Lunar Calendar Effect: Evidence of the Chinese Farmer’s

Calendar on the Equity Markets in East Asia

Wei-han Liu

School of Economics and Finance

Faculty of Law and Management

La Trobe University

Bundoora, VIC 3086, Australia

+ 61 3 94793129 (O)

+61 3 94791654 (F)

Email: weihanliu2002@yahoo.com

1
Title: Lunar Calendar Effect: Evidence of the Chinese Farmer’s Calendar on the Equity Markets

in East Asia

Abstract: This paper examines the statistical relation between the Chinese Farmer’s Calendar

(CFC) and public market information in ten East Asian equity markets during 1995-2004.

CATREG and CART, two data mining techniques, are employed and the implications of the

outcomes are discussed. The outcomes confirm that the CFC plays a supplementary role to

market information in predicting market rate of return. In addition to confirmation of lunar

calendar effect by CFC, CATREG outperforms in the three markets: Taiwan, South Korea, and

Singapore. According to CART analyses, all the three markets value Funeral category of CFC

advice and this pattern coincide the traditional wisdom of astrological knowledge. The latter two

markets share the same set of CFC items while Taiwan has its own. This contrast indicates the

market-dependent observance and different interpretation of CFC items for determining an

auspicious date for equity trading.

Keywords: Lunar Calendar Effect, Chinese Farmer’s Calendar, CATREG, Classification Tree

Analysis, CART

2
I. Introduction

Life is full of uncertainty and compromises, so is equity trading behavior. Investors seek

help from modern finance theory, e.g. capital asset pricing model, to avoid risk and pursue

higher return. However, there are significant discrepancies between market dynamics and

predictions of the models. One major reason is that market behavior does not conform to the

rational behavior specified in modern finance theory. Behavioral finance has emerged as a

superseding alternative. After all, the concept of bounded rationality prevails in explaining

individual behavior (Cho 2005). Hirshleifer (2001) summarizes that human judgment and

decision biases are derived from the following four common roots: fast and frugal heuristics,

overconfidence, mood and time-inconsistent preference, and social conformity. All these four

roosts are manifested in our daily life and the way of manifestation of human bounded rationality

has thus evolved as a field-specific cultural issue. Market anomalies accordingly exhibit similar

properties. Further, if equity market behavior is regarded as an expression of social psychology,

culture or superstition will influence outcome and exercise certain weight during the process.

The number of documented anomalies in equity markets is large and continues to grow.

While various types of market anomalies exist, calendar effects serve as one of the major

examples. Calendar effects are confirmed at various time series patterns across markets, e.g.

weekday, weekend, month, holiday, and festival, etc. (Campbell, Lo, and MacKinlay 1997;

Schwert 2003), in contrast to cross sectional return patterns which discuss the effects of size,

value, momentum and other things,. The current related literatures focus exclusively on solar

calendar though. Among the literatures on calendar effects, some studies have analyzed different

3
superstitions across different cultures. For example, American market participants do not like the

number 13, even Friday the 13 th (Kolb and Rodriguez 1988; Chamberlain, Cheung, and Kwan

1991). Chinese avoid the number 4 (Brown, Chua, and Mitchell 2002) and this dislike is also

revealed in many aspects, e.g. plate number selection (Woo et al. 2008). However, no study so

far has thoroughly analyzed the effects of days of lunar calendar on the large markets with

participants who observe lunar calendar. Yuan, Zheng, and Zhu (2006) and Dichev and Janes

(2002), have reported significant effects of the moon phases on stock returns across global

markets. These papers confirm that the lunar effect is independent of other ordinary solar

calendar-related anomalies. However, the lunar calendar effect means more than tracking the

moon phase because lunar calendar itself contains profound yet unexplored information.

There are four major types of lunar calendar: Chinese, Islamic, Hebrew, and Hindu.

Among them, the Chinese type is the most influential one partly because it has the largest

number of observers. In addition to its popular observance, this calendar has been observed for

over tens of centuries, acknowledged and practiced as an invigorating part of Chinese culture.

The Chinese type is integrated with Chinese tradition and this lunar calendar remains culturally

essential today in Chinese communities. The Chinese type is still used for making traditional

East Asian holidays and in astrology, e.g. for choosing the most auspicious date for a wedding or

the opening of a building. For traditional households, the Chinese type is used to pick auspicious

dates for important events, such as weddings, funerals, and business deals. The cultural

influences brought by the Chinese lunar calendar are pervasive mostly in some East Asian

countries.

4
 The Chinese type has its special properties; most importantly, on a daily basis, it gives

specific advices of favorable and unfavorable issues for our daily life. These daily advices are

usually regarded as sign of luck for specific daily activities and accordingly may be treated as

superstitions. Its detailed coverage extends to seven essential fields of human activities, e.g.

construction, marriage, worship, funeral, commerce, life, and farming. In imperial days, the

Chinese type is certified, decreed, and published exclusively by the central government, initially

for farming purpose. This calendar is nicknamed as the Chinese Farmer’s Calendar (CFC, 中國

農民曆). This lunar calendar is also called the Yellow Calendar (黃曆) or Imperial Calendar (皇

曆) because yellow is esteemed as the royal color, representing the emperor. This lunar calendar

is also called the Old Calendar (舊曆), while the solar system or the Gregorian calendar adopted

since the early 20 th century in China is called the New Calendar (新曆). The general populace

continues to use the traditional CFC, even though the Gregorian calendar was adopted for

official business (Chinese calendar, http://en.wikipedia.org/wiki/Chinese_calendar). Distinct

from other lunar calendars, CFC functions more than an ordinary calendar whose exclusive

purpose is to help keep track of time.

The Chinese lunar calendar can be regarded as a handy encyclopedia for making decision

in everyday life for traditional households. In general, Chinese customs value the fate and timing

of selective date for its specific purpose because superstitious belief holds that heaven, earth, and

human being are interconnected (天地人三才). The CFC directly gives explicit advices for the

daily life, and those daily advices cover lifetime major occasions as well as daily activities.

Based on these daily advices, observance of CFC facilitates the recognition of

favorable/unfavorable trend of observers for the specific date. This daily recognition can serve as
5
a hunch or a guideline for CFA observers to decide if it is a lucky or unlucky date to take or

avoid certain activity. In addition to being part of the culture, following the CFC is undeniably

superstition. This daily practice has certain degrees of influence on those who observe it,

especially on their mood and heuristics. It can be expected that the daily advices from CFC help

observers form their fast and frugal heuristics and expectedly CFC can exercise certain influence

on the behavior of equity trading of its observers. Interesting enough, the logics to decide the

daily advice is based the calculation of the power and number of certain luck stars to appear on a

specific date. Yet this mystical or esoteric know-how is only shared only among prestigious

astrologists, not in open media. The printed material covering the CFC is only available in

traditional Chinese, to the best of my knowledge and search. The book source covers only the

outcomes at the introductory level of interpretation for practical purpose, nothing of the theory or

calculation details. Yet, the popular observance of the CFC is quite significant in certain areas in

East Asia such as Taiwan, South Korea, and Singapore.

Human attention is sporadic and we pay attention especially to those which are followed

by most others, implicitly or explicitly. It is confirmed that culture or superstition may influence

the forming of heuristics or mood (Gilovich, Griffin, and Kahneman 2002). Either recognized as

culture or superstition, CFC falls within the paradigm of behavioral finance. The lunar calendar

effect is worthy of detailed investigation at least for the following reasons if we follow the four

common roots of human judgment biases outlined by Hirshleifer (2001).

First, the concept of mental accounting (Thaler 1985) claims that human decisions are

usually based on an efficient heuristic method which compares possible outcomes by state of the
6
world, rather than evaluating the expected utility of each alternative. The detailed daily advice

of the CFC helps the observers compare the favorable and unfavorable signs and anchor their

mood to judge if it is a favorable date to trade. The investors’ mood is changeable and vulnerable

to many factors which may not directly linked to equity market but are confirmed at work, such

as weather (Saunders 1993; Loughran and Schultz 2004), temperature (Cao and Wei 2005), etc.

The anchoring effect from the CFC affects its observers to determine their mindset and to make

them subconsciously fell confident or overconfident about the daily fate. Consequently this

mechanism assists mental accounting. In addition, people tend to judge the probability of an

event by finding a 'comparable known' event and assuming that the probabilities will be similar,

rather than essential characteristics. The detailed advice of the CFC can serve as a comparable

known event on a daily basis and facilitate people to form their heuristics accordingly by

assuming the probabilities will be similar. This leads to overweight some intuitive memory or

similar pattern but ignore the actual probability of a specific event. Kahneman and Tversky

(1973) name the cognitive shortcut as representativeness heuristic but it leads to lead to a bias

that “emotionally relevant events ought to have emotionally relevant causes.”

