 0 Akaike info criterion Log likelihood Restr.85067 47.928 0 0 0 0 0 0 0 0 0 2.006924 0.112367 Limit Points LIMIT_2:C(3) LIMIT_3:C(4) LIMIT_4:C(5) LIMIT_5:C(6) LIMIT_6:C(7) LIMIT_7:C(8) LIMIT_8:C(9) LIMIT_9:C(10) LIMIT_10:C(11) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.000917 z-Statistic 90.010998 0.006894 0.01721 0.666276 0.538779 0.037526 2.788901 0.17413 122.259114 2.129331 Std.1.7475 -162.Ordered Probit (Quadratic hill climbing) Sample (adjusted): 2 308749 Included observations: 307943 after adjustments Number of ordered indicator values: 10 Convergence achieved after 7 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) Coefficient 0.4 Ordered Probit Results: 10 Ordered Values Dependent Variable: AJ Method: ML .8258 96.64 0 0.258844 27118.00944 0.329696 0.206482 -1. Error 0.694678 2.006894 0. LR statistic Prob(LR statistic) -2.1396 100.9984 46.5758 Prob.A.322974 0.

1.55753 -35.007086 0.43857 216.56097 -39.02521 4.6283 0 0 0 0 0 0 0 0 0 2.251959 -0.034829 2.000938 0.277872 -0.020522 0.008402 -178.41 0 0. 0 0 0 Akaike info criterion Log likelihood Restr.132496 Std.017039 0.007086 0.00091 z-Statistic 54. log likelihood Avg.265485 2.006997 0.251127 1.006998 0.096148 Limit Points LIMIT_2:C(4) LIMIT_3:C(5) LIMIT_4:C(6) LIMIT_5:C(7) LIMIT_6:C(8) LIMIT_7:C(9) LIMIT_8:C(10) LIMIT_9:C(11) LIMIT_10:C(12) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.393098 -0.00038 0.1147 -86. Error 0.605993 -2.26519 25169.A.007077 0.26293 -35.501637 105.3 -361329 -1.7287 -251.5 Ordered Probit (with Returns) Results: 10 Ordered Values Dependent Variable: AJ Method: ML .04669 -2.1. LR statistic Prob(LR statistic) -3.605194 -0.598756 -0.00953 0.49427 -85.004221 0.812 -236.011008 0.6836 Prob.820135 0.Ordered Probit (Quadratic hill climbing) Sample (adjusted): 2 308749 Included observations: 307943 after adjustments Number of ordered indicator values: 10 Failure to improve Likelihood after 9 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) 10000*DLOG(ASK) Coefficient 0.26507 -348744. log likelihood 64 .

2955 Prob.6037 328.003998 z-Statistic 102.6436 108.129908 0.96 0 0.1.6 Ordered Logit Results: 4 Ordered Values Dependent Variable: AJ Method: ML . LR statistic Prob(LR statistic) 0.05487 34153.A.009948 0. 0 0 Akaike info criterion Log likelihood Restr.051821 2.0225 0 0 0 2. Error 0.Ordered Logit (Quadratic hill climbing) Sample (adjusted): 2 304131 Included observations: 304130 after adjustments Number of ordered indicator values: 4 Convergence achieved after 4 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) Coefficient 0.001266 0.009847 0.1 -329538. log likelihood 65 .483887 1.14077 110.1 -1. log likelihood Avg.432993 Limit Points LIMIT_2:C(3) LIMIT_3:C(4) LIMIT_4:C(5) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.1.05482 -312461.100335 4.014927 49.054994 2.027393 Std.89648 0.

004354 0.31565 152.6104 174.010468 0. 0 0 0 Akaike info criterion Log likelihood Restr.001297 0.131878 0. LR statistic Prob(LR statistic) 0.3301 341.955897 1.7 Ordered Logit (with Returns) Results: 4 Ordered Values Dependent Variable: AJ Method: ML .94 0 0.3 -329536.673158 0.016938 91. Error 0.944327 1. log likelihood 66 .5674 0 0 0 1.0656 Prob.78533 0.944178 67823.320523 Limit Points LIMIT_2:C(4) LIMIT_3:C(5) LIMIT_4:C(6) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.Ordered Logit (Quadratic hill climbing) Sample (adjusted): 3 304131 Included observations: 304128 after adjustments Number of ordered indicator values: 4 Convergence achieved after 5 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) 10000*DLOG(ASK) Coefficient 0.626336 5.010676 0.1.001841 z-Statistic 101.2 -0.1.944117 -295624.683 154.102908 1.972039 Std.A. log likelihood Avg.

