You are on page 1of 29

Module 1: Building Blocks of Finance Books

The Random Behaviour of Assets


Different types of financial analysis

Examining time-series data to model returns

Random nature of prices

The need for probabilistic models

1. Applied Stochastic
2. Notes on Brownian
motion and related
The Wiener process, a mathematical model of randomness
phenomena
3. Random Walk 2
4. Random Walk

The lognormal random walk- The most important model for


equities, currencies, commodities and indices
PDE's and Transition Density Functions
Taylor series

A trinomial random walk

Transition density functions

Our first stochastic differential equation

Similarity reduction to solve partial differential equations

Fokker-Planck and Kolmogorov equations

Applied Stochastic Calculus 1


Moment Generating Function

Construction of Brownian Motion/Wiener Process

Functions of a stochastic variable and Itô’s Lemma

Applied Itô calculus

Stochastic Integration

The Itô Integral

Examples of popular Stochastic Differential Equations

Applied Stochastic Calculus 2


Extensions of Itô’s Lemma

Important Cases - Equities and Interest rates

Producing standardised Normal random variables

The steady state distribution

Binomial Model
A simple model for an asset price random walk

Delta hedging

No arbitrage

The basics of the binomial method for valuing options

Risk neutrality
Discrete Martingales
Binomial Model extended

The Probabilistic System: sample space, filtration, measures

Conditional and unconditional expectation

Change of measure and Radon-Nikodym derivative

Continuous Martingales
What is a continuous time martingale?

Martingales and Itô calculus

A detour to explore some further Ito calculus

Exponential martingales, Girsanov and change of measure


The martingale zoology in a hurry: martingale, local martingale,
supermartingale, submartingales, semimartingales
On the average, how far is the walker from the starting point?
The displacement x in the direction of the X-axis that a particle experiences on the average or, more exactly, the square root of the arithmetic mean of the square of
the displacement in the direction of the X-axis; it is
x= Square root(2Dt), t is time , D is diffusion Coeffiecient D= 1/T Integration from - infinifty to + Infinity (Deltasquare / 2) X Pfobality density function ) integrated
accross space)

What is the probability that at a particular time the walker is at the origin?

The probability that a (one-dimensional) simple random walker returns to the origin infinitely often is one.

More generally, what is the probability distribution for the position of the walker?

The probability of being at position m after N jumps is given by binomial distribution.

Does the random walker keep returning to the origin or does the walker eventually leave forever?

A return to the origin, often referred to as an equalization, occurs when Sn equals 0 for some n greater than 0. If an in nite number of equalizations occur, then the
walk is called recurrent. If only a nite number of equalizations occur, then the walk is called transient.
Module 2: Quantitative Risk & Return

Portfolio
Management
Measuring risk and return

Benefits of diversification
Modern Portfolio Theory
and the Capital Asset
Pricing Model
The efficient frontier

Optimising your portfolio


How to analyse portfolio
performance
Alphas and Betas

Fundamentals of
Optimization and
Application to
Portfolio Selection
Fundamentals of portfolio
optimization
Formulation of
optimization problems
Solving unconstrained
problems using calculus
Kuhn-Tucker conditions

Derivation of CAPM

Risk Regulation and


Basel III
Definition of capital

Evolution of Basel

Basel III and market risk

Key provisions

Value at Risk &


Expected Shortfall
Measuring Risk

VaR and Stressed VaR


Expected Shortfall and
Liquidity Horizons
Correlation Everywhere
Frontiers: Extreme Value
Theory
Liquidity Asset
Liability
Management
Gap analysis, of assets,
liabilities and
contingencies, both static
and dynamic
The role of derivative and
non-derivative
instruments in liquidity
Liquidity Coverage Ratio
(LCR)
Net Stable Funding Rate
(NSFR)
Collateral and
Margins
Expected Exposure (EE)
profiles for various types
of instruments
Types of Collateral
Calculation Initial and
Variation Margins
Minimum transfer amount
(MTA)
ISDA / CSA
documentation
Module 3: Equities & Currencies

Black-Scholes Model
The assumptions that go into the
Black-Scholes equation
Foundations of options theory: delta
hedging and no arbitrage
The Black-Scholes partial
differential equation
Modifying the equation for
commodity and currency options
The Black-Scholes formulae for
calls, puts and simple digitals
The meaning and importance of the
Greeks, delta, gamma, theta, vega
and rho
American options and early
exercise
Relationship between option values
and expectations
Martingale Theory –
Applications to Option
Pricing
The Greeks in detail

Delta, gamma, theta, vega and rho

Higher-order Greeks

How traders use the Greeks

Martingales and PDEs:


Which, When and Why
Computing the price of a derivative
as an expectation
Girsanov's theorem and change of
measures
The fundamental asset pricing
formula
The Black-Scholes Formula

The Feynman-K_ac formula


Extensions to Black-Scholes:
dividends and time-dependent
parameters
Black's formula for options on
futures
Understanding Volatility
The many types of volatility
The market prices of options tells us
about volatility
The term structure of volatility