Second, in fact, superstition has its actual values (Ng, Chonga, and Du 2010). Rightly or

wrongly, superstitions can change economic activity (Ng, Chonga, and Du 2010) and their robust

influence is invoked to bring about good luck and to fend off bad luck. A myriad of marketing

research has confirmed that superstitious beliefs can influence consumer expectations and

consequently consumer behavior (Block and Kramer 2009; Kramer and Block 2008). Equity

trading is one of the major forms of consumer behavior and can be strongly influenced by

superstition and culture.

7
 Third, in addition to heuristics or mood, the daily advice from the CFC helps make its

observer confident or overconfident about their trading decisions. The daily advice of the CFC

can serve as a supplement to public market information, e.g. opening price, closing price,

turnover rate, such that the observers are inclined to believe they are more certain about their

daily fate before they make their trading decision. Consequently overconfidence and implied

overoptimism is promoted. The degree of social conformity to the CFC varies across markets

and in some cases its effect can be significant enough to push the non-observers to comply,

especially when the market pattern coincides with what CFC has specifies. This is most likely to

be the case for equity markets which are emerging and whose participants are predominantly

individual investors and those participants are mostly novices, less dependent on modern finance

theory, professional advice, or investment information. Roughly speaking, in the period before

2004 the East Asian markets are in general in their introductory and inactive stage; after 2004,

those markets are more influenced by international hot cash flows because of series measures of

financial liberalization. The equity markets in East Asia during the 1995-2004 periods are the

most appropriate candidates for this study.

Fourth, if we regard market trading behavior as an expression of social psychology, it is

even more interesting to investigate the effect of the CFC in the process. When the momentum of

market herd behavior reaches a certain level, the non-followers cannot afford to neglect the

plausibly significant “band wagon effect.” The consensus on market status expectation is thus

implicitly formed. The observers may exert their willpower to control the uncontrollable market

outcome. The actual market status can be accordingly affected. People in East Asia possess a

8
high degree of collectivism (Oyserman, Coon, and Kemmelmeier 2002), which may imply that

the if the influence of the CFC is significant enough then this herd behavior may play a

potentially important role in determining group-behavior among investors and the status of

equity market outcomes in these countries. As the CFC is mostly observed in East Asia, it is

interesting to study the respective effect of the CFC across the equity markets in that region. It is

expected that the degree of influence may vary across the market due to different intensity of

observance. After all, the tradition of observing the CFC can be diluted in some ways due to

local non-observers, colonist culture, and political reform. For example, Philippine and Malaysia

are dominated by their local cultures. Hong Kong and Singapore were once British colonies but

are different in their market structure in terms of composition of market participants and market

regulation. Japan has tried to sever the ties with Chinese culture since the Meiji era (1868-1912).

Japan has officially adopted the Gregorian calendar and abandoned the CFC since then. China

campaigned to abolish the old Chinese culture during the Cultural Revolution (1966-1975) and

the CFC has not been circulated until recent years. The rich information content of the CFC’s

daily advices should be studied so as to select the principal factors at work. Also, we can explore

the respective recognition patterns of the local investor with respect to the CFC’s daily advices in

that region and discuss their implications.

It can be thus expected that the CFC can create a significant lunar calendar effect, which

is distinct from solar calendar anomalies as pointed out by Yuan, Zheng, and Zhu (2006) and

Dichev and Janes (2002). However, the previous literature has not discussed the specific contents

of the lunar calendar effect on market anomalies. It is extremely interesting to check the

statistical relationship between market status and CFC advice. That is, can CFC advice affect its

9
observers in aggregate to have auspicious or inauspicious expectations on a certain date for stock

trading? What are the implications from those relevant lunar calendar conditions?

Before the quantitative analysis, I choose to avoid arbitrary assumption on the CFC

advices. There are possible complex relationships among the CFC advices such as correlation

and endogeneity. These properties exclude the traditional regression approaches as the

appropriate quantitative methods to employ. I select two innovative techniques in data mining for

this study: CATREG and CART (Classification and Regression Trees). CATREG first quantifies

categorical variables using optimal scaling and then employs alternating least squares to attain an

optimal linear regression equation for the transformed variables. The variables involved can be

given mixed optimal scaling levels and no distributional assumptions about the variables are

made. Through optimal scaling and alternating least squares, CATREG helps transform the

regression of categorical variables into an ordinary linear regression. Optimal scaling gives

flexible transformation of the nonlinear relations. Alternating least squares help the

transformations of optimal scaling chosen so that the (univariate, multiple, or canonical)

correlation between the transformed predictor(s) and the transformed dependent variable(s) is as

large as possible (William D. Perreault and Young 1980). Thanks to CATREG, we can run

regression with mixed types of data and this study (nominal, ordinal, or numeric). We can thus

compare the explanatory performance of possible regressions, i.e. various combinations of

explanatory variables and dependent variables. This will help us to find out the best-performing

regression model for lunar calendar effect. On the other hand, CART is one of the major

techniques in classification tree analysis which helps searching for and isolating significant

patterns and relationships which may remain hidden using other analytical tools. CART can

10
quickly reveal important data relationships that could remain hidden using other analytical tools.

We can expect CART will help in sorting of the principal set of explanatory variables among the

CFC advices and market information.

My empirical analysis of the East Asian equity markets during 1995-2004 confirms there

are the lunar calendar effects of the CFC. CATREG analyses confirm that CFC advices have

certain levels of explanatory power for market performance, especially in three markets: Taiwan,

South Korea, and Singapore. The explanatory performance is significant for predicting market

rate of return especially when CFC data, market information, and market status serve together as

predictors in a CATREG regression. In addition to significant price-volume relation,

classification tree analyses help select the significant CFC items. The outcomes indicate that

South Korea and Singapore follow the traditional wisdom and share the same CFC item set:

categories of Construction, Worship, and Funeral. Taiwan has its own calibrated and diverse set.

This contrast of sets of selected CFC items shows the different level of observance in respective

local markets and disparity in CFC recognition in terms of aggregate equity trading behavior.

The remaining sections of the paper are organized as follows. Section II introduces

CATREG method, and Section III overviews the methods for classification tree analysis. Section

IV summarizes the empirical analysis results and discussion. Section V concludes.

Section II: CATREG

11
 Among regression approaches for nonlinearly related data, Regression with

Transformations approach is known for its linearizing the relation between the response and the

predictors, allowing for flexible transformation of nonlinear relations. CATREG, one of the

Regressions with Transformations approaches, applies optimal scaling methodology to transform

the response and the predictor variables. The variables in CATREG can be nominal, ordinal, or

numeric. Nominal and ordinal variables have some shortfalls that keep them often difficult to

implement and interpret definitively. Many statistical procedures cannot be undertaken

accordingly. The drawbacks of nominal and ordinal variables could be circumvented by

converting nominal and ordinal variables to interval (numeric) variables. To some degree,

optimal scaling can fulfill this purpose. That is, optimal scaling derives interval measures from

nominal and ordinal measures. Optimal scaling is composed two processes: quantification and

transformation. The former is designed to find optimal numeric values to transform categorical

data to numeric data. The latter is on scaling which is nonlinear transformation that treats

numeric variables as categorical variables, with the number of categories equal to the number of

distinct values of the variable. Thus, optimal scaling is applicable to both categorical variables

and to numerical variables and this transformation linearlize the relation between the response

and predictor variables. For estimation, the transformed categorical variables are estimated

simultaneously with the estimation of the regression coefficients, using an alternating least

squares procedure that maximizes the coefficient of determination for linear regression on

transformed variables (Kooij 2007; van der Kooij, Meulman, and Heiser 2006). CATREG has

the advantages of both the Regression with Transformations approach and the optimal scaling

methodology, and serves the best candidate in this study because of the mixed data types of the

involved dataset. That is, nominal data include the status variables of CFC advices, ordinal data

12
include the variable of market status, and numeric data include market rate of return and major

market information.

CATREG model is a classical linear regression model of the numerical and transformed

variables and can be specified as

J
r  y     j j  x j   e
i 1 (2.1)

with the loss function

J 2

L r ; 1 ,..., r ; 1 ,...,  j   r  y     j j  x j 

i 1 (2.2)

xj
where J denotes the number of variables, y is the observed or discretized dependent variable,

j
denotes the observed or discretized predictor variables, is the regression coefficients,  r is

j
the transformations of the dependent variables, denotes the transformations of the predictor

variables, and e is the error vector

The form of the transformations depends on the optimal scaling level of each variable.