0075 60.011771 0.155742 0.2591 100. log likelihood 67 .6399 112.6803 0 0 0 0 0 0 0 0 0 2.762068 -3.137085 -3.158783 0.1.23552 2. Error 0.053891 0.011771 0.7518 313.3013 -141.1.159926 Limit Points LIMIT_2:C(3) LIMIT_3:C(4) LIMIT_4:C(5) LIMIT_5:C(6) LIMIT_6:C(7) LIMIT_7:C(8) LIMIT_8:C(9) LIMIT_9:C(10) LIMIT_10:C(11) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.23525 34384.341132 5.98525 61. LR statistic Prob(LR statistic) -5.011879 0.29396 1. log likelihood Avg.2 0 0.8 Ordered Logit Results: 10 Ordered Values Dependent Variable: AJ Method: ML . 0 0 Akaike info criterion Log likelihood Restr.95536 108.23514 -344136.022402 0.016561 -95.717831 0.001594 z-Statistic 115.04758 2.9279 112.028652 0.729267 1.339747 1.3382 Prob.011894 0.9 -361329 -1.Ordered Logit (Quadratic hill climbing) Sample (adjusted): 2 308749 Included observations: 307943 after adjustments Number of ordered indicator values: 10 Convergence achieved after 7 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) Coefficient 0.001351 0.011895 0.194799 0.32363 -131.117534 Std.A.

37753 -113. 0 0 0 Akaike info criterion Log likelihood Restr.3644 104. log likelihood 68 . Error 0.663842 -3.286633 -2.985893 6.9311 -118.2477 Prob.984385 1.2 -361329 -1. LR statistic Prob(LR statistic) -4.028848 0.5792 155.3672 173.3148 330.1.2614 105.012554 0.9 Ordered Logit (with Returns) Results: 10 Ordered Values Dependent Variable: AJ Method: ML .3986 0 0 0 0 0 0 0 0 0 2.012556 0.7732 146. log likelihood Avg.053994 0.66 0 0.018747 -86.001399 0.308867 1.681408 1.093349 2.063831 Std.127739 -327599.93451 1.012786 0.321246 1.318154 Limit Points LIMIT_2:C(4) LIMIT_3:C(5) LIMIT_4:C(6) LIMIT_5:C(7) LIMIT_6:C(8) LIMIT_7:C(9) LIMIT_8:C(10) LIMIT_9:C(11) LIMIT_10:C(12) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.012786 0.127859 67459.A.128154 2.012762 0.194035 0.2298 151.00173 0.253184 0.Ordered Logit (Quadratic hill climbing) Sample (adjusted): 2 308749 Included observations: 307943 after adjustments Number of ordered indicator values: 10 Convergence achieved after 7 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) 10000*DLOG(ASK) Coefficient 0.2056 155.15917 0.022654 0.001836 z-Statistic 113.1.

A.1.2 A.1.2.1

Dataset 2: USD-USD April 1989 to June 1989: MA(1)-GARCH(1,1) Results:

Dependent Variable: 10000*DLOG(ASK) Method: ML - ARCH Sample (adjusted): 2 8445 Included observations: 8444 after adjustments Convergence achieved after 37 iterations MA Backcast: 1 Presample variance: backcast (parameter = 0.7) GARCH = C(3) + C(4)*RESID(-1)^2 + C(5)*GARCH(-1) Variable C MA(1) Coefficient 0.045205 -0.3281 Variance Equation C RESID(-1)^2 GARCH(-1) R-squared Adjusted R-squared S.E. of regression Sum squared resid Log likelihood Durbin-Watson stat Inverted MA Roots 0.683349 0.265029 0.653991 0.046483 0.04637 2.549771 54884.24 -18761.29 1.819609 0.33 0.025552 0.008055 0.007513 26.7437 32.9008 87.05354 0 0 0 0.001809 2.611027 4.444882 4.449051 4.446305 Std. Error 0.013771 0.011435 z-Statistic 3.282516 -28.693 Prob. 0.001 0

Mean dependent var S.D. dependent var Akaike info criterion Schwarz criterion Hannan-Quinn criter.

69

A.1.2.2

Ordered Probit Results: 4 Ordered Values

Dependent Variable: AJ Method: ML - Ordered Probit (Quadratic hill climbing) Sample (adjusted): 2 8445 Included observations: 8444 after adjustments Number of ordered indicator values: 4 Convergence achieved after 5 iterations Covariance matrix computed using second derivatives Variable GARCH01 AJ(-1) Coefficient 0.009127 0.942999 Limit Points LIMIT_2:C(3) LIMIT_3:C(4) LIMIT_4:C(5) Pseudo R-squared Schwarz criterion Hannan-Quinn criter. LR statistic Prob(LR statistic) 0.61897 1.002232 4.898957 0.115395 0.585632 0.582885 639.1806 0 0.123303 0.121611 0.134626 5.019921 8.241273 36.38941 0 0 0 0.581462 -2449.933 -2769.524 -0.290139 Std. Error 0.000855 0.042231 z-Statistic 10.67187 22.3298 Prob. 0 0

