HU Credits:
1
Degree/Cycle:
2nd degree (Master)
Responsible Department:
Business Administration
Semester:
2nd Semester
Teaching Languages:
English
Campus:
Mt. Scopus
Course/Module Coordinator:
Prof ROGER STEIN
Coordinator Office Hours:
Teaching Staff:
Prof ROGER STEIN
Course/Module description:
This short seminar focuses on the practical challenges that arise in implementing a variety of datadriven
models for discrete choice problems in finance, as well as several frameworks for thinking developing FinTech
platforms that use these building blocks.
Datadriven methods for predicting discrete choice are now in wide use in both traditional banks and new financial firms platform
companies. However, machine learning and statistical algorithms are only a small part of what is involved in building
robust analytics. And analytics, and technology more generally, form only one of the buildingblocks for successful FinTech
businesses.
Although the approaches we describe are applicable to a widevariety of financial and insurance problems, throughout this
short seminar, we will use default (e.g., bankruptcy and default models retail and commercial entities) as a prediction as a
working example throughout this short seminar. With a focus on large data sets, we explore a number of datadriven approaches
to modeling binary outcomes. I will draw heavily on my experiences building and evaluating some of the most
widely used and commercially successful datadriven credit evaluation tools in the industry. This seminar will tend heavily
towards discussions of practical model implementations and the “frictions” that make these implementations difficult in realworld
settings. We pay special attention to validating discretechoice models in realworld settings.
We will take the view that an effective, practical modeling framework will sometimes be rough around the edges with the odd
inconsistency (usually to deal with available data or the lack thereof). This implies that seemingly incompatible models can
each have value in specific contexts, resulting in retention of several models despite their theoretical inconsistency. Because
the focus is applied, we will discuss model validation and calibration in detail and highlight data issues in estimation and validation.
Lectures will to focus on conceptual themes and practical issues, with much of the technical detail underlying these to
be found in the readings.
I will also provide suggested “miniprojects” for those students who are more technically
inclined. These projects serve to provide motivation and, if you do them,
you will leave the seminar with some very useful tools for applying this subject
matter in practice. These “miniprojects” are described using R syntax, though they
may be implemented in any language in which you work (Python, SAS, Matlab,
etc.). I will go over a “solution” to at least one of these during the seminar.
Course/Module aims:
To expose students to the practical challenges associated with building and testing datadriven discrete
choice models introducing several of the modeling techniques that can be used to build them and to provide a framework for
building robust fintech platforms that use these tools.
Learning outcomes  On successful completion of this module, students should be able to:
be familiar with modeling techniques that can be used to build them and to provide a framework for
building robust fintech platforms that use these tools.
Attendance requirements(%):
Teaching arrangement and method of instruction:
Course/Module Content:
Day 1
Thursday
March 26,
2020
Introduction to discrete choice models, credit risk modeling concepts, the challenge of data analytics and the
nature of FinTech platforms
• How can we add value in developing datadriven analytics
• Data problems and resolutions
• Key components of credit risk – PD, LGD, (EAD), correlation, size
ACPMIP: Chapter 1, pp. 216; 1923; 3234; 38; 4243. Chapter 2, pp. 6062; 7274.
Supplemental readings:
• Dhar, V. and
Day 2
Friday
March 27,
2020
Introduction to PD model validation
• The role of trust in FinTech
• Validating model power using ROC curves
• Validating model calibration using probabilitybased measures
ACPMIP: Chapter 7, pp. 361397.
• Dhar, Vasant. (2016) “When to Trust Robots with Decisions, and When Not To”, Harvard Business Review. May.
• Stein, R. M., A. E. Kocagil, J. Bohn and J. Akhavain (2003). “Systematic and Idiosyncratic Risk in MiddleMarket Default
Prediction: A Study of the Performance of the RiskCalc and PFM Models.” Moody’s KMV.
For those interested, see if you can write this tool: Function to calculate the AUC ROC for two different subsets of a single data set.