Volatility skews and smiles


Volatility arbitrage: Should you
hedge using implied or actual
volatility?
Introduction to Numerical
Methods
The justification for pricing by Monte
Carlo simulation
Grids and discretization of
derivatives

The explicit finite-difference method

Further Numerical Methods


Implicit finite-difference methods
including Crank-Nicolson schemes
Douglas schemes

Richardson extrapolation
American-style exercise
Explicit finite-difference method for
two-factor models
ADI and Hopscotch methods

Exotic Options
Characterisation of exotic options
Time dependence (Bermudian
options)
Path dependence and embedded
decisions
Asian options

Derivatives Market Practice

Option traders now and then

Put-Call Parity in early 1900

Options Arbitrage Between London


and New York (Nelson 1904)

Delta Hedging

Arbitrage in early 1900

Fat-Tails in Price Data

Some of the Big Ideas in Finance

Dynamic Delta Hedging

Bates Jump-Diffusion

Advanced Greeks
The names and contract details for
basic types of exotic options
How to classify exotic options
according to important features
How to compare and contrast
different contracts
Pricing exotics using Monte Carlo
simulation
Pricing exotics via partial differential
equations and then finite difference
methods
Advanced Volatility
Modelling in Complete
Markets
The relationship between implied
volatility and actual volatility in a
deterministic world
The difference between 'random'
and 'uncertain'
How to price contracts when
volatility, interest rate and dividend
are uncertain
Non-linear pricing equations
Optimal static hedging with traded
options
How non-linear equations make a
mockery of calibration
Market-Based Valuation of
Equity Index Options
Stylized Facts of Equity & Options
Markets
Numerically efficient valuation of
equity index options
Calibration of option pricing models
to market data
Simulation of option pricing models
for European & American options
Canonical Example
Euro Stoxx 50 index and options

Merton (1976) jump-diffusion model

Python implementation
Module 4: Fixed Income
Fixed Income Products
and Analysis
Names and properties of the
basic and most important fixed-
income products
Features commonly found in
fixed-income products
Simple ways to analyze the
market value of the instruments:
yield, duration and convexity
How to construct yield curves and
forward rates
Swaps
The relationship between swaps
and zero-coupon bonds
Stochastic Interest Rate
Modeling
Stochastic models for interest
rates

How to derive the pricing equation


for many fixed-income products

The structure of many popular


one-factor interest rate models

The theoretical framework for


multi-factor interest rate modeling

Popular two-factor models

Calibration and Data


Analysis
How to choose time-dependent
parameters in one-factor models
so that
Today’s yield curve is an output of
the model
The advantages and
disadvantages of yield curve
fitting
How to analyze short-term
interest rates to determine the
best model for the volatility and
the real drift
How to analyze the slope of the
yield curve to get information
about the market price of risk

Probabilistic methods for


interest rates
The pricing of interest rate
products in a probabilistic setting
The equivalent martingale
measures
The fundamental asset pricing
formula for bonds
Application for popular interest
rates models
The dynamics of bond prices

The forward measure


The fundamental asset pricing
formula for derivatives on bonds
Heath Jarrow and Morton
Model
The Heath, Jarrow & Morton
(HJM) forward rate model
The relationship between HJM
and spot rate models
The advantages and
disadvantages of the HJM
approach
How to decompose the random
movements of the forward rate
curve into its principal
components
Fixed Income Market
Practices
Basics: discount factors, FRAs,
swaps, and other delta products
Basic curve stripping, bucket
deltas, and managing IR risks
Interpolation methods

Risk bleeding
Scenarios-based risks and
hedging (wave method)
Current Market Practices

Advanced stripping

Volatility Smiles and the


SABR Model
Vanilla options: European
swaptions, caps, and floors
Arbitrage Free SABR

The Libor Market Model


The Libor Market model

The market view of the yield curve

Yield curve discretisation


Standard Libor market model
dynamics
Numéraire and measure

The drift

Factor reduction

Further Monte Carlo


The connection to statistics
The basic Monte Carlo algorithm,
standard error and uniform
variates
Non-uniform variates, efficiency
ratio and yield
Co-dependence in multiple
dimensions
Wiener path construction;
Poisson path construction
Numerical integration for solving
SDEs
Variance reduction techniques

Sensitivity calculations

Weighted Monte Carlo

Energy Derivatives
Participants' risk exposure in
electricity markets
Price volatility
Fluctuation of electricity
production costs
Hedging with electricity futures for
generators, marketers and end-
users
Risks of hedging with electricity
futures
"Stack and Roll" hedging for the
long-term periods
"Hedging with electricity options
and crack spreads
Speculation (views taking) using
electricity futures
Politics of Speculation. Dodd-
Frank Act

Restrictions to oil speculation and


high frequency commodity trading

Energy trading case study. Sparks


spread
Module 5: Credit Products and Risk
Books
Introduction to Credit
Derivatives & Structural
Models
Introduction to credit risk

