A2007
Title: Geolocation-aware credit risk modeling
Authors: Margot Geerts - KU Leuven (Belgium) [presenting]
Jochen De Weerdt - KU Leuven (Belgium)
Seppe vanden Broucke - U Gent (Belgium)
Abstract: Although financial and nonfinancial data are considered key to corporate credit risk modeling, network data and supply chain data have been recently adopted as well. However, more recent research on corporate credit risk shows that credit ratings also depend on location. Corporate credit risk is approached as a geospatial problem. The geolocation of firms' headquarters allows for assessing the location effect of a company's credit risk at a finer level than in previous research. Tree-based methods consistently rank among the best-performing models for tabular data, including this application. Yet, currently available decision tree learning algorithms are suboptimal for geospatial problems. First, conventional decision tree learners are restricted to axis-parallel boundaries. For data sets including $X-Y$-coordinates, this leads to unnatural decision boundaries. Second, decision tree learners are insufficiently tailored to operate well on heterogeneous data. This creates a strong need for a tailored geospatial decision tree learning algorithm with more appropriate splits. Two multivariate decision tree splitters are proposed: diagonal splits and Gaussian splits. The former includes linear combinations of features in the set of candidate splits, and the latter approximates the decision boundary by a Gaussian and splits around it. As this introduces intractability in finding the optimal split, heuristic optimization is leveraged to achieve higher performance and scalability.