• Definition: subROC<function(x, split.val, split.on, score, outcome,...) where
o x is a dataframe
o split.val is a scalar, factor value, date or string used to divide the data frame
o split.on is a vector of length nrow(x) of the same type as split.val; split.on <&eq; split.val goes to one data subset
while the remainder goes to the other
o score is a numerical vector of length nrow(x) for calculating the ROC AUC
o outcome is a binary numerical vector of length nrow(x) for calculating the ROC AUC
o ... additional parameters
• Return: The function should return a list with three elements:
o A vector of length 2 with the ROC AUC for each data subset.
o A vector of length nrow(x.subset1) giving the indices of x that for the records included in x.subset1
o A vector of length nrow(x.subset2) giving the indices of x that for the records included in x.subset2
• Tasks:
• Implement subROC
• Describe how you would make subROC more general so that it could take in an arbitrary one or two variable statistic (function) as an
input and return the appropriate data
Day 3
Sunday
March 29,
2020
Data driven models for discrete choice problems
• Discrete choice models
• Survival models
Treebased models
• CART
• RandomForests
ACPMIP: Chapter 4, pp. 183215, 238252.
Supplemental readings:
• Dhar, V. and Stein, R. (1997), Seven Methods for Transforming Corporate Data into Business Intelligence, Ch. 10.
• Friedman, Hastie and Tibshirani (2013), Elements of Statistical Learning, Section 9.2, pp. 305311.
Day 4
Thursday
April 2,
2020
Introduction to discrete choice model calibration
• Calibrating to empirical data using calibration curves
• Adjusting for differing baseline rates
• Mapping between ordinal scales and PDs and back again
• The features of successful of FinTech platforms
ACPMIP: Chapter 4, pp. 215233.
Supplemental readings:
• Dhar, V. and Stein R. M. (2017) “Economic and Business Dimensions on FinTech Platforms and Strategy.”, Communications
of the ACM, 60, 10, October, pp. 3235.
• Stein, R. M., A. E. Kocagil, J. Bohn and J. Akhavain (2003). “Systematic and Idiosyncratic Risk in MiddleMarket Default
Prediction: A Study of the Performance of the RiskCalc and PFM Models.” Moody’s KMV.
For those interested, see if you can write this tool:
• Function to create a calibration curve mapping a variable to a default rate.
• Function to use the calibration curve to map a variable to a default rate (including interpolation).
• Definition: estimateCalibCurve<function(x, outcome, k, ...) where
o x is a vector of model scores
o outcome is a is a binary numerical vector of length length(x)
k is a scalar, denoting the number of “buckets” to use in the mapping
• Return: The function should return a list with three elements:
o A list containing
§ map a dataframe of length k with two columns
• the cutoff (on the same scale as x)
• the mapped PD corresponding to the cutoff
§ baseline a scalar containing the baseline PD for outcome
• Definition: applyCalibCurve <function(x, map, baseline&eq;NULL, ...) where
o x is a vector of model scores
o map is a is a dataframe returned in map by buildCalibCurve
baseline is a scalar, to be used if baseline adjustment is to be applied after mapping
• Return: The function should return a vector of length length(x) containing the mapped PD for each element of x .
• Tasks:
• Implement estimateCalibCurve
• Implement applyCalibCurve
Required Reading:
Bohn, J. R. and R. M. Stein, (2009) Active Credit Portfolio Management in Practice, NY, Wiley. (ACPMIP).
Additional Reading Material:
12 papers per course session may be recommended (see outline).
Course/Module evaluation:
End of year written/oral examination 0 %
Presentation 0 %
Participation in Tutorials 100 %
Project work 0 %
Assignments 0 %
Reports 0 %
Research project 0 %
Quizzes 0 %
Other 0 %
Additional information:
This seminar is a highly compressed version of a fullsemester course I give. Much of that course focuses on implementing
the ideas through handson data science projects with large data sets. While I will allude to this work from
timetotime, much of the technical detail will be omitted. No programming is required for this seminar.