1. CREDIT RISK
Rating based 1. Sturctural models model credit risk based on assuming a stochastic
modeling of credit process for the value of the firm and the term structure of interest rates.
Modelling credit risk risk, Theory and Clearly the problem is to determine the value and volatility of the firm’s
application of assets and to model the stochastic process driving
migration the value of the firm adequately.
matricces.pdf
Basic structural models: Merton
Model, Black and Cox Model
Advanced structural models

Intensity Models
Modelling default by Poisson
Process
Relationship between intensity
and arrival time of default
Risky bond pricing: constant vs.
stochastic hazard rate
Bond pricing with recovery

Theory of affine models


Affine intensity models and use of
Feynman-Kac
Two-factor affine intensity model
example: Vasicek
Credit Default Swaps
An Introduction to CDS
Default Modelling Toolkit.
Inhomogenous Poisson Process
CDS Pricing: Basic and
Advanced Models
Bootstrapping intensity from CDS
market quotes
Accruals and upfront premium in
CDS pricing
X-Valuation Adjustment
(CVA, DVA, FVA, MVA)
Theory
Historical development of OTC
derivatives and xVA
Credit and debt value
adjustments (CVA and DVA)
Funding value adjustment (FVA)
Margin and Capital Value
Adjustments (MVA and KVA)
Current market practice and
application
X-Valuation Adjustment
(CVA, DVA, FVA, MVA)
Implementation
Implementation of counterparty
credit valuation adjustment (CVA)

Review the numerical


methodologies currently used to
quantify CVA in terms of
exposure and Monte Carol
simulation and the Libor Market
Model
Illustrate this methodology as
well as DVA, FVA and others
CDO & Correlation Sense
CDO market pricing and risk
management
Loss Function and CDO Pricing
Equation
Motivation from loss distribution

What is Copula Function


Classification of Copula
Functions
Simulating via Gaussian Copula

3 Gaussian Copula Factor mode

The meaning of correlation.


Intuition and timescale

Linear Correlation and its misuse

Rank Correlation

Correlation in exotic options


Uncertain correlation model for
Mezzanine tranche
Compound (implied) correlation
in Loss Distribution
Statistical Methods in
Estimating Default
Probability
Sources of default probability
information
Capital Structure Arbitrage
Generalized Linear Models
(GLM): theory, estimation and
inference

Exponential family of distributions

Logit and probit regressions. Link


functions
Estimation of default probability
for an enterprise with a logit
model

Sovereign credit rating transitions


and the ordered probit model

Other Topics
1. They are more general than structural models and assume that an
exogenous random variable drives default and that the probability of default
(PD) over any time interval is non-zero. An important input to determine the
default probability and the price of a bond is the rating of the company. Thus,
to determine the risk of a credit portfolio of rated issuers one generally has to
consider historical average defaults and transition probabilities for current
rating classes. Quite often in reduced form approaches the migration from
Reduced Form Models
one rating state to another is modeled using a Markov chain model with a
migration matrix governing the changes from one rating state to another.
Besides the fact that they allow for realistic short-term credit spreads,
reduced form models also give great flexibility in specifying the source of
default.

2.
1. Analysts expect financial information about the company consisting of five
years of audited annual financial statements,the last several interim financial
statements, and narrative descriptions of operations and products. The
meeting with corporate management can be considered an important part of
an agency’s rating process. The purpose is to review in detail the company’s
key operating and financing plans, management policies, and other credit
1. CREDIT RISK
factors that have an impact on the rating.
Rating based
modeling of credit
A:
risk, Theory and
Although credit rating and credit score may be used interchangeably in
application of
some cases, there is a distinction between these two phrases. A credit
migration
rating, often expressed as a letter grade, conveys the creditworthiness of a
matricces.pdf
business or government. A credit score is also an expression of
creditworthiness, but it is expressed in numerical form and only used for
2.219148718-The-
individuals. Both ratings and scores are designed to show creditors a
Credit-Scoring-
borrower's likelihood of repaying a debt.
Toolkit-R-
Anderson.pdf