Scaling level refers to the level at which a particular variable is analyzed. Optimal scaling level

depends on the measurement level of the variable, which describes the scale properties of the

13
observed variable. Variable with nominal measurement level only have the grouping property,

that is, the category values only serve to code the observations into classes. Ordinal variables

have the properties of grouping and ordering. Interval (numeric) variables have all the three

properties: grouping, ordering and equal relative spacing. The scaling level is related to the

number of degrees of freedom, the number of categories minus one, of the transformation and

then to the fit of the model. Nominal and ordinal scaling level result in transformations that are

step functions, which are suitable for variables with limited number of categories. For variables

with large number of categories, spline functions are more appropriate.

In terms of estimation, CATREG uses alternating least squares to estimate the regression

model and the quantifications in a simultaneous and iterative manner such that to minimize the

loss function (2.2). Because the transformed variables are centered, loss function (2.2)

maximizes the (squared) multiple correlation. Alternating least squares involves a process of

alternating back and forth between two different least squares phases: (1) a model estimation

phase and (2) an optimal scaling phase. That is, CATREG alternates between estimation of the

transformation of the response variable and estimation of the transformations and regression

weight. The iteration refers to the transformation of the response is estimated from the linear

combination of the transformed predictors from the previous iteration. The whole estimation

process uses the backfitting procedure (Buja, Hasti, and Tibshirani 1989) in which the

transformations are estimated from the partial residual when the response is predicted from all

predictor, except the predictor for which the transformation is being estimated. Specifically, the

CATREG algorithm consists of two steps. In the first step, keeping the quantifications of the

predictor variables and the regression coefficients fixed, the quantification of the response

variable is estimated. In the second step, holding fixed the quantification of the response variable,
14
the quantifications of the predictor variables and the regression coefficients are estimated for one

variable at a time.

For the parameter estimation in regression, CATREG employs regularization methods

because ordinary least squares regression does not perform well with respect to prediction

accuracy, due to highly instable parameter estimates. Regularization methods include Ridge

regression, Lasso (Least Absolute Shrinking and Selection Operator) (Tibshirani 1996), and

Elastic net (Zou and Hastie 2005). Ridge regression reduces coefficient estimates variability by

shrinking the coefficients, resulting in more prediction accuracy at the cost of only a small

increase of bias. Lasso improves prediction accuracy and model interpretability by combining

the nice features of Ridge regression and subset selection. Elastic net further improves Lasso by

encouraging grouping of highly correlated variables while Lasso would select only one variable

in the group. Zou and Hastie (2005) show that whenever Ridge regression improves on ordinary

least squares regression, the Elastic net will improves on Lasso. Thus, Elastic net is selected as

the estimation method in the CATREG algorithm for regularized nonlinear regression in this

study.

The performance of a fitted regression model can be determined by its future

observations. The optimality of the quantifications can be obtained from the observed data to

predict future response and the usual criterion is the expected prediction error. For a linear

regression, the apparent prediction error is the average loss of the observed data, which is

minimized over the regression weights. The apparent prediction error is not an appropriate

estimate for expected prediction error because the expected prediction error is estimated from the
15
same data that were used for fitting the model and consequently leads to an optimistic estimate.

To obtain a better estimate, resampling methods such as cross validation or a bootstrap procedure

could be used. The goal of CATREG is to minimize the apparent prediction error for the dataset

at hand and resampling methods like cross validation or bootstrap are used for this purpose.

These resampling methods can serve as out of sample tests to avoid the possible over-fitting due

to optimal scaling. However, those two resampling methods are handicapped with issues of bias

and variability. The .632 bootstrap (Efron 1983), a combination of both bootstrapping and cross

validation, is shown to outperform the aforementioned two resampling methods.

Section III: Classification Tree Analysis

Classification tree analysis is one of the major data mining techniques. It is designed to

predict membership of cases or objects in the classes of a categorical dependent variable from

their measurement on one or more predictor variables, so as to reveal the patterns and

relationship among variables. This nonparametric and efficient technique has become an

attractive analysis option for its more flexibility in data requirement and less stringent theoretical

and distributional assumptions than the traditional approaches such as linear discriminant

analysis, cluster analysis, logistic regression, nonparametric additive logistic regression, partial

least squares classification, or neural networks. Its hierarchical nature can help simplify

interpretation of the effects of the predictor variables one at a time even for a complex tree, while

it is difficult to incorporate possible interactions in the traditional methods. Those traditional

methods make parametric assumptions which may not be valid. It is designed to analyze large

datasets with high dimensionality, mixed data type, missing data, different relationships between

16
variables in different parts of measurement space, and outliers. Its flexibility in processing

categorical, numerical, or mixture variables helps outperform other traditional approaches

(Hastie, Tibshirani, and Friedman 2009; Izenman 2008; Meulman and Heiser 2007).

The major classification tree analysis techniques include: CART, Chi-Squared Automatic

Interaction Detection (CHAID), and Quick, Unbiased, Efficient Statistical Tree (QUEST).

Feldman and Gross (2005) document that the other aforementioned traditional methods are not

objective enough and recommend CART (Breiman et al. 1998) even though this method is not

perfect, e.g. recursive nature of its algorithm makes it time-consuming. CART is noted for its

recursive consideration of one feature at a time, instead of working on multiple features at a time

as most other parametric and nonparametric methods do. CART is selected in this study because

this technique splits the data into segments that are as homogeneous as possible with respect to

the dependent variable. CART can thus help find the most important advices from the CFC for

predicting market status or market index return, while keeping the model as parsimonious as

possible.

A classification tree contains nodes and branches. The nodes include interior nodes (the

categories formed during classification) and terminal nodes (the last category formed during

classification). The dataset is classified into major categories based on the data diversity. The

sample is divided to maximize the differences (variances) between its subgroups on the

dependent variable. Then, the data in each category is further divided into branches. The splitting

is repeated within formed subgroups recursively until the preset limit is reached. The best

candidate tree structure is identified by recursively expanding from a root node, and specific
17
branches are gradually pruned. Consequently, the data diversity is minimized within categories

and data with the same attributes will be grouped in the same node. Although the splits are

selected by maximizing the local reduction in diversity, the procedure also minimizes the overall

tree diversity. The relationship among the variables is then mapped as a tree. The recursive

computational process can be categorized into four steps: (1) specifying the criteria for predictive

accuracy; (2) selecting splits; (3) determining when to stop splitting; and (4) selecting the “right-

sized” tree. The CART algorithm is a recursive procedure; starting at the root node, and then at

every internal node, it selects a single feature, and a threshold value to split the group of

individuals at the node into two groups to be placed at two new daughter nodes. CART grows the

largest tree possible, called a maximal tree, that is the tree whose terminal nodes cannot be split

any further. A node may not be split any further either because it contains only cases that belong

to a single class, or because no reduction in total diversity can be obtained by further splitting.

A splitting rule is derived from a diversity function which measures the diversity of the

node that is to be split. Suppose that the node is split along a certain attribute into K classes. Let

C i | k 
the cost, , of misclassifying a case that belongs to class k into class i, when in fact it

C i | k   0 C i | i   0 p k | t
belongs to class k, satisfying and . Let be the proportion of class

0  p  k | t   1 k  1,..., K
j cases present at node t of the tree, obeying , . For each node t,

 p  j | t   1 . CART provides three binary recursive splitting criteria: Gini, Entropy, and
j 1

Twoing. It still remains an open question of which criterion is most appropriate for a given

dataset. In the end, all these three criteria give exceptionally similar final trees. It is

recommended to use Gini criterion when the number of feature categories is kept to a minimum.
18
I select the splitting criterion Gini, relatively robust to the effects of outliers, because this

requirement is met in this study. The Gini index of diversity under uniform cost is

K k 1 K
1 
d G  t    p  k | t  p  i | t   1   p 2  k | t   , and
k 1 i 1 2  k 1 

K k 1
d G  t    p  k | t  p  i | t  C  k | i   C  i | k  under non-uniform costs.
k 1 i 1

In terms of selecting and pruning a tree, it is devised to select a tree that minimizes the

diversity and the best split for that feature for each node. The diversity is defined as

D T    d  s, t  , where tree T has been generated with terminal nodes T t . A maximal tree is
tT t

initially produced by splitting nodes until each node contains only cases that belong to a single

class or contains nodes whose diversity cannot be reduced by further splitting. The resulting tree

is the tree which minimizes the diversity D(T) but it is not necessarily the best tree from the point

of view of misclassification. The goodness of the tree is evaluated by its estimated

misclassification rate and the expected misclassification cost is defined as

K K
R T     C i | k  Q i | k    k  , where Q  i | k  denotes the proportion of class k cases
k 1 i 1, i  k

misclassified into class i, and   k  is the prior probability of a case being in class k. Due to

possibly underestimated misclassification rates during the splitting process, two alternatives are

developed: the cross-validation method and the test-sample method. When the dataset is

sufficiently large, test-sample method is preferred and the performance is satisfactory enough.