Akaike info criterion Log likelihood Restr. log likelihood Avg. log likelihood

70

A.1.2.3

Ordered Probit (with Returns) Results: 4 Ordered Values

Dependent Variable: AJ Method: ML - Ordered Probit (Quadratic hill climbing) Sample (adjusted): 2 8445 Included observations: 8444 after adjustments Number of ordered indicator values: 4 Convergence achieved after 5 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) 10000*DLOG(ASK) Coefficient 0.00933 1.008975 0.08268 Limit Points LIMIT_2:C(4) LIMIT_3:C(5) LIMIT_4:C(6) Pseudo R-squared Schwarz criterion Hannan-Quinn criter. LR statistic Prob(LR statistic) 0.775747 1.166192 5.148438 0.142475 0.568939 0.565644 789.1746 0 0.125149 0.12345 0.138288 6.198581 9.446642 37.22976 0 0 0 0.563936 -2374.936 -2769.524 -0.281257 Std. Error 0.000852 0.043056 0.006736 z-Statistic 10.94538 23.43378 12.27428 Prob. 0 0 0

Akaike info criterion Log likelihood Restr. log likelihood Avg. log likelihood

71

463308 0.644992 0.A.837742 0. LR statistic Prob(LR statistic) -1.251299 0.568 -2976.65859 -2769.000829 0.195539 0.925167 -0.1.162967 -4.010588 0.293255 Limit Points LIMIT_2:C(3) LIMIT_3:C(4) LIMIT_4:C(5) LIMIT_5:C(6) LIMIT_6:C(7) LIMIT_7:C(8) LIMIT_8:C(9) LIMIT_9:C(10) LIMIT_10:C(11) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.2452 0 0.07658 Prob.159031 0.661721 413.Ordered Probit (Quadratic hill climbing) Sample (adjusted): 2 8445 Included observations: 8444 after adjustments Number of ordered indicator values: 10 Convergence achieved after 7 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) Coefficient 0.334317 27.77686 0.191 -0.2.91332 3.908998 5.158252 0.327992 Std.843638 4.71307 0.308537 4.018241 z-Statistic 12.069425 0.104586 -3.0036 0. log likelihood 72 .0009 0 0 0 0 0. 0 0 Akaike info criterion Log likelihood Restr.296765 5.225398 0. log likelihood Avg.4 Ordered Probit Results: 10 Ordered Values Dependent Variable: AJ Method: ML .0003 0.158821 0.046148 -0.667763 0.525466 0.7664 16.646689 2.158153 0. Error 0.158161 0.167892 -4.66655 0 0 0.

647126 540.018435 0.92919 0.823252 6.586622 0.159566 0. LR statistic Prob(LR statistic) -0.160333 0.804416 4.643716 -3.320434 Std.197098 0.074509 Limit Points LIMIT_2:C(4) LIMIT_3:C(5) LIMIT_4:C(6) LIMIT_5:C(7) LIMIT_6:C(8) LIMIT_7:C(9) LIMIT_8:C(10) LIMIT_9:C(11) LIMIT_10:C(12) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.524124 -2.Ordered Probit (Quadratic hill climbing) Sample (adjusted): 2 8445 Included observations: 8444 after adjustments Number of ordered indicator values: 10 Convergence achieved after 7 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) 10000*DLOG(ASK) Coefficient 0.160125 0.170386 -3. log likelihood 73 .010826 0.0001 0 0 0 0 0 0.2.745 -2976. log likelihood Avg.159477 0.226892 0. 0 0 0 Akaike info criterion Log likelihood Restr.251929 28.976292 3.0004 0.673407 0.1.56432 0. Error 0.00409 11.205514 5.159469 0.64371 -2705.A.609972 0.0029 0.866967 0.09087 0.991005 0.0003 0.5 Ordered Probit (with Returns) Results: 10 Ordered Values Dependent Variable: AJ Method: ML .313473 0.653717 0.33749 Prob.000828 0.799597 -0.191 -0.214104 6.08258 17.006572 z-Statistic 13.920766 -0.2527 0.996988 4.8901 0 0.

2.584483 625.208446 9.052707 9.285628 0.A. 0 0 Akaike info criterion Log likelihood Restr.587229 0.252771 5.215518 0.Ordered Logit (Quadratic hill climbing) Sample (adjusted): 2 8445 Included observations: 8444 after adjustments Number of ordered indicator values: 4 Convergence achieved after 7 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) Coefficient 0.290938 Std.6 Ordered Logit Results: 4 Ordered Values Dependent Variable: AJ Method: ML .524 -0. log likelihood Avg.677 -2769.023145 1. Error 0.73541 0 0 0 0.234157 23.15919 Prob.078415 z-Statistic 9. LR statistic Prob(LR statistic) 1.6929 0 0.816022 Limit Points LIMIT_2:C(3) LIMIT_3:C(4) LIMIT_4:C(5) Pseudo R-squared Schwarz criterion Hannan-Quinn criter. log likelihood 74 .583059 -2456.151811 2.221143 0.1.11296 0.002506 0.524505 36.

014658 z-Statistic 8. 0 0 0 Akaike info criterion Log likelihood Restr.12276 12.66822 37.524 -0.002679 0.A.565974 -2383.7 Ordered Logit (with Returns) Results: 4 Ordered Values Dependent Variable: AJ Method: ML .Ordered Logit (Quadratic hill climbing) Sample (adjusted): 2 8445 Included observations: 8444 after adjustments Number of ordered indicator values: 4 Convergence achieved after 8 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) 10000*DLOG(ASK) Coefficient 0.772378 24.139367 0.567682 771. Error 0.570977 0. log likelihood 75 .282276 Std.080338 0.1.2.219477 0.023498 1.431605 2.25201 Prob.36562 0 0 0 0.371611 10.542 -2769.224685 0. LR statistic Prob(LR statistic) 1.770045 0. log likelihood Avg.261472 6.937969 0.341427 9.9628 0 0.179594 Limit Points LIMIT_2:C(4) LIMIT_3:C(5) LIMIT_4:C(6) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.