3. Credit Risk
2. Rating Facotrs: Business Risk || Financial Risk
Scorecards
Industry Characteristics || Financial Characteristics
Rating process Developing and
Competitive Position || Financial Policy
Implementing
Marketing || Profitability
Intelligent Credit
Technology || Capital Structure
Scoring.pdf
Efficiency || Cash Flow Protection
Regulation || Financial Flexibility
4.74973105-
Management
Consumer-Credit-
Source: S&P’s Corporate Ratings Criteria (2000)
Models-Pricing-
Profit-and-
In the world of emerging markets, rating agencies usually also incorporate
Portfolios-2009
country and sovereign risk to their rating analysis. Both business risk factors
such as macroeconomic volatility, exchange-rate risk, government
5. 242637835-Model-
regulation, taxes, legal issues, etc., and financial risk factors such as
Risk
accounting standards, potential price controls, inflation, and access.
6.rating_models_tc
3.The borrowers who share a similar risk profile are assigned to the same
m16-22933
rating grade. Afterwards a probability of default (PD) is assigned. Very often
the same PD is assigned to all borrowers of the same rating grade. For such
a rating methodology the PDs do not discriminate between better and lower
creditworthiness inside one rating grade. Consequently, the probability to
migrate to a certain other rating grade is the same for all borrowers having
the same rating grade.
1.A point-in-time (PIT) estimate refers to immediate In some cases, using several
probabilities, typically one-year, that will fluctuate up scorecards for a portfolio provides
and down over the course of an economic cycle. In Objectives better risk differentiation than using
contrast, a through-the-cycle (TTC) estimate is one that one scorecard on everyone. This is
approximates a stressed bottom-of-the-cycle scenario, • Reduction in bad usually the segmentation case
with a horizon of five years or more. debt/bankruptcy/claims/fraud where a population is made up of
• Increase in approval rates or market share distinct subpopulations, and
2, According to Aguais (2005), no risk estimate will ever in areas such as secured where one scorecard will not work
be purely PIT or TTC, but will+D56 loans, where low delinquency presents efficiently for all of them (i.e., we
always be a combination of the two. For example, expansion opportunities assume that different characteristics
default estimates based upon account performance • Increased profitability are required to predict risk for the
or the value of traded securities tend towards PIT, while • Increased operational efficiency (e.g., to different subpopulations in our
rating agency grades tend better manage workflow portfolio).
towards TTC. All of them will vary over an economic in an adjudication environment)
cycle, some more than others. • Cost savings or faster turnaround through 1. Generating segmentation ideas
Companies’ own internal grades are made up of a automation of adjudication based on experience and industry
combination of ‘subjective assessments, using scorecards knowledge, and then validating
statistical models, market information, and agency • Better predictive power (compared to these ideas using analytics
ratings’ that have a mixture of different existing custom or bureau 2. Generating unique segments
time horizons. While it may be ideal to provide separate scorecard) using statistical techniques such as
PIT and TTC estimates for each clustering or decision trees
obligor, this is beyond the capabilities of today’s banks, Data
and has not been required by Basel II. Typical segmentation areas used in
The following data items are usually the industry include those based on:
3. Stages of Scorecard Development collected for applications from the previous
two to five years, or from a large enough • Demographics. Regional
Feasibility study sample: province/state, internal definition,
• Account/identification number urban/rural, postal-code based,
Data—Will there be sufficient data available to develop • Date opened or applied neighborhood), age, lifestyle code,
a model? • Arrears/claims history over the life of the time at bureau, tenure at bank
Resources—Will there be money and people available? account
Technology—Is the technology available to support us? • Accept/reject indicator • Product Type. Gold/platinum cards,
• Product/channel and other segment length of mortgage, insurance type,
Player Identification identifiers secured/unsecured, new versus
• Current account status (e.g., inactive, used leases for auto, size of loan.
Project manager—Reports to the champion, and must closed, lost, stolen, fraud, etc.
advise of any resource requirements • Sources of Business (channel).
or shortfalls. As a rule of thumb, for application corecard Store-front, take one, branch,
Scorecard developer—Develops the actual scorecard. development there should be pproximately internet, dealers, brokers
Internal analysts—Assist in assembling and 2,000 “bad” accounts and 2,000 “good” segmentation
understanding data. accounts that can be randomly selected for
Functional experts—Assist in understanding the each proposed scorecard, from a group of
business and affected areas, and will be key when approved accounts opened within a defined Statistically-Based Segmentation
deciding upon the strategies to be employed. time frame. For behavior scorecards, these
Technical resources—Responsible for final would be from a group of accounts that Clustering
implementation of the scorecard. were current at a given point in time, or at a
certain delinquency status for collections Clustering is a widely used
Data Preparation scoring. A further 2,000 technique to identify groups that
declined applications may also be required are similar to each other with
Project scope—Which cases are to be included? for application scorecards where reject respect to the input variables.
Good/bad definition—What is to be predicted? inference is to be performed. Clustering, which can be used to
Sample windows—What will be the observation and segment databases, places objects
outcome periods? .Exclusions into groups, or “clusters,” suggested
The first step is to measure the Behavioural scores estimate the The use of nonlinear model functions as well as the maximum likelihood
improvement in predictive power risk that the borrower will default method to optimize those functions means that regression models also
through segmentation. This can be in the next 12 months. They are make it possible to calculate membership probabilities and thus to
done using a number of statistics obtained by taking a sample of determine default probabilities directly from the model function.
such as the Kolmogorov-Smirnov previous borrowers and relating
(KS), c-statistic, and so on. Exhibit
their characteristics, including their The curves of the model functions and their mathematical representation
4.13 shows an example of this repayment, arrears and usage are shown in chart 17. In this chart, the function denotes the cumulative
analysis using the c-statistic (details
during a performance period, with standard normal distribution, and the term (P) stands for a linear
on their default status 12 months after combination of the factors input into the rating model; this combination can
the c-statistic are covered later, in
the end of that performance also contain a constant term. By rescaling the linear term, both model
Chapter 6). period.Some of these functions can be adjusted to yield almost identical results. The results of the
characteristics indicate whether two model types are therefore not substantially different.
We assume that a proper or the borrower can afford to repay
sufficient score s(x) captures as the loan, but the most important Due to their relative ease of mathematical representation, logit models are
much information for predicting the characteristics are usually those used more frequently for rating modeling in practice. The general manner in
probability of a performance from the credit bureau and which regression models work is therefore only discussed here using the
outcome, say good/bad, as does the information on the arrears status of logistic regression model (logit model) as an example.
original data vector, x. the borrower.
P = 1/ (1+ exp ( - (b0 + b1 K1 + b2 K2 + bnKn)))
Now we consider three approaches A borrower is usually assumed to
to modelling the credit risk of have defaulted if their payments In this formula, n refers to the number of financial indicators included in the
portfolios of consumer loans, all of on the loan are more than 90 days scoring function, Ki refers to the specific value of the creditworthiness
which are based on the behavioural overdue. If we define those who criterion, and bi stands for each indicator s coefficient within the scoring
scores of the individual borrowers have defaulted as “bad” (B) and function (for i ¼ 1; :::n). The constant b0 has a decisive impact on the value
who make up the portfolio. those who have not defaulted as of p (i.e. the probability of membership).
“good” (G), then the behavioural
The three models have analogies score is essentially a sufficient Selecting an S-shaped logistic function curve ensures that the p values fall
with the three main approaches to statistic of the probability of the between 0 and 1 and can thus be interpreted as actual probabilities. The
corporate credit-risk modelling: a borrower being good. typical curve of a logit function is shown again in relation to the result of the
structural approach, a reduced-form exponential function (score) in chart 18.
defaultmode approach and a Thus, if x are the characteristics of
ratings-based reduced-form the borrower, a score s(x) has the Years @ address
approach. property that P(G | x) = P(G | s(x)) <3 years
3–6 years
>6 years
Blank