Standard errors of the overall misclassification cost are calculated. The search for the best pruned

tree is terminated when a subtree is found that is within one standard error of the minimum

estimated misclassification cost subtree. Thus, we can use CART technique to help select and
19
prune the minimum set of significant CFC advices in this study. Accordingly, the tree structure

outcome can help us get a concise overview of all the involved variables of mixed data types.

Section IV: Data Analysis

The purpose of this study is to examine whether the daily advices in the CFC has a

significant statistical relationship with equity market performance, filter out the significant pieces

of advice, and derive their implications. The CFC covers seven major aspects of daily life and

lifetime occasions: Construction, Farming, Worship, Funeral, Commerce, Life, and Marriage. Its

seventy-nine items and their definitions, respective code, and original Chinese are summarized in

Table 1. For East Asian countries, the CFC is the most popularly observed lunar calendar system

and it serves as the primary public source for a calendar and daily activity advice source for tens

of centuries. The degree of local observance may vary due to the local cultural situations.

Although the CFC was originally used to develop recommended actions for farming purposes,

the range of daily advice has been extended over time to include almost every aspect of daily life.

The study of the influence of CFC presented in this paper is based on the following ten

East Asian equity market indices: Taiwan (Taipei Weighted Price Index, code: TW), Hong Kong

(Hang Seng Index, code: HK), Shanghai (Shanghia A-shares, code: SHA), Shenzhen (Shenzhen

A-shares, code: SZ), Singapore (Singapore Straits Times Industrial Index, code: SG), Philippine

(Manila Composite Index, code: PH), South Korea (Seoul Composite Index, code: SK), Japan

(Nikkei Average, code: JP), Indonesia (Jakarta Stock Exchange Composite Index, code: INDO),

20
and Malaysia (Kuala Lumpur Composite Index, code: MA)1. The sample period for the daily

data ranges from 1995 to 2004. During the data period, most of the East Asian equity markets

are in their preliminary stage and dominated by novice investors. The selected equity markets

exhibit different degrees of market openness and market efficiency, and Chinese culture has

varying degrees of influence in these countries. The time series plots of the aggregate index of

the respective markets are summarized in Figure 1. Among the ten panels, three patterns have

emerged for their respective aggregate market indexes: increasing (HK, SHA, SZ, INDO),

decreasing (JP, ML, PH), and balanced (SG, SK, TW). The market outcomes for the Shanghai

and Shenzhen markets are limited to the A-shares, which are only available to local citizens, not

foreign investors. This restriction helps evaluate the impact of local market psychology

exclusive of foreign investors.

The set of seventy-nine major CFC activity items that provide daily advice are listed in

Table 1. The seventy-nine activity items are classified into seven categories (A through G):

Construction, Marriage, Worship, Funeral, Commerce, Life, and Farming. For every day, the

CFC states if the date is viewed as “auspicious” (宜) or “inauspicious” (忌) for any of the

activity item in Table 1. I assign the nominal weight of “+1” to that day for activities for which

1
The detailed information for these ten capitalization-weighted indexes are available in the following respective
websites: Taiwan Stock Exchange Corporation, http://www.twse.com.tw/en/; Hang Seng Indexes Company,
http://www.hsi.com.hk/HSI-Net/; Shanghai Stock Exchange, http://www.sse.com.cn/sseportal/en/home/home.shtml;
Shenzhen Stock Exchange, http://www.szse.cn/main/en/; Singapore Exchange,
http://www.sgx.com/wps/portal/marketplace/mp-en/home; Philippine Stock Exchange, http://www.pse.com.ph/;
Korea Exchange, http://eng.krx.co.kr/; Tokyo Stock Exchange, http://www.tse.or.jp/english/; Indonesia Stock
Exchange, http://www.idx.co.id/; Bursa Malaysia, http://www.klse.com.my/website/bm/.

21
the day is labeled “auspicious,” “-1” for activities for which the day is labeled “inauspicious,”

and the day is assigned “0” for an item if the day is neutral or unmentioned for that item.

Figures 2 and 3 summarize the frequencies of the CFC items and CFC categories,

respectively. CFC advices show neutral signs most of the time. In general, label “inauspicious” is

more frequent than label “auspicious,” except in category D (Funeral, twelve items: digging

grave, collecting skeleton, renovating tomb, erecting tombstone, laying in coffin, moving coffin,

burial, digging grave during life, preparing coffin, day and hour of birth, sweeping clean,

mourning dress). For funeral activity, the observers are encouraged to take specific action at the

lucky date to bring peace to the deceased and blessings to the children. In category C (Worship,

eleven items: worship, praying for good, giving food, installing incense tray, praying for kids,

consecration, sculpture and painting, gratitude, exorcism, visiting grave, erecting altar), the

frequencies of “inauspicious” and “auspicious” are quite similar. Worship activity can be

positioned as a daily necessary practice because the highest frequency of “auspicious” is the C1

item: worship. In general, the CFC provides more precautious advices than encouraging advices.

However, in the both categories of Funeral and Worship, the CFC advices are inclined to show

positive signs when the date is right. That is, the observers are advised to take specific activities

of both categories in the specific data when the date is lucky enough. This contrast underlines the

significance of both categories of Funeral and Worship. In addition to the categorical data

provided in CFC, two types of numerical market data were collected from the DataStream

database2: market rate of return and market information. The market rate of return is a numeric

data which is the observed daily rate of return for consecutive trading days based on aggregate

2
http://product.datastream.com/extranet
22
market index. The daily market information variables are numerical and include the highest

price (HIGH), the low price (LOW), opening price (OPEN), closing price (CLOSE), turnover

volume (TURNVO)3. The market status variable (WINORLOSE) is ordinal and is assigned the

value “1” if the daily rate of return is positive or “0” if daily rate of return is zero or negative.

Figure 4 summarizes the market status of each market. Distinct from other markets, SZ and SHA

exhibit feisty market trend with more upticks than downticks.

This study investigates if the market participants follow the CFC advices to predict the

equity market performance in terms of market status and market rate of return. Market status and

market rate of return are selected as the dependent variables for they are most investors want to

know about the aggregate market conditions in the trading days. Based on aforementioned

variables, I attempt to examine the statistical relationship between dependent variables (the

ordinal variable, market status; the numeric variable, market rate of return), and the explanatory

variables (the nominal variables, CFC items; the numeric variables, market information

variables). Considering all possible combinations, I set up the following ten regressions to

examine if CFC advice and/or market information can exercise significant explanatory power for

market status and/or market rate of return.

1. Dependent: market rate of return; Predictors: CFC items, market information

2. Dependent: market rate of return; Predictors: CFC items, market information, market

status

3. Dependent: market status; Predictors: CFC items, market information

3
Some of the ten equity markets are in their preliminary stage during the data period. Not all the markets carry the
same market information set. I choose the representative five items which are available for all the ten markets.
23
 4. Dependent: market status; Predictors: CFC items, market information, market rate of

return

5. Dependent: market rate of return; Predictors: market information

6. Dependent: market rate of return; Predictors: CFC items

7. Dependent: market status; Predictors: market information

8. Dependent: market status; Predictors: CFC items

9. Dependent: market status; Predictors: market information, market rate of return

10. Dependent: market status; Predictors: market information

Those ten regressions consider all possible combinations to investigate if CFC items can

have significant explanatory performance with regard to market status and market rate of return.

Regressions 1, 2, 5, and 6 have the rate of return of market index as the dependant variables

while the other four regressions have market status as the dependant variable. The candidates of

predictor variables include: CFC items (nominal data), market information (numeric data),

market rate of return (numeric data), and market status (ordinal data). CATREG is employed for

these regressions of mixed types of data. Traditional regression methods are not appropriate for

this purpose because of mixed types of data, e.g. ordinary least squares regression, logit

regression, linear and nonlinear regression. Numeric data (market information and market rate of

return) are transformed by grouping which I follow the default and assume normal distribution of

seven categories. Market status variable is transformed by grouping method with one equal

interval because there are only two categories. Step functions are used for nominal and ordinal

scaling because there are limited number of categories in the CFC items and market status.