399721 -2.310773 -4. Error 0.662368 407.030535 0.810792 8. LR statistic Prob(LR statistic) -3.285436 0.66841 0.282811 0.845043 0.00282 0.66173 1.756259 0.068507 0.282485 0.0012 0.46145 0 0 0 0.797737 1.72951 Prob.63663 0.8 Ordered Logit Results: 10 Ordered Values Dependent Variable: AJ Method: ML .563651 Limit Points LIMIT_2:C(3) LIMIT_3:C(4) LIMIT_4:C(5) LIMIT_5:C(6) LIMIT_6:C(7) LIMIT_7:C(8) LIMIT_8:C(9) LIMIT_9:C(10) LIMIT_10:C(11) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.Ordered Logit (Quadratic hill climbing) Sample (adjusted): 2 8445 Included observations: 8444 after adjustments Number of ordered indicator values: 10 Convergence achieved after 6 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) Coefficient 0.1.7821 0 0.2.41083 28.77277 5.286528 0.233876 3.282458 0.993661 -2.703637 3.82827 16.3 -2976.875773 6.702359 -4.926596 1.A.659237 -2772.0002 0 0 0 0 0. log likelihood Avg.328316 Std.488651 0. log likelihood 76 .364009 6.191 -0.033692 z-Statistic 10.495444 -4.298438 0. 0 0 Akaike info criterion Log likelihood Restr.076884 1.

108606 -2.112879 2. log likelihood Avg.285722 0.440867 29.811259 6.236004 1.109446 -4.002984 0.908731 7.394317 7.271391 4.605872 0. log likelihood 77 .004677 1.653921 0.2.387414 1.976024 2. LR statistic Prob(LR statistic) -3.64733 539.286018 0. Error 0.62409 17.336144 0.48917 0.289368 0.63687 0.320536 Std.72067 11.38731 Prob.756454 0.098116 4.09058 0.607 -2976.288368 0.285744 0.Ordered Logit (Quadratic hill climbing) Sample (adjusted): 2 8445 Included observations: 8444 after adjustments Number of ordered indicator values: 10 Convergence achieved after 8 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) 10000*DLOG(ASK) Coefficient 0.9 Ordered Logit (with Returns) Results: 10 Ordered Values Dependent Variable: AJ Method: ML .643914 -2706.167709 Limit Points LIMIT_2:C(4) LIMIT_3:C(5) LIMIT_4:C(6) LIMIT_5:C(7) LIMIT_6:C(8) LIMIT_7:C(9) LIMIT_8:C(10) LIMIT_9:C(11) LIMIT_10:C(12) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.014728 z-Statistic 10. 0 0 0 Akaike info criterion Log likelihood Restr.1.1676 0 0.A.191 -0.702375 -2.243211 -4.03419 0.126018 9.349 0 0 0 0 0 0 0 0 0 0.031698 0.318108 -4.

2785 1015.657286 3.000558 0.075531 -911.902484 0.000813 0.78 0.204 Prob.1.3.781719 Variance Equation C RESID(-1)^2 GARCH(-1) R-squared Adjusted R-squared S.002535 -0.3 -910333.15628 113.7) GARCH = C(3) + C(4)*RESID(-1)^2 + C(5)*GARCH(-1) Variable C MA(1) Coefficient 0.302185 3.3 A.2 1.057826 0.313106 950669. dependent var Akaike info criterion Schwarz criterion Hannan-Quinn criter.ARCH (Marquardt) .A.000858 z-Statistic 7.000425 1. Error 0.1) Results: Dependent Variable: 10000*DLOG(ASK) Method: ML .372224 1.1 Dataset 3: USD-USD April 2006 to May 2006: MA(1)-GARCH(1.D.302287 3.000889 71.000358 0. of regression Sum squared resid Log likelihood Durbin-Watson stat Inverted MA Roots 0.1.372225 0.04 0 0 0 -0. 78 . 0 0 Mean dependent var S.302214 Std.E.Normal distribution Sample (adjusted): 2 551356 Included observations: 551355 after adjustments Convergence achieved after 14 iterations MA Backcast: 1 Presample variance: backcast (parameter = 0.97678 0.063219 0.