Years @ employer
<2 years
2–8 years
9–20 years
>20 years
Blank

Home phone
Y/N

Accom. status
Own
Rent
Parents
Other
Application score—Used for new business In retail credit, lenders strive to automate as many
origination, and combines data from the decisions as possible. The motivation for human input
customer, past dealings, and the credit arises only where the models are known to be weak, the
bureaux. value at risk is large, and/or the potential profit is high.
Behavioural score—Used for account
management (limit setting, over-limit Instances where judgmental assessments dominate are
management, authorisations), and usually insovereign, corporate, and project-finance lending, where
focuses upon the behaviour of an individual the borrowers’ financial situation is complex, the information
account. is not standard and/or difficult to interpret, and volumes are
Collections score—Used as part of the extremely low.If there is a scoring model, it will only be
collections process, usually to drive predictive used for guidance, and the underwriter will assess other
diallers in outbound call centres, and information that has not been incorporated in the score.
incorporates behavioural, collections, and
bureau data. The internal versus external rating issue is primarily the
Customer score—Combines behaviour on bespoke versus generic debate:
many accounts, and is used for both account Most internal ratings are provided by bespoke models
management and cross-sales to existing developed specifically for a lender, while external ratings
customers. are generics based upon the experience of many lenders,
Bureau score—A score provided by the credit that are also available to competitors. For retail credit, the
bureau, usually a delinquency or bankruptcy latter may be provided by credit bureaux, or co-operatives.
predictor that summarises the data held by Generics are often used by lenders that: (i) are small, and
them. cannot provide sufficient data for a bespoke development;
(ii) are looking to enter new markets, where they have no
Markets where credit scoring is used today experience; or (iii) lack the technological sophistication to
include, but are not limited to: develop and implement a bespoke system.

Unsecured—Credit cards, personal loans, Application scoring traditionally has assessed a very
overdrafts. specific default risk. The most common risk considered is
Secured—Home loan mortgages, motor the chance that an applicant will go 90 days overdue on
vehicle finance. their payments in the next 12 months. What would happen
Store credit—Clothing, furniture, mail order. over other time periods and whether the customer is
Service provision—Phone contracts, proving profitable to the lender are aspects not considered
municipal accounts, short-term insurance. in this assessment.
Enterprise lending—Working-capital loans,
trade credit. Ascore, s(x), is a function of the characteristics’ attributes x
of a potential borrower which can be translated into the
Request -> Application probability estimate that the borrower will be good.The
Time -> Behavioural critical assumption in credit scoring is that the score is all
Entry -> Recovery that is required for predicting the probability of the applicant
Transaction -> Fraud being good. In some ways the score is like a sufficient
Event warning -> Behavioural/fraud statistic. It is also usual to assume that the score has a
Campaign -> Response monotonic increasing relationship with the probability of
being good – in that case the score is called a monotonic
Request—Customer application for a specific score.
product, or increased facilities.
Time—Regular recalculation, such as
monthly.
Entry—Calculated on first entry into a specific
stage
of the CRMC, and the resultant score is
retained for future use.
Transaction—Customer already has the
product, and scores are calculated each time
Module 6: Big Data and
Machine Learning
Big Data in Finance
What is Data science?