Elastic net is used for parameter estimation for this method generally outperforms Ridge

24
regression and Lasso. The .632 bootstrap is employed for optimal quantification, while avoiding

possible over-fitting and the number of sample for bootstrapping is set as 2000 for all markets.

Since there are different numbers of predictors in the aforementioned regressions, adjusted R

squares is used as the criterion for model performance comparison. The outcomes for each

market by each regression are summarized in Figure 5. Listed below are the observations from

Figure 5:

1. In general, either CFC items or market information functioning alone as the predictors do not

give impressive performance (Regressions 5, 6, 7, 8, 10) but the combinations may make

substantial contribution. The high adjusted R squares in Equation 9 can be due to the high

collinearity between the variables of market rate of return and market status.

2. Noteworthy, in Regression 8, CFC items alone can give adjusted R square level between

18~28% in the markets of TW, SK, and SG in terms of predicting market status. CFC items

show certain levels of contribution in explaining market status in specific markets. Meanwhile,

the adjusted R square levels in the other seven markets are close to zero. This contrast

indicates that CFC items alone can have significant explanatory power especially over the

three equity markets: TW, SK, and SG in terms of predicting market status.

3. Market rate of return is more predictable than market status due to higher adjusted R squares

level. Regression 2 outperforms especially in the three markets: TW, SK, and SG (in

descending order), with adjusted R squares level around 70%. However, the high adjusted R

squares may contribute to the significant correlation between the variables of market rate of

return and market status. Regression 3, which include CFC items and market information as

the predictors and market status as the dependent variable, gives the performance level over

50%.

25
4. The results for the three Chinese dominant markets (HK, SZ, and SHA) indicate that CFC

hardly shows explanatory power. The People’s Republic of China has a long traumatic history

of denouncing traditional Chinese culture, e.g. the Cultural Revolution, let alone CFC

observance. The CFC was categorized as one of the Four Olds4 to be destroyed during the

Cultural Revolution. In addition to the opening-up policy since Deng Xiaoping’s era, it takes

a long time to recover the Chinese cultural tradition and the data period falls in its infant stage

of rebirth. The CFC has not gained much circulation in China even today. Meanwhile, the

rapid economic growth and limited investment options have made the stock markets in SZ

and SHA overheated. The zealous local investors hardly value any sort of advice as mass

media has extensively reported the madness in the equity markets. Hong Kong is dominated

by institutional investors and the aggregate individual investors are prone to be market

followers, instead of trend makers.

5. The other four markets do not show significant statistical relationship with CFC: Indonesia,

Malaysia, Philippine, and Japan. The former three markets are more immersed with local non-

Chinese culture and colonist culture. In Japan, the government and the populace have

officially embraced Western culture and abandoned Chinese culture since the Meiji era by

adopting the Gregorian calendar and stopping using the CFC. Thus, it is understandable that

the CFC hardly shows any significant relationship with their respective market index.

Figure 5 suggests that CFC items have more explanatory power in the markets of TW,

SK, and SG in. I accordingly focus the further discussion on these three markets. Table 2

summarizes the CATREG analysis outcomes which contains two parts: coefficient estimate and

4 The Red Guards in the Cultural Revolution campaigned to “Destroy the Four Olds,” intriguing to break the old
thought, culture, custom, and habit of the traditional Chinese society.
26
correlation matrix for original predictors and for transformed predictors. Because the majority of

the coefficient estimates are close to zero, I summarize those coefficient estimates which are

significant different from zero. The coefficient estimates of the independent variables are rarely

significant except for some numeric variables of market information. However, there are

significant correlations among all the explanatory variables in these three respective markets.

The highly complex correlation structure among the variables is part of the esoteric astrological

science. It indicates that it is not appropriate to employ traditional regression technique in this

study which is beyond the scope of this study. It is hard to expect that, in general, the investors

have a good command of the CFC advices. It is advisable to work on the CFC category, rather

than the specific item, to grasp a more comprehensible picture. Thus, due to the highly complex

correlation structure, I need to resort an approach that can serve this purpose. I proceed to filter

out the significant CFC items in the three markets by employing CART, one of the classification

tree methods. 80% of the data series are used for training purpose for applying CART and the

remaining 20% are reserved for testing the model. Table 5 summarizes the risk estimate and its

standard error of the three respective markets. The statistics in the cells serve as a measure of the

tree’s predictive accuracy. For the market rate of return as a numeric dependent variable, the risk

estimate is within-node variance. The lower value of the estimate, the better model performance.

In other words, the CART model performs better in SG and SK markets5 than in TW market.

The difference of performance indicates that there is more diversity exhibited in the selected

variable set of TW.

5
Classification table to show the misclassification cost is not available for numeric dependent variable.
27
 According to Figures 6 through 86, the included independent variables, the predictors in

the nodes of the tree diagram, for the specific markets are specified below:

TW: WINORLOSE, HIGH, E4, F1, E9, TURNOVER, OPEN, B3, D2

SK: WINORLOSE, LOW, C8, TURNOVER, OPEN, D7, C3, A11

SG: WINORLOSE, LOW, C8, TURNOVER, OPEN, D7, C3, A11

Some points of discussion can be summarized as follows:

1. Significant price-volume relations are exhibited in the aforementioned three markets for some

market information variables are selected.

2. SG and SK share a common set of predictors while TW has its own. Specifically, the

categories of CFC items in SG and SK lie in Construction (eighteen items: demolishing,

building, break ground, launching, door or window, restroom, moving out, moving in, moving

incense tray, post, stove, grindstone, temple, well, pond, wood, embankment, bridge),

Worship, and Funeral. For TW, the CFC categories are Marriage, Funeral, Commerce, and

Life. In other words, SG and SK value more on items of ceremonial occasions for the

deceased and new establishment. TW considers more diverse aspects. Consequently, more

variance can be introduced in the CART analysis of the TW market. This also leads to better

explanatory power and higher adjusted R squares in TW as Figure 5 shows. This contrast

indicates the local investors in TW are more familiar with CFC items and know what to fine-

tune their market expectation based on the complex item set. Overall, the category of Funeral

is the common one in all the three market. This outcome coincides the traditional wisdom that

6
Figures 3 through 5 show the concise tree structure diagrams without the detailed statistics so as to avoid the reader
being overwhelmed, e.g. Chi-square statistics and the number of cases in each node. The full set of output can be
available upon request.
28
 funeral activities take the most luck. This major lifetime activity is important to the deceased

and the offspring.

3. The intricate principle for the classification of an “auspicious” day is so esoteric that is

apprehended to a small group of CFC experts. Its complexity highlights the highly nonlinear

and complicated relationships among the CFC advices. Its discussion is beyond the scope of

this study. According to the CFC professionals, this principle depends on the power, size, and

number of “lucky stars” assigned to the CFC activity item(s) on that day. In general, it takes

more and dominant lucky stars to be an auspicious day, so is the situation for the Construction

or Funeral categories. Accordingly, both the Construction or Funeral categories play a more

decisive role in differentiating an auspicious date or not. The dominance of these two

categories in the outcomes may suggest that people accordingly place more emphasis on those

activity items under the two categories in addition to market information. The local investors

of SG and SK significantly follow the traditional view of CFC items.

4. All the selected predictor sets of the three markets do not cover Category G. This reflects that

the local investors in the three markets hardly consider equity trading relevant to farming

activities.

5. This contrast also indicates that the interpretation of CFC items varies across markets. The

CFC can be recognized as either superstition or culture, and the comprehension of CFC is

market-dependent.

Section V: Conclusion

29
 There are extensive literature confirming calendar anomalies, one form of market

anomaly, but they focus on solar calendar exclusively. This paper attempts to be the first piece to

study lunar calendar effect. Among the major lunar calendars, the CFC in East Asia is known for

its daily advices covering daily activities and lifetime occasions. Recognized either as culture or

superstition, this calendar is popularly observed partly because it helps observers form their

recognition about the fate of the date. It is interesting to know to what degree the influence the

CFC can have to its observers in aggregate in respective equity markets. Among the seven-nine

items of the CFC advices, what are the significant pieces that people care and observe for their

equity trading activity? What are the implications out of the selected set of variables?

This study investigates the ten equity markets in East Asia during 1995-2004 when most

of the developing markets were largely participated by novice investors then. In addition to CFC

advices, I include public market information for empirical analysis. Traditional regression

techniques are inappropriate for this study because of mixed data types and highly complex

structure among the variables. Thus, I resort to data mining techniques: CATREG and CART.