24124 -69.004577 0.036661 -58. 0 0 Akaike info criterion Log likelihood Restr.122714 -585179. log likelihood 79 .001588 0.4 -587919.624999 0.A.87791 0 0 0 2.004557 0.652 98.3.267391 0.130554 Limit Points LIMIT_2:C(3) LIMIT_3:C(4) LIMIT_4:C(5) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.2 Ordered Probit Results: 4 Ordered Values Dependent Variable: AJ Method: ML .67524 108. LR statistic Prob(LR statistic) -0.851 0 0.3 -1.061348 Std. log likelihood Avg.1.001879 z-Statistic 29.122815 2. Error 0.00466 2.497267 3.122742 5479.046421 -0.Ordered Probit (Quadratic hill climbing) Sample (adjusted): 2 551356 Included observations: 551355 after adjustments Number of ordered indicator values: 4 Convergence achieved after 4 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) Coefficient 0.46221 Prob.

002468 0.700133 -468682.005446 0.850056 Std.4 -587919. LR statistic Prob(LR statistic) 0.803072 1.624067 Limit Points LIMIT_2:C(4) LIMIT_3:C(5) LIMIT_4:C(6) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.520271 0.700167 238473.4683 317. log likelihood 80 .032475 0.3 -0.8 0 0.84724 6. 0 0 0 Akaike info criterion Log likelihood Restr. Error 0.Ordered Probit (Quadratic hill climbing) Sample (adjusted): 2 551356 Included observations: 551355 after adjustments Number of ordered indicator values: 4 Convergence achieved after 6 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) 10000*DLOG(ASK) Coefficient 0.682 0 0 0 1.1.202812 1.12882 210.001388 z-Statistic 19.7702 449.00582 0.3 Ordered Probit (with Returns) Results: 4 Ordered Values Dependent Variable: AJ Method: ML .051481 147.3.001698 0.4066 121.264271 0. log likelihood Avg.700255 1.5486 Prob.A.

1. log likelihood Avg.89255 45.351617 -888807.000655 z-Statistic 19.9 -1.040361 Limit Points LIMIT_2:C(3) LIMIT_3:C(4) LIMIT_4:C(5) LIMIT_5:C(6) LIMIT_6:C(7) LIMIT_7:C(8) LIMIT_8:C(9) LIMIT_9:C(10) LIMIT_10:C(11) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.004831 0.472672 3.20169 -61.001543 0. 0 0 Akaike info criterion Log likelihood Restr.4 Ordered Probit Results: 10 Ordered Values Dependent Variable: AJ Method: ML .029623 -0. LR statistic Prob(LR statistic) -2.36043 0 0 0 0 0 0 0 0 0 3.005154 0.675788 Std.6 -890838.A.005776 0.218854 0.123895 -1.292156 -0.696 0 0.00484 0.076775 0.Ordered Probit (Quadratic hill climbing) Sample (adjusted): 2 551356 Included observations: 530382 after adjustments Number of ordered indicator values: 10 Convergence achieved after 5 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) Coefficient 0.008137 0.770052 -2.351849 3.004827 0.351682 4062.4275 -367.037289 -340.33503 95. log likelihood 81 .5938 Prob.004841 0.00228 3.70278 97.463249 0. Error 0.0033 -60.34619 -15.004841 0.63469 96.693 -299.593189 0.3.541017 -0.

005688 0.56212 298.00565 0.018353 180859.509118 Std.648621 0.Ordered Probit (Quadratic hill climbing) Sample (adjusted): 2 551356 Included observations: 530382 after adjustments Number of ordered indicator values: 10 Failure to improve Likelihood after 17 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) 10000*DLOG(ASK) Coefficient -0.9393 0 0 0 0 0 0 0 0 0 3.69168 22. Error 0.365238 Limit Points LIMIT_2:C(4) LIMIT_3:C(5) LIMIT_4:C(6) LIMIT_5:C(7) LIMIT_6:C(8) LIMIT_7:C(9) LIMIT_8:C(10) LIMIT_9:C(11) LIMIT_10:C(12) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.018282 -800409.5115 -262.005598 0. LR statistic Prob(LR statistic) -2.001223 z-Statistic -22.035043 0.005795 0.018535 3.60886 112.768453 -2.000821 0.3.A.7437 -26.089688 -1.148442 0.025918 -325.008501 0. 0 0 0 Akaike info criterion Log likelihood Restr.6782 -338.01852 0.1.1 -890838.659757 4.6 0 0.006173 0.084178 0.0419 113.5 Ordered Probit (with Returns) Results: 10 Ordered Values Dependent Variable: AJ Method: ML .395925 0.484632 -0.03649 69. log likelihood 82 .101511 3.6273 Prob.74997 15.001544 0.005789 0.041679 0.9 -1.005549 0.8488 155. log likelihood Avg.

41091 0. Error 0.003055 z-Statistic 28.061638 Std.007581 0.004388 2.123397 2.204018 Limit Points LIMIT_2:C(3) LIMIT_3:C(4) LIMIT_4:C(5) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.123324 5159.1.007658 0.00277 0.20336 108.Ordered Logit (Quadratic hill climbing) Sample (adjusted): 2 551356 Included observations: 551355 after adjustments Number of ordered indicator values: 4 Convergence achieved after 4 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) Coefficient 0.7 -587919.830244 9. LR statistic Prob(LR statistic) -0. log likelihood 83 .6 Ordered Logit Results: 4 Ordered Values Dependent Variable: AJ Method: ML . log likelihood Avg.4198 63.21045 0 0 0 2. 0 0 Akaike info criterion Log likelihood Restr.3 -1.080152 -0.135262 0.3.93061 -66.A.321 0 0.79117 Prob.144521 -54.123295 -585339.