Supervised and unsupervised learning

Structured and unstructured data

Introduction to Classification
Bayesian Models and inference using
Markov chain Monte-Carlo
Introduction to graphical models: Bayesian
networks, Markov networks, inference in
graphical models
Optimisation techniques
Examples: Predictive analytics/trading &
Pricing
Classification, Clustering and
filtering
Classification: K-nearest neighbours,
optimal Bayes classifier, naïve Bayes, LDA
and QDA, reduced rank LDA, Logistic
regression, Support Vector Machines

Cluster analysis: BIRCH, Hierarchical, K-


mean, Expectation-maximization, DBSCAN,
OPTICS and Mean-shift
Kalman filtering

Examples (2 worked practical examples)

Machine Learning & Predictive


Analytics
Regression: liner regression, bias-variance
decomposition, subset selection, shrinkage
methods, regression in high dimensions

Support Vectors Machines: Classification


and regression using SVM’s and kernel
methods
Dimension reduction: Principal component
analysis (PCA), kernel PCA, non-negative
matrix decomposition, PageRank
Examples (2 worked examples)

Big Data Lab


Data analytics sandbox

Examples

Asset Returns: Key, Empirical


Stylised Facts
Volatility clustering: the concept and the
evidence
Properties of daily asset returns

Properties of high-frequency returns

Volatility Models: The ARCH


framework
Why ARCH models are popular

The original GARCH model

What makes a model an ARCH model?

Asymmetric ARCH models

Econometric methods

Co-Integration using R
Multivariate time series analysis
Financial time series: stationary and unit
root

Vector Autoregression, a theory-free model

Equilibrium and Error Correction Model

Eagle-Granger Procedure

Cointegrating relationships and their rank


Estimation of reduced rank regression:
Johansen Procedure
Stochastic modelling of equilibrium: Orstein-
Uhlenbeck process
Statistical arbitrage using mean reversion
Additional Content
Time Series Distribution
Poisson's Distribution
chart 18
m - number of characteristics included in the model
factor - scaling parameter based on formula presented
previously
offset - scaling parameter based on formula presented
previously

Neutral score is the calculated as:


Neutral score = sum(scorei*distri) where i = 1 to k, k- number
of bins
scorei- scoring assigned to the ith bin
distri- percentage distribution of the total cases in the
ith bin

Gini Cofficient = abs ( 1- sum((Gx(i) + Gx(i-1))*(Bx(i)-Bx(i-1))) i =


1 to k
where k number of categories of analyzed predictor
Gx(i) cumulative distribution of “good” cases in the ith
category
Bx(i) cumulative distribution of “bad” cases in the ith
category
Gx(0)=Bx(0)=0
Gini Coefficient = 2.AUC - 1

Monitoring the score card:


System Stability/Population stability report between recent
applicants and
expected (from development sample):
index = sum of ((% Actual -% Expected ) × ln ( % Actual / %
Expected))
An index of less than 0.10 shows no significant change,0.10–
0.25 denotes a small change that needs to be investigated, and
an index greater than 0.25 points to a significant shift in the
applicant population.
Characteristic Analysis Report:
“Expected %” and “Actual %” again refer to the distributions of
the development and recent samples, respectively. The index
here is calculated simply by:
sum of ((%Actual - % Expected) * Points))
of default.
Now it’s time to identity the drivers of the LGD modelling. There
are categorized under four categories
1. Obligor characteristics: In case of retail obligors, you can also
include socio-demographic variables, such as marital status,
gender (if allowed by the regulation), salary, time at address,
time at job, etc. Also, variables measuring the intensity of the
bank relationship such as number of years client, number of
products with the bank, can be considered. In case of
corporate, you can include industry sector, sector indicators,
size of the company, legal form of the company, age of the
company, balance sheet information such as revenues, total
assets, solvency, profitability, liquidity ratios, etc
2. Loan characteristics: Popular examples here are real estate,
cash, inventories, guarantees, etc. It is highly advised to qualify
the collateral as precisely as possible. So, in case of real estate,
you can further categorize the collateral into, for example, flat,
apartment, villa, detached/semi-detached house.
3. Country-specific regulations
4. Macroeconomic factors: Here you can think of gross
domestic product (GDP), which reflects economic growth,
aggregated default rates, inflation, unemployment rate, and the
interest rates, which obviously will have an impact on the
discount factor.
Modelling approach:
There are various approaches for modeling depending upon the
distribution of dependent and independent variable. It could be
one stage model or two stage model.
1. Linear regression
2. Linear regression with Beta transformation
3. Logistic regression
4. Decision tree
Downturn LGD: To calculate the downturn LGD, we take the
LGD of last 5 year and take the average of them for each year.
Take the LGD, which is highest as downturn LGD. This LGD is
used for calculation for expected losses for each pool.
a. Impute the missing value in training datasets with median
using proc stdize
b. There could be nominal variable that needs to be collapsed
using proc cluster and proc tree, they merge levels using chi
square statistics , proc means generates level and proportions
of events statistics
c. Remove redundancy: using proc varclus, it create variable
cluster based on second eignvalue, default is 1.00 you could
specify 0.7
d. Pick variable based on R square ratios from cluster. Rquare
ratio is ratio of (1-Rquare within dataset)/ (1- Rsquare value
with other clusters)
e. Run the proc Reg with VIF option to detect the co-linearity,
remove the variable with VIF greater than 5.
f. Now run corr and create a table of speaman and hoffding
value, remove values with low rank of spearsman and high
rankof hoffding value, they indicate nonlinearity, its called
nonlinearity screening
g. Next step to adjust outlier using percentile treatment