CATREG applies optimal scaling to transform dependent and predictor variables so as to handle

all types of data, thus resulting in an optimal linear regression is feasible based on the

transformed variables. Alternative least squares regression technique helps the parameter

estimation. We can compare the model performance among the markets and determine in which

market CFC advices have higher explanatory power. CART is one of the major classification

tree analysis techniques. CART helps filter out the significant variables through iterative splitting

process of the variables into homogeneous groups. This technique helps avoid working on the

regression structure and present the significant factors in an intuitive tree structure.

30
 The empirical analyses indicate that the CFC advices have the higher explanatory

performance in the three markets: TW, SK, and SG (in descending order of adjusted R squares),

for their respective adjusted R squares are significant higher than the other seven markets. This

indicates that CFC advices together with market information and market rate of return have

significant explanatory performance in predicting market status, i.e. uptick or downtick of daily

index return, in these three markets. CART analysis outcomes indicate the significant price-

volume relationship in the three markets. Also, CART helps select different sets of CFC advices

for the respective market. SK and SG share the same CFC categories: Construction, Worship,

and Funeral. For TW, the CFC categories are Marriage, Funeral, Commerce, and Life. The

market participants in SK and SG follow the traditional way by observing the fundamental set of

the CFC categories for determining the daily fate. The investors in TW are more familiar in

articulating the CFC advices since the selected categories are more diverse. Overall, the category

of Funeral is the common one in all the three market. These outcomes coincide the traditional

wisdom that funeral activities take the most luck and this major lifetime activity matters the

family fate as well. Funeral category can serve as a significant indicator of daily fate

Culture or superstition is an issue at work in our daily life but it is not much studied in

behavioral finance. Either recognized as culture or superstition, the Chinese Farmer’s Calendar is

confirmed to have influence in some of the East Asian equity markets. The interpretation pattern

is market-dependent. I expect this study can help give new insights to decision bias in behavioral

finance from the perspectives of culture or superstition.

31
32
Table 1：Details on CFC Items

Category A: Construction

Code Item Chinese Definition

A1 demolishing 拆卸 tearing down door, wall, or building
A2 building 修造 building or renovating house
A3 Break ground 動土 launching a construction project
A4 launching 起基 launching foundation construction jobs
A5 door or window 安門 Fixing door or window
A6 restroom 作廁 building or renovating restroom
A7 moving out 移徙 moving out to other residence
A8 moving in 入宅 moving in to new residence
A9 moving incense tray 出火 moving incense tray
A10 post 豎柱上樑 placing posts or beams
A11 stove 作灶 renovating or moving stove
A12 grindstone 安碓磑 renovating or installing grindstone
A13 temple 造廟 building or renovating temple
A14 well 開渠穿井 building aqueduct or digging well
A15 pond 開池 digging pond or building dam
A16 wood 伐木 cutting tree
A17 embankment 築堤防 building bank
A18 bridge 造橋 building bridge or viaduct

Category B: Marriage
Code Item Chinese Definition
B1 hat and gown 冠笄 dress up for maturity ceremony
B2 name 問名 asking names and ages of each side
B3 engagement 訂盟 undergoing first-stage engagement
B4 completing proposal 納采 undergoing last-stage engagement
B5 suit 裁衣 tailoring dress suit for bride
B6 mosquito net 合帳 installing or making mosquito net for bed
B7 bed 安床 buying or moving bed for newly wedded couple
B8 wedding 嫁娶 the bridegroom marrying bride
B9 taking wife’s surname 納婿 the bridegroom marrying into the bride’s family

Category C: Worship
Code Item Chinese Definition
C1 worship 祭祀 worshiping ancestors or gods
C2 praying for good 祈福 praying for good or redeeming a pledge
C3 giving food 齋醮 giving food to monastery
C4 installing incense tray 安香 installing incense tray for ancestors or gods
C5 praying for kids 求嗣 praying for the birth of a son
C6 consecration 開光 consecrating newly completed idol
C7 sculpture and painting 塑繪 sculpturing and painting idol figure
C8 gratitude 謝土 gratitude toward completion of construction or

33
 funeral
C9 exorcism 解除 exorcising evil spirits
C10 visiting grave 祭墓 visiting grave
C11 erecting altar 設醮 setting up altar for prayer rites

Category D: Funeral
Code Item Chinese Definition
D1 digging grave 破土 digging grave and building tomb
D2 collecting skeleton 啟攢 collecting skeleton and storing in jar
D3 renovating tomb 修墳 renovating grave
D4 erecting tombstone 立碑 erecting tombstone or monument
D5 laying in coffin 入殮 laying body in a coffin
D6 moving coffin 移柩 moving coffin to main hall
D7 burial 安葬 burying in a grave
digging grave during tomb erected by person or for oneself, while
D8 開生墳
life living
coffin prepared by person or for oneself, while
D9 preparing coffin 合壽木
living
D10 day and hour of birth 進壽符 placing tally of day and hour of birth in a tomb
D11 sweeping clean 掃舍宇 sweeping clean of the whole residence
D12 mourning dress 成除服 putting on/off mourning dress

Category E: Commerce
Code Item Chinese Definition
E1 opening 開市 opening for business or work
buying asset or inventory, asking payment of
E2 money 納財
debt
E3 contract 立券 batch transaction of goods, signing a contract
transaction involving movable or immovable
E4 deal 交易
assets
E5 dyeing 作染 dyeing cotton or silk goods
E6 brewing 醞釀 brewing wine or vinegar
E7 vessel 造船 making a ship
E8 vehicle 造車器 making a land conveyance
E9 shop signboard 掛匾 placing horizontal tablet over door
E10 delivery 開倉出貨 production and delivery of goods

Category F: Life
Code Item Chinese Definition
F1 travel 出行 travel, going outdoors
F2 kotow 入學 bowing in salute to be a pupil to learn art
F3 bath 沐浴 fasting and bathing before a religious observance
F4 haircut 剃頭 cutting off hair for a baby or to be a monk
F5 meeting 會親友 meeting friends for business or party
F6 take-office 上官赴任 taking office or inaugurating
F7 doctor 求醫療病 seeing a doctor or having surgery
34
 F8 adoption 進入口 adopting children

Category G: Farming
Code Item Chinese Definition
G1 planting 栽種 sowing or planting
G2 extermination 捕捉 Exterminating insects
G3 honey 割蜜 harvesting honey
G4 fish 取魚 casting a net or fishing
G5 net 結網 making or mending a net
G6 hunting 畋獵 going on a hunting expedition
G7 domestication 牧養納畜 buying and domesticating fouls or animals
G8 enclosure 造畜欄 building enclosures for domestic animals
G9 training 教牛馬 training horse or cow for transportation

35
Table 2: CATREG Analysis Outcomes of the Markets of TW, SK, and SG

Market: TW
ANOVA

Sum of Squares df Mean Square F Sig.

Regression 1801.332 2 900.666 3084.525 .000

Residual 741.668 2540 .292

Total 2543.000 2542

Coefficients

Standardized Coefficients

Bootstrap (1000) Estimate of Std.

Beta Error df F Sig.

C9 .000 .001 0 .000 .

TURNOVER .024 .014 1 3.047 .081
WINORLOS .772 .005 1 20499.149 .000
E

Correlations and Tolerance

Correlations Tolerance

Before
Zero-Order Partial Part After Transformation Transformation

Category A
Max 0.034 0.05 0.026 0.616 0.615

Min -0.008 -0.018 -0.01 0.133 0.133

Average 0.00977778 0.0093333 0.0048333 0.3843889 0.3841667

Category B
0.032 0.02 0.011 0.728 0.728
Max
-0.006 -0.023 -0.012 0.032 0.032
Min

36
 0.01055556 0.0056667 0.0031111 0.3615556 0.361
Average

Category C
0.046 0.045 0.024 0.936 0.764
Max
-0.008 -0.025 -0.013 0.01 0.01
Min
0.01190909 0.0101818 0.0053636 0.399 0.3532727
Average

Category D
Max 0.036 0.022 0.012 0.596 0.597

Min 0.001 -0.003 -0.002 0.148 0.148

Average 0.01441667 0.0090833 0.00475 0.39025 0.3900833

Category E
Max 0.014 0.05 0.026 0.576 0.577

Min -0.005 -0.033 -0.018 0.037 0.037

Average 0.0008 0.0046 0.0023 0.2302 0.2303

Category F
Max 0.022 0.032 0.017 0.589 0.589

Min -0.006 -0.025 -0.013 0.159 0.158

Average 0.00366667 0.0076667 0.0041111 0.357 0.3565556

Category G
Max 0.016 0.034 0.018 0.454 0.454

Min -0.008 -0.036 -0.019 0.023 0.023

Average 0.002 0.0054 0.0028 0.2081 0.208

.001 -.091 -.048 .011 .011

OPEN
.018 .028 .015 .009 .009
HIGH
.023 .002 .001 .009 .009
LOW
.039 .059 .031 .011 .011
CLOSE
.141 .069 .036 .856 .854
TURNOVER
.844 .832 .786 .908 .908
WINORLOSE

Market: SK
ANOVA

37
 Sum of Squares df Mean Square F Sig.

Regression 1496.784 1 1496.784 3889.140 .000

Residual 935.216 2430 .385

Total 2432.000 2431

Coefficients

Standardized Coefficients

Bootstrap (1000) Estimate of Std.