6294 0. LR statistic Prob(LR statistic) 1.453473 13.671801 1.600267 3.671678 -460838. log likelihood 84 .009958 0.671713 254162.7016 312.1 -587919.41909 0 0 0 1.7483 Prob.3.7 Ordered Logit (with Returns) Results: 4 Ordered Values Dependent Variable: AJ Method: ML .216154 1.85304 229.053246 1.167398 160.3 0 0.003264 0.011048 0.048475 1.1. log likelihood Avg.6602 409.3 -0. Error 0. 0 0 0 Akaike info criterion Log likelihood Restr.Ordered Logit (Quadratic hill climbing) Sample (adjusted): 2 551356 Included observations: 551355 after adjustments Number of ordered indicator values: 4 Convergence achieved after 6 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) 10000*DLOG(ASK) Coefficient 0.835828 Std.002966 z-Statistic 14.5789 81.A.004586 0.215211 Limit Points LIMIT_2:C(4) LIMIT_3:C(5) LIMIT_4:C(6) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.

008281 0.001114 z-Statistic 22.4939 -290. 0 0 Akaike info criterion Log likelihood Restr.841 0 0.78632 Prob.349936 4988.73051 9.7677 -339.681481 -3. Error 0.008229 0.009268 0.011679 0.002706 0.3095 0.714603 0.1. LR statistic Prob(LR statistic) -5.690845 -0.61812 86.3.059869 -0.3316 -62.Ordered Logit (Quadratic hill climbing) Sample (adjusted): 2 551356 Included observations: 530382 after adjustments Number of ordered indicator values: 10 Convergence achieved after 5 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) Coefficient 0.168038 0.965107 -2.5 -890838.41918 37.34987 -888344.0028 3.12349 -20.008227 0.350102 3.147643 -265.A. log likelihood Avg.008257 0.674915 Std. log likelihood 85 .8 Ordered Logit Results: 10 Ordered Values Dependent Variable: AJ Method: ML .32978 88.008278 0.07549 Limit Points LIMIT_2:C(3) LIMIT_3:C(4) LIMIT_4:C(5) LIMIT_5:C(6) LIMIT_6:C(7) LIMIT_7:C(8) LIMIT_8:C(9) LIMIT_9:C(10) LIMIT_10:C(11) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.021378 0.036028 0.512979 -0.20197 0 0 0 0 0 0 0 0 0 3.12459 -67.9 -1.22038 61.

Error 0.6353 286.3.067262 0.225475 3.846774 3.8036 -165.872262 13.8341 231.1774 329.528082 3.011712 0. log likelihood 86 .9 Ordered Logit (with Returns) Results: 10 Ordered Values Dependent Variable: AJ Method: ML .54475 190.189108 Limit Points LIMIT_2:C(4) LIMIT_3:C(5) LIMIT_4:C(6) LIMIT_5:C(7) LIMIT_6:C(8) LIMIT_7:C(9) LIMIT_8:C(10) LIMIT_9:C(11) LIMIT_10:C(12) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.402995 1.4841 Prob.0968 -66.114505 -0. 0 0 0 Akaike info criterion Log likelihood Restr.9007 434.1 0 0.867156 -760331.A.81305 0 0 0 0 0 0 0 0 0 2. LR statistic Prob(LR statistic) -3.18593 251.6322 81.9 -890838.0016 0.894364 -2.022026 0.04924 2.433555 Std.9 -1. log likelihood Avg.011692 0.012808 0.1.0122 330.867227 261014.010767 0.003175 0.94908 0.011271 0.010738 0.Ordered Logit (Quadratic hill climbing) Sample (adjusted): 2 551356 Included observations: 530382 after adjustments Number of ordered indicator values: 10 Convergence achieved after 6 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) 10000*DLOG(ASK) Coefficient 0.7165 2.010914 0.002737 z-Statistic 21.170499 -176.146499 2.867409 2.

dependent var Akaike info criterion Schwarz criterion Hannan-Quinn criter. 87 .001 Prob.E. Error 0.000386 1.919658 0.00021 0.D.602816 3.A.196957 3.274136 -1276.78 0.1) Results: Dependent Variable: 10000*DLOG(ASK) Method: ML .051345 0.0686 3502. 0 0 Mean dependent var S.272934 1631271 -1609241 1.7) GARCH = C(3) + C(4)*RESID(-1)^2 + C(5)*GARCH(-1) Variable C MA(1) Coefficient 0.002329 -0.045284 0.000251 0.196973 Std.36927 0.970126 0.000263 201.000611 z-Statistic 9.ARCH Sample (adjusted): 2 1006737 Included observations: 1006736 after adjustments Convergence achieved after 21 iterations MA Backcast: 1 Presample variance: backcast (parameter = 0. of regression Sum squared resid Log likelihood Durbin-Watson stat Inverted MA Roots 0.197016 3.369269 1.78024 Variance Equation C RESID(-1)^2 GARCH(-1) R-squared Adjusted R-squared S.1 Dataset 4: USD-USD April 2009 to May 2009: MA(1)-GARCH(1.001 0 0 0 -0.1.4.4 A.1.000225 0.6421 244.

