New SNSF Grant for Prof. Thomas Weber

© 2018 EPFL
The Swiss National Science foundation awarded a 3-year grant (from 2018 to 2021) to Prof. Weber in support of his research project on “Optimal Credit Collections: Theory and Practice.” In Europe, the United States, and other developed countries, significant amounts of unsecured consumer debt, worth many billions of dollars, are in default, requiring special collection efforts to retrieve an acceptable fraction of the outstanding loans. For the lending institutions, the performance of debt collections determines the value of their nonperforming loan portfolio and ultimately their capital-reserve requirements. The project seeks to develop techniques to practically identify repayment processes and then optimize the collections process using stochastic dynamic programming, interfacing with field data.
Aims of the Project: While in practice the approach to credit collections has become increasingly data-driven, until recently no effective stochastic models of credit repayments were available to accurately predict the behavior and value of defaulted accounts, as well as their reactions to a variety of account-treatment actions. In the first phase of the project, we examine the identification of both self-exciting point processes and hidden Markov processes, which enables predictions about state trajectory, e.g., in terms of attaining certain action/inaction regions in the (intensity,balance)-space relevant for credit collections. The corresponding probability metric can be used for valuation purposes and for decisions about interventions such as settlement offers. In the second phase of the project, we tackle the optimal control of such point processes, which despite the recent progress remains—from a practical point of view—an important open problem. The reason is, for example, that in practice the number of implementable interventions is usually limited, thus resulting in a nonconvex action region. In the third phase, we investigate, based on earlier work in the static context, how to dissolve the boundaries between the identification problem and the control problem using a robust optimization approach.
Scientific and societal context: The results have broad implications for collections management. First, given an identified model, a collections manager can use the optimal policy to determine the nature and timing of value-maximizing account-treatment actions. Second, the optimized (dynamic) net present value of an account implies at any point in time a threshold for early account settlement at less than 100% of the outstanding balance. Third, using the expected value of treated accounts in defaults, the bank can provide more precise estimates for the expected loss given default (LGD), which in turn enhances the bank's compliance with capital-reserve requirements as introduced in the Basel II accords. Moreover, the dependence of the minimum expected LGD on account attributes (and holder characteristics) may be used at the underwriting to curb excess exposure to risky clients and enhance the quality of the bank's loan portfolio. Beyond credit collections, the identification and subsequent control of marked point processes with observable or partially observable state arises naturally whenever there are monotone random progressions, such as—for instance—knowledge accumulation or growth of stock-pollutants in the atmosphere.
Chehrazi, N., Weber, T.A. (2015) “Dynamic Valuation of Delinquent Credit-Card Accounts,” Management Science, Vol. 61, No. 12, pp. 3077—3096. [DOI 10.1287/mnsc.2015.2203] [Download]