3. Now for each numeric variable create intervals and for each
category variable with different levels calculate WOEs and IVs.
Each bucket/bin should contain 5% of observations. The
variables with IV less than 0.1 and more than 0.5 should be kept
of analysis; they are under predictive or over predictive.

4. Now do the modelling (attrition -1, non attrition - 0) using


stepwise logistic regression (forward selection, backward
selection and stepwise selection) and subset selection (where
selection = score), Select the model, which predicts best
and explanatory. Cutoff is calculated using ROC.

5. Validate the model using the validation dataset, in case you


have done the oversampling, correct the sample bias by using
prievent option in Logistic Regression step. Use ROC, KS
statistics and Gini coefficient for judging the model
performance. KS should be nearly .5. For good behavioral score
card, it should be above .65
2. customers with 30 DPD, will go to stage 2
3. relative ratio of current PD and PD at origination, higher that a
predefined limit, depending upon certain tests as specified by
IFRS 9, will go to stage 2
4. all customer with forbearance status will go to stage 2
5. Rest of the customer, which are neither in stage 3 and stage 2,
will go stage 1
The PD, LGD and EAD models are built on initially for stage 1,
based on PD, we make decision for which customer, we should
build lifetime model. Life time means the remaining time in the
contract period.
The below description of LGD and EAD model is based on
simplest approach towards LGD at EAD calculation.
EAD Model:
In case of mortgage, there are mainly three kinds of products
-Amortization type: Where product is based on annuity
-linear
-Interest only mortgage
Depending upon the product type, EAD is calculated as
remaining balance at end of each year. There are two
adjustments applied each year:
-Prepayment rate at end of each year
-Drop in customer balance due to interest reset at interest reset
period, interest reset period could be 5 year, 10 years and 15
years
Amortization can be based on monthly interest rate or yearly
interest rate. At the end of contract duration, EAD is set to zero.
LGD Model:
Once EAD is calculated, LGD can be computed easily. Mortgage
is backed by collateral. For different scenarios- best, base and
worst, using the forward looking information collateral is
calculated for remaining years of product life. Two adjustments
are applied to calculated collateral
-Haircut is applied to collateral
-Cure rate is applied
-Interests and costs of sale added to EAD
LGD is calculated as follows:
LGD = min (1, max(Collateral-EAD,0)/EAD *CureRate)
measures considered here are the R-square value, Mean
Squared Error (MSE) and Mean Absolute Error (MAE). E.g.
model parameters are loan-to-value (LTV) ratio at time of loan
application (start of loan), a binary indicator for whether this
account has had a previous default, time on book in years, ratio
of valuation of property to average in that region (binned), type
of security, i.e. detached, semi-detached, terraced, flat or other,
age group of property and region.
LGD calculations for one year datasets: We calculate the total
loss by summing NPV of pending payments and future
payments. We calculate the total exposure by summing up on
balance and off balance. LGD is calculated by dividing total
losses by total exposure for each observation. For each
observation calculate the LGD as below
Collateral = market value* predicted haircut
Outstanding Balance = NPV of (Payment dues from the past
since default ) + NPV of (sum of future cash flows)
Repossession LGD = (Outstanding Balance at default -
Collateral)/ Total Exposure
Non Repossession LGD = Outstanding Balance/Total Exposure
Sum and average the LGD to calculate the LGD for portfolio.
Downturn LGD
In recognition that banks may be unable to derive their own
downturn LGD estimates, the Board of Governors of the Federal
Reserve System proposed the following generic formula for
deriving downturn LGD from long-term-average or through-the-
cycle LGD estimates:
LGDDownturn = .08 + .92 LGD
where LGD equals the long-term average LGD and LGDDownturn
equals expected downturn LGD.5 Under this rule, downturn LGD
can be anywhere from zero to eight percentage points higher
than the long-term average LGD. For example, a debt with a
long-term average LGD of 100% would have a corresponding
downturn LGD also equal to 100%, while an instrument with a
LGD of 0% would have a corresponding downturn LGD equal to
8%. For a debt with LGD equal to 50%, the corresponding
downturn LGD equals 54%.
very low in boom times. Mortgage industry is trillion
pounds industry; with capital requirement going high in
downturn would threat the economic stability. With
TTCs, main issue is that drivers to it, are very static in
nature and might not reflect the actual economic cycle.
So PRA in UK has proposed that organization should
neither use the PITs approach nor TTCs approach.
Rather organization should be using the hybrid
approach.
Now with hybrid approach, PRA has introduced the
concept of cyclicality. PRA defines cyclicality as measure
of PITness of the model. It would be very specific to
model. Additionally PRA has put a cap on cyclicality, at
no point model cyclicality should exceed more that 30%.
PRA has purposed two different formulas to calculate
the cyclicality.
Formula (1) cyclicality = (PDt – CT)/(DRt-CT)
Formula (2) cyclicality = (PD(t) – PD(t-1) )/ (DR(t)-
DR(t-1))
Both the formulas have their own issues while
calculating cyclicality. Formula (1) gives very high
number for cyclicality, when default rate gets close to
CT. Formula (2) gives very high number for cyclicality
when default rates of two consecutive year becomes
closer. To deal with these problems of cyclicality
formula, I purpose a new formula.
Formula (3) cyclicality = SD(ΔPD/ ΔDR)
Now question is why this formula is appropriate for
cyclicality calculation. When we are looking to capture
changes of series, standard deviation of it should
capture it. Next question is why we need model
cyclicality; CP2916 says that we should uplift the model’s
long run average PDs in proportion of cyclicality. So first
step is to calculate model cyclicality and then uplift the
model long run PD in proportion of model cyclicality. It
will complete the CP2916 calibration.
For more information on CP2916 implementation,
please contact me by email
levels using chi square statistics , proc means generates
level and proportions of events statistics
c. Remove redundancy: using proc varclus, it create
variable cluster based on second eignvalue, default is
1.00 you could specify 0.7
d. Pick variable based on R square ratios from cluster.
Rquare ratio is ratio of (1-Rquare within dataset)/ (1-
Rsquare value with other clusters)