003373 0.50314 532.003801 99.8891 629.8863 177.1.069791 52977. Error 0.4.366 0 0 0 2. log likelihood 88 .262245 Limit Points LIMIT_2:C(3) LIMIT_3:C(4) LIMIT_4:C(5) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.Ordered Probit (Quadratic hill climbing) Sample (adjusted): 2 1006737 Included observations: 1006736 after adjustments Number of ordered indicator values: 4 Convergence achieved after 3 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) Coefficient 0.00148 z-Statistic 113.000867 0.1729 Prob.A. log likelihood Avg.024794 2. 0 0 Akaike info criterion Log likelihood Restr.797373 2.392075 0.003019 0.300383 1.45 0 0.034882 Std.069833 2. LR statistic Prob(LR statistic) 0.098754 0.2 Ordered Probit Results: 4 Ordered Values Dependent Variable: AJ Method: ML .069774 -1041853 -1068342 -1.

Error 0.003441 0.001785 0.040619 0.741812 2. 0 0 0 Akaike info criterion Log likelihood Restr.415817 Limit Points LIMIT_2:C(4) LIMIT_3:C(5) LIMIT_4:C(6) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.7718 146.874046 Std.1 -1068342 -0.7 0 0.748122 376817. log likelihood 89 .1.748173 1.000907 0.Ordered Probit (Quadratic hill climbing) Sample (adjusted): 2 1006737 Included observations: 1006736 after adjustments Number of ordered indicator values: 4 Convergence achieved after 12 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) 10000*DLOG(ASK) Coefficient 0.10602 429.6876 Prob.4.1621 517.3 Ordered Probit (with Returns) Results: 4 Ordered Values Dependent Variable: AJ Method: ML .004675 21.A.072632 1.00095 z-Statistic 44.420409 0. log likelihood Avg.260624 0.0215 437.176356 1.004059 0.7516 0 0 0 1.748103 -879933. LR statistic Prob(LR statistic) 0.

000415 z-Statistic 117.855945 1.002718 0.4852 -62.630826 1.002717 0.142757 0.0911 212.003573 0.9753 170.4.002965 0. 0 0 Akaike info criterion Log likelihood Restr.38151 52.799457 2.2099 0 0 0 0 0 0 0 0 0 4.463947 0. log likelihood 90 .029392 Std.0266 308.8967 Prob.67088 117.47 0 0.000823 0.Ordered Probit (Quadratic hill climbing) Sample (adjusted): 2 1006737 Included observations: 1006736 after adjustments Number of ordered indicator values: 10 Convergence achieved after 4 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) Coefficient 0.002774 0.466867 -0.05884 67525.32063 0.003495 -410. LR statistic Prob(LR statistic) -1.169471 0.00271 0.088372 Limit Points LIMIT_2:C(3) LIMIT_3:C(4) LIMIT_4:C(5) LIMIT_5:C(6) LIMIT_6:C(7) LIMIT_7:C(8) LIMIT_8:C(9) LIMIT_9:C(10) LIMIT_10:C(11) Pseudo R-squared Schwarz criterion Hannan-Quinn criter. Error 0.058805 -2043062 -2076824 -2.0574 685.016257 4.9868 593.058934 4.003034 0.4 Ordered Probit Results: 10 Ordered Values Dependent Variable: AJ Method: ML .5626 549.A.394888 0. log likelihood Avg.002729 0.096416 0.1.

LR statistic Prob(LR statistic) -1.540899 -0. 0.014555 0.003013 0.0001 0 0 Akaike info criterion Log likelihood Restr.2568 -182.5 Ordered Probit (with Returns) Results: 10 Ordered Values Dependent Variable: AJ Method: ML .889585 -0.003187 0.4845 548.0877 413.1. Error 0.A.843887 188.000853 z-Statistic -3.481455 1.8559 0 0 0 0 0 0 0 0 0 3.729748 -1877424 -2076824 -1.729787 398801.000502 0.5108 Prob.094896 0.Ordered Probit (Quadratic hill climbing) Sample (adjusted): 2 1006737 Included observations: 1006736 after adjustments Number of ordered indicator values: 10 Convergence achieved after 9 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) 10000*DLOG(ASK) Coefficient -0.1 0 0. log likelihood Avg.00377 0.399614 Limit Points LIMIT_2:C(4) LIMIT_3:C(5) LIMIT_4:C(6) LIMIT_5:C(7) LIMIT_6:C(8) LIMIT_7:C(9) LIMIT_8:C(10) LIMIT_9:C(11) LIMIT_10:C(12) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.096012 3.8783 453.753915 49.4.003113 0.729889 3.003697 0.09902 -4.003255 0.211211 -0.153294 0.864862 Std.003579 0.002964 0.676514 2.24206 189.40264 0. log likelihood 91 .000829 0.9157 468.003062 0.615502 1.004378 -501.4773 -70.