Beta Error df F Sig.

CLOSE .000 .001 0 .000 .

TURNOVER .000 .001 0 .000 .
WINORLOSE .709 .005 1 17653.601 .000

Correlations and Tolerance

Correlations Tolerance

Before
Zero-Order Partial Part After Transformation Transformation

Category A
Max 0.051 0.043 0.026 0.766 0.492

Min 0.008 -0.012 -0.007 0.04 0.044

Average 0.02638889 0.0082778 0.005 0.5168889 0.2577778

Category B
0.034 0.017 0.01 0.669 0.622
Max
0.01 -0.008 -0.005 0.024 0.011
Min
0.01877778 0.0051111 0.003 0.357 0.2165556
Average

Category C
0.051 0.035 0.021 0.825 0.68
Max
0.013 -0.021 -0.012 0.164 0.003
Min
0.02645455 0.0056364 0.0033636 0.4134545 0.2117273
Average

38
 Category D
Max 0.032 0.035 0.021 0.763 0.62

Min -0.002 -0.01 -0.006 0.143 0.052

Average 0.011 0.0101667 0.0061667 0.4484167 0.2360833

Category E
Max 0.044 0.049 0.029 0.824 0.47

Min 0.007 -0.037 -0.022 0.065 0.013

Average 0.0234 0.0149 0.0088 0.4507 0.1356

Category F
Max 0.051 0.042 0.025 0.851 0.39

Min 0.007 -0.028 -0.016 0.123 0.062

Average 0.021 0.011 0.0065556 0.5165556 0.1922222

Category G
Max 0.021 0.043 0.026 0.745 0.283

Min 0.008 -0.016 -0.01 0.02 0.006

Average 0.014 0.0142 0.0084 0.3663 0.0995

0.047 0.116 0.07 0.013 0.013

OPEN
-0.003 -0.102 -0.061 0.014 0.014
HIGH
0.021 -0.03 -0.018 0.012 0.012
LOW
0.025 0.02 0.012 0.014 0.014
CLOSE
0.08 0.033 0.02 0.757 0.758
TURNOVER
0.789 0.772 0.723 0.9 0.903
WINORLOSE

Market: SG

ANOVA

Sum of Squares df Mean Square F Sig.

Regression 1365.898 1 1365.898 3113.333 .000

Residual 1066.102 2430 .439

Total 2432.000 2431

39
 Coefficients

Standardized Coefficients

Bootstrap (1000) Estimate of Std.

Beta Error df F Sig.

TURNOVER .000 .001 0 .000 .

WINORLOSE .665 .005 1 16088.368 .000

Correlations and Tolerance

Correlations Tolerance

Zero-Order Partial Part After Transformation Before Transformation

Category A
Max 0.014 0.047 0.03 0.491 0.491

Min -0.033 -0.03 -0.019 0.044 0.044

Average -0.01194444 0.0024444 0.0014444 0.2573889 0.2573889

Category B
-0.002 0.038 0.024 0.62 0.62
Max
-0.014 -0.034 -0.021 0.011 0.011
Min
-0.00822222 0.001 0.0006667 0.2162222 0.2162222
Average

Category C
0.005 0.015 0.009 0.679 0.679
Max
-0.032 -0.025 -0.016 0.003 0.003
Min
-0.01418182 -0.0046364 -0.0029091 0.2115455 0.2115455
Average

Category D
Max 0.015 0.03 0.019 0.619 0.619

Min -0.013 -0.051 -0.032 0.052 0.052

Average -0.00075 -0.0054167 -0.0035833 0.23575 0.23575

Category E
Max -0.002 0.024 0.015 0.469 0.469

40
 Min -0.021 -0.036 -0.023 0.013 0.013

Average -0.0096 -0.0018 -0.0014 0.1354 0.1354

Category F
Max -0.002 0.029 0.018 0.39 0.39

Min -0.028 -0.05 -0.032 0.062 0.062

Average -0.01144444 -0.0027778 -0.0017778 0.1921111 0.1921111

Category G
Max 0.009 0.039 0.025 0.283 0.283

Min -0.016 -0.032 -0.02 0.006 0.006

Average -0.0065 0.0053 0.0035 0.0994 0.0994

0.036 0.116 0.074 0.013 0.013

OPEN
-0.014 -0.11 -0.07 0.014 0.014
HIGH
0.01 -0.017 -0.011 0.012 0.012
LOW
0.013 0.012 0.008 0.014 0.014
CLOSE
0.077 0.034 0.022 0.758 0.758
TURNOVER
0.755 0.737 0.692 0.903 0.903
WINORLOSE

Note:
1. In the Coefficient sections, the coefficient estimates only show those are different from zero.
2. In the Correlation and Tolerance section, for each predictor:
(1) Sub-column Zero-Order refers to the zero-order correlation which is the standard Pearson correlation between each
quantified variable and the response.
(2) Sub-column Partial refers to the partial correlation which is the correlation after removing the effect of the other
predictors from both the quantified variable and the response.
(3) Sub-column Part refers to the part correlation is the correlation after removing the effect of the other predictors from
the quantified variable.
(4) Column Tolerance refers to the tolerance for optimally scaled predictor variable which represents the degree of
independence. Small tolerance means multi-collinearity and possibly unstable regression weights.
(5) The detailed information of CFC advices is summarized and presented in CFC categories by three statistics: maximum
(Max), minimum (Min), and mean (Average).

Table 3: Risk Estimate and Its Standard Error of CART Analysis

Market SG SK TW
Sample Estimate Std. Error Estimate Std. Error Estimate Std. Error
Training .000 .000 .000 .000 1.111 .056
Test .000 .000 .000 .000 1.274 .133

41
 0.00
200.00
400.00
600.00
800.00
1,000.00
1,200.00

0.00
0.00

10,000.00
15,000.00
20,000.00
25,000.00
10,000.00
12,000.00
14,000.00
16,000.00
18,000.00
20,000.00

5,000.00
2,000.00
4,000.00
6,000.00
8,000.00
12/30/94
12/30/94 30/12/94
12/30/95 12/30/95 09/10/95
12/30/96 16/07/96
12/30/96 23/04/97
12/30/97 12/30/97 29/01/98

JP
HK

12/30/98 12/30/98 06/11/98

INDO
12/30/99 16/08/99
12/30/99
23/05/00
12/30/00 12/30/00 28/02/2001
12/30/01 12/30/01 6/12/2001
12/30/02 12/30/02 13/09/2002
23/06/2003
12/30/03 12/30/03 30/03/2004
12/30/04 12/30/04
Figure 1: Time Series Plot of the Aggregate Market Index

1,000.00
1,200.00
0.00

0.00
200.00
400.00
600.00
800.00
1,000.00
1,200.00

200.00
400.00
600.00
800.00
1,000.000
1,500.000
2,000.000
2,500.000

0.000
500.000
12/30/94 01/05/95 12/30/94
12/30/95 01/05/96 12/30/95
12/30/96 01/05/97 12/30/96
12/30/97 01/05/98 12/30/97

SK
12/30/98
SG

01/05/99 12/30/98
SHA

12/30/99 01/05/00 12/30/99

12/30/00 01/05/01 12/30/00
12/30/01 01/05/02 12/30/01
12/30/02 01/05/03 12/30/02
12/30/03 01/05/04 12/30/03

42
12/30/04
12/30/04
 0.00
100.00
150.00
200.00
250.00
300.00
350.00
400.00

50.00

0.00
500.00
1,000.00
1,500.00
2,000.00
2,500.00
3,000.00
3,500.00
4,000.00
12/30/94
12/30/94
12/30/95
12/30/95
12/30/96
12/30/96
12/30/97
12/30/97
12/30/98