e. Run the proc Reg with VIF option to detect the co-
linearity, remove the variable with VIF greater than 5.
f. Now run corr and create a table of speaman and
hoffding value, remove values with low rank of
spearsman and high rankof hoffding value, they indicate
nonlinearity, its called nonlinearity screening
g. Next step to adjust outlier using percentile treatment

3. Now for each numeric variable create intervals and


for each category variable with different levels calculate
WOEs and IVs. Each bucket/bin should contain 5% of
observations. The variables with IV less than 0.1 and
more than 0.5 should be kept of analysis; they are under
predictive or over predictive.

4. Now do the modelling (attrition -1, non attrition - 0)


using stepwise logistic regression (forward selection,
backward selection and stepwise selection) and subset
selection (where selection = score), Select the model,
which predicts best
and explanatory. Cutoff is calculated using ROC.

5. Validate the model using the validation dataset, in


case you have done the oversampling, correct the
sample bias by using prievent option in Logistic
Regression step. Use ROC, KS statistics and Gini
coefficient for judging the model performance. KS
should be nearly .5. For good behavioral score card, it
should be above .65
We now have two redictions
In credit scoring, one can analyse the historical data on past for odds of a good, one
borrowers and use it to predict the likely future behaviour based on age and the other
of prospective borrowers. To do this one assembles relevant based on residential status.
data on the past borrowers from application forms, credit
bureaus, accounting and financial records, as well as They are very different from
marketing information. one another. How do we
combine these different
There are three interrelated ways of describing the chance results to obtain a prediction
of a borrower being good. The first is the familiar one of which makes use of the
estimating the probability that an event occurs; the second information in both age and
is to consider the odds of its occurrence; and the third is to own/rent status. The
prescribe a score or index, which contains all the simplest approach is to use
information needed to estimate these odds.We want to Bayes’ rule in a very naïve
understand how probabilities, odds, and scores are related way :
to one another. When dealing with probabilities it is often
easier in practice to consider the odds of the event – the Naïve Bayes’ scorecard
chance of the event happening divided by the chance of it building
not happening. Horse racing is a good example of where
this occurs. Similarly with default risk one can assess the
chance of a good or bad outcome by the odds of a good or
bad, where we define the odds as the ratio of the
probability of a good (bad) outcome to a bad (good)

outcome:

Population odd : Opop = p (G) / p (B)

Information odd :

O (G | X) = p (X) . p (G | X) / p(X) . p (B | X)
= p (G | X) / p (B | X)

Using bayes theorem,

p (G & X) = p (G|X) . p (X) = p (X|G) . p (G)

O (G | X) = p (G | X) / p (B | X)