002593 z-Statistic 109.301442 0.070184 -1042059 -1068342 -1.0702 52565.035087 Std.01 0 0.006332 0. log likelihood Avg. log likelihood 92 .001959 0.007555 111.1.574 169.3912 491. Error 0.114943 4.Ordered Logit (Quadratic hill climbing) Sample (adjusted): 2 1006737 Included observations: 1006736 after adjustments Number of ordered indicator values: 4 Convergence achieved after 4 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) Coefficient 0.9238 569.A. 0 0 Akaike info criterion Log likelihood Restr.4684 Prob.43948 Limit Points LIMIT_2:C(3) LIMIT_3:C(4) LIMIT_4:C(5) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.214626 0.070243 2.024601 2.4. LR statistic Prob(LR statistic) 0.6 Ordered Logit Results: 4 Ordered Values Dependent Variable: AJ Method: ML .00533 0.3761 0 0 0 2.593662 3.

LR statistic Prob(LR statistic) 2. Error 0.532543 593849.11273 494.27793 1.4.964362 1.6543 585.38648 Limit Points LIMIT_2:C(4) LIMIT_3:C(5) LIMIT_4:C(6) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.010304 0.003971 0.002369 z-Statistic 96.002315 0.A.796605 8.1713 Prob.1. 0 0 0 Akaike info criterion Log likelihood Restr.7 Ordered Logit (with Returns) Results: 4 Ordered Values Dependent Variable: AJ Method: ML .532523 -771417. log likelihood 93 .Ordered Logit (Quadratic hill climbing) Sample (adjusted): 2 1006737 Included observations: 1006736 after adjustments Number of ordered indicator values: 4 Convergence achieved after 5 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) 10000*DLOG(ASK) Coefficient 0.6204 696.2264 659.992082 6.766256 Std.222491 1. log likelihood Avg.532594 1.007366 0.2 -1068342 -0.5 0 0.504201 0.012206 406.7327 0 0 0 1.

log likelihood 94 .Ordered Logit (Quadratic hill climbing) Sample (adjusted): 2 1006737 Included observations: 1006736 after adjustments Number of ordered indicator values: 10 Convergence achieved after 4 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) Coefficient 0.07788 129.4. 0 0 Akaike info criterion Log likelihood Restr.630813 0.5268 297.863371 1.059952 4.007967 0.1. log likelihood Avg.059823 -2043574 -2076824 -2.4985 0 0 0 0 0 0 0 0 0 4.000728 z-Statistic 110.004866 0.812 612.187268 0.0571 Prob.005538 0.005024 0.9948 205.297395 0.804875 3.4991 543.004859 0.149279 Limit Points LIMIT_2:C(3) LIMIT_3:C(4) LIMIT_4:C(5) LIMIT_5:C(6) LIMIT_6:C(7) LIMIT_7:C(8) LIMIT_8:C(9) LIMIT_9:C(10) LIMIT_10:C(11) Pseudo R-squared Schwarz criterion Hannan-Quinn criter.007016 -350.204722 0.029901 Std.A.004839 0.059859 66500.112486 4.48186 70.5 0 0.8 Ordered Logit Results: 10 Ordered Values Dependent Variable: AJ Method: ML .8518 -38. LR statistic Prob(LR statistic) -2.004891 0.01601 4.795295 -0. Error 0.8187 176.005723 0.496184 2.001844 0.339075 0.8132 506.

575528 1.158021 3. LR statistic Prob(LR statistic) -1.008209 0.85286 6.3935 382.959922 0.1 0 0.4871 Prob.288018 2.005979 0. log likelihood 95 .001961 z-Statistic 124.736943 Std. Error 0.010167 -177.6506 443.006343 0.006163 0.734323 3. 0 0 0 Akaike info criterion Log likelihood Restr.002158 0.499334 1.268841 0.9462 765.3314 654.5703 267.4.034615 5.473909 -1748643 -2076824 -1.008444 0.473948 656363.8769 594.285226 7.100859 4.2899 741.283457 Limit Points LIMIT_2:C(4) LIMIT_3:C(5) LIMIT_4:C(6) LIMIT_5:C(7) LIMIT_6:C(8) LIMIT_7:C(9) LIMIT_8:C(10) LIMIT_9:C(11) LIMIT_10:C(12) Pseudo R-squared Schwarz criterion Hannan-Quinn criter. log likelihood Avg.9249 0 0 0 0 0 0 0 0 0 3.A.006789 0.Ordered Logit (Quadratic hill climbing) Sample (adjusted): 2 1006737 Included observations: 1006736 after adjustments Number of ordered indicator values: 10 Convergence achieved after 5 iterations Covariance matrix computed using second derivatives Variable GARCHVAR AJ(-1) 10000*DLOG(ASK) Coefficient 0.005795 0.007889 0.6451 488.549476 2.1.9 Ordered Logit (with Returns) Results: 10 Ordered Values Dependent Variable: AJ Method: ML .5891 574.001002 0.6376 782.47405 3.