Evolving measures of credit risk
Εξελισσόμενα μέτρα πιστωτικού κινδύνου
View/
Keywords
Credit risk ; GPTIPS ; Genetic programmingAbstract
Credit scoring constitutes a quintessential element of economic risk management
allowing financial agencies to quantify the probability of default for a future loan. However,
acclaimed contemporary credit risk measures such as the scores provided by FICO or Vantage
are not publicly accessible.
The severity of the underlying problem is manifested by the limited amount of knowledge
which can be obtained for both the exact analytical formula and the complete set of credit-specific
features that underpin the computation of FICO score. The proposed measure will be derived by
exploiting a limited amount of entry-level information submitted by each candidate borrower
without requiring the accumulation of historical credit data for each consumer over large periods
of time.
This thesis addresses the problem of developing an alternative credit scoring measure
that approximates the behavior of the original FICO score in a large-scale collection of loan-
related data available from Lending Club. We are particularly interested in expressing the
acquired credit risk measure in a closed analytical form of adjustable complexity. For this purpose,
we utilize a symbolic regression technique which operates within the framework of Genetic
Programming (GP). In this context, we harness the notion of Occam’s razor to apply evolutionary
pressure towards the preservation of models associated with reduced complexity and higher
degree of human interpretability.
In order to verify the validity of our approach we compare the approximation ability of the
GP-based symbolic regression against state-of-the-art machine learning-based regression
methods such as Support Vector Machines (SVMs), Multi-Layer Perceptrons (MLPs) and Radial
Basis Function Networks (RBFNs). Our experimentation demonstrates that GP-based symbolic
regression achieves comparable accuracy with respect to the aforementioned benchmark
techniques. At the same time, the acquired analytical model can provide valuable insights
concerning the credit risk assessment mechanisms that underlie the computation of FICO based
on a significantly reduced set of credit-related features.