PH
12/30/98
ML

12/30/99 12/30/99

12/30/00 12/30/00

12/30/01 12/30/01
12/30/02 12/30/02
12/30/03 12/30/03
12/30/04 12/30/04

0
10000
12000

2000
4000
6000
8000
100.00
200.00
300.00
400.00
500.00
600.00
700.00

0.00

01/05/95 01/05/95
01/05/96 01/05/96
01/05/97 01/05/97
01/05/98 01/05/98
01/05/99
SZ

01/05/99
TW

01/05/00 01/05/00

01/05/01 01/05/01

01/05/02 01/05/02
01/05/03
01/05/03
01/05/04
01/05/04

43
 1000
1500
2000
2500

500

0
A1
A2

Frequency
A3
A4
A5
A6
A7
A8
A9
A10
A11
A12
A13
A14
A15
A16
A17
A18
B1
B2
B3
B4
Figure 2: Frequency of CFC Item

B5
B6
B7
B8
B9
C1
C2
C3
C4
C5
C6
C7
C8
C9
C10
C11
D1
D2
D3
D4

CFC Item
D5
D6
D7
D8
D9
D10
D11
D12
E1
E2
E3
E4
E5
E6
E7
E8
E9
E10
F1
F2
F3
F4
F5
F6
F7
F8
F9
G1
G2
G3
G4
G5
NEUTRAL

G6
G7
AUSPICIOUS

G8
G9
INAUSPICIUOS

44
G10
Figure 3: Frequency of CFC Category

90.00%

80.00%

70.00%

60.00%
NEUTRAL

AUSPICIOUS

INAUSPICIUOS
50.00%
Percentage

40.00%

30.00%

20.00%

10.00%

0.00%
A B C D E F G
CFA Category

45
Figure 4: Frequency of Market Status

58.00%

56.00%

0
1
54.00%

52.00%

50.00%
Percentage

48.00%

46.00%

44.00%

42.00%

40.00%
HK INDO JP ML PH SG SHA SK SZ TW
Market

Note: “1” indicates that the daily rate of return is positive, and “0” indicates that the daily rate of return is zero and negative.

46
Figure 5: Summary of Adjusted R Squares of Ten Regressions among the Ten Markets

0.8

0.7

0.6

0.5
Adjusted R squares

0.4
Regression 1
Regression 2
Regression 3
Regression 4
0.3
Regression 5
Regression 6
Regression 7
Regression 8
0.2
Regression 9
Regression 10

0.1

0
TW SK SG SZ SHA PH MA JP INDO HK

47
Figure 6: CART Analysis of TW Market

Note: The rectangle represents the node of a selected variable. The number of bars attached below the rectangle represents the
number of classified categories under the specific variable in the rectangle. The circle represented the terminal node. The number
of layer of the tree is defaulted as five.

48
Figure 7: CART Analysis of SK Market

Note: The rectangle represents the node of a selected variable. The number of bars attached below the rectangle represents the
number of classified categories under the specific variable in the rectangle. The circle represented the terminal node. The number
of layer of the tree is defaulted as five.

49
Figure 8: CART Analysis of SG Market

Note: The rectangle represents the node of a selected variable. The number of bars attached below the rectangle represents the
number of classified categories under the specific variable in the rectangle. The circle represented the terminal node. The number
of layer of the tree is defaulted as five.

50
References

Block, Lauren, and Thomas Kramer. 2009. The effect of superstitious beliefs on performance
expectations. Journal of the Academy of Marketing Science 37 (2):161-169.
Breiman, L., J. H. Friedman, R.A. Olshen, and C. J. Stone. 1998. Classification and Regression
Trees. New York, NY: Chapman & Hall.
Brown, Philip, Angeline Chua, and Jason Mitchell. 2002. The influence of cultural factors on
price clustering: Evidence from Asia–Pacific stock markets. Pacific-Basin Finance
Journal 10 (3):307-332.
Buja, A., T. J. Hasti, and R. J. Tibshirani. 1989. Linear smoothers and additive models. Annals of
Statistics 17:453-555.
Campbell, John Y., Andrew W. Lo, and A. Craig MacKinlay. 1997. The Econometrics of
Financial Markets. New Jersey: Princeton University Press.
Cao, Melanie, and Jason Wei. 2005. Stock market returns: A note on temperature anomaly.
Journal of Banking and Finance 29 (6):1559-1573.
Chamberlain, Trevor W., C. Sherman Cheung, and Clarence C. Y. Kwan. 1991. The Friday the
thirteenth effect: Myth or reality? Quarterly Journal of Business and Economics 30
(2):111-117.
Cho, In-Koo. 2005. Introduction to learning and bounded rationality. Journal of Economic
Theory 124 (2):127-128.
Dichev, Ilia D., and Troy D. Janes. 2002. Lunar cycle effects in stock returns. Journal of Private
Equity 6 (4):8-29.
Efron, R. E. 1983. Estimating the error rate of a prediction rule: Improvement on cross-validation.
Journal of the American Statistical Association 78:316-313.
Feldman, David, and Shulamith Gross. 2005. Mortgage default: Classification trees analysis.
Journal of Real Estate Finance and Economics 30 (4):369-396.
Gilovich, Thomas, Dale Griffin, and Daniel Kahneman, eds. 2002. Heuristics and Biases: The
Psychology of Intuitive Judgment. New York: Cambridge University Press.
Hastie, Trevor, Robert Tibshirani, and Jerome Friedman. 2009. The Elements of Statistical
Learning- Data Mining, Inference, and Prediction. 2nd ed: Springer.
Hirshleifer, David. 2001. Investor psychology and asset pricing. Journal of Finance 56 (4):1533-
1597.
Izenman, Alan Julian. 2008. Modern Multivariate Statistical Techniques- Regression,
Classification, and Manifold Learning: Springer.
Kahneman, D., and A. Tversky. 1973. On the psychology of prediction. Psychologicaal Review
80:237-251.
Kolb, R. W., and R. J. Rodriguez. 1988. Friday the Thirteen: Part VII- A note. Journal of
Finance 42:1385-1387.
Kooij, Anita J. van der. 2007. Prediction Accuracy and Stability of Regression with Optimal
Scaling Transformations, Leiden University.
Kramer, Thomas, and Lauren Block. 2008. Conscious and nonconscious components of
superstitious beliefs in judgment and decision making. Journal of Consumer Research 34
(6):783-793.
Loughran, Tim, and Paul Schultz. 2004. Weather, stock returns, and the impact of localized
trading behavior. Journal of Financial & Quantitative Analysis 39 (2):343-364.
Meulman, Jacqueline J., and Willem J. Heiser. 2007. PASW Categories 18: SPSS Inc.
Ng, Travis, Terence Chonga, and Xin Du. 2010. The value of superstitions. Journal of Economic
Psychology 31 (3):293-309.
51
Oyserman, D., H. M. Coon, and M. Kemmelmeier. 2002. Rethinking individualism and
collectivism: Evaluation of theoretical assumptions and meta-analysis. Psychological
Bulletin 128 (1):3-72.
Saunders, E. 1993. Stock prices and Wall St. weather. American Economic Review 83:1337-1345.
Schwert, G. William. 2003. Anomalies and Market Efficiency. In Handbook of the Economics of
Finance, edited by G. Constantinides, M. Harris and R. M. Stulz. North-Holland.
Thaler, R. H. 1985. Mental accounting and consumer choice. Marketing Science 4:199-214.
Tibshirani, R. J. 1996. Regression shrinkage and selection via Lasso. Journal of the Royal
Statistical Society B 58:267-288.
van der Kooij, A.J., Jacqueline J. Meulman, and Willem J. Heiser. 2006. Local minima in
categorical multiple regression. Computational Statistics and Data Analysis 50 (2):446-
462.
William D. Perreault, Jr., and Forrest W. Young. 1980. Alternating least squares optimal scaling:
Analysis of nonmetric data in marketing research. Journal of Marketing Research 17
(1):1-13.
Woo, C.-K., I. Horowitz, S. Luk, and A. Lai. 2008. Willingness to pay and nuanced cultural cues:
Evidence from Hong Kong’s license-plate auction market. Journal of Economic
Psychology 29 (1):35-53.
Yuan, Kathy, Lu Zheng, and Qiaoqiao Zhu. 2006. Are investors moonstruck? Lunar phases and
stock returns. Journal of Empirical Finance 13 (1):1-23.
Zou, H., and T. J. Hastie. 2005. Regularization and variable selection via the elastic net. Journal
of the Royal Statistical Society B 67:301-320.

52