SSA Vic Branch May Meeting
On the 29th of May, we welcomed Ed Stokes as our guest speaker at our monthly meeting. Ed, previously a senior analyst at NAB and now a Manager in consulting at PwC, provided us with some useful insights into how he and his team build credit scoring models. Credit scoring typically refers to a predictive analysis of an individual’s ability, or lack thereof, to meet their obligations of a loan. Put simply, credit providers use credit scoring models to assess someone’s creditworthiness and use that assessment to make a decision on whether to provide a loan. The art of credit scoring has matured over the last two decades where the fitting of generalised linear models (GLMs) has been favoured. However, the GLM approach requires a great deal of data cleansing, stratification and feature engineering prior to model fitting and this can take up to six months. Added to this the need for regulatory sign-off, it may be up to 12 months before a model can be implemented in practice. A lot can change in that time and the trained model may no longer be relevant. This leads us to the use of gradient boosted trees (GBTs) which are popular in machine learning. GBTs use decision trees in series where a model is trained on the residuals from the previous fitted models. The GBT approach is much quicker that the GLM method and, as Ed pointed out, often results in credit scoring models with greater predictive strength. And this leads us to the title of the talk, “Credit scoring: improved predictions or improved interpretability?”. Unlike GLMs where model coefficients can be used to interpret the contributions of covariates to the model, GBTs are difficult to interpret due to their reliance on many decision trees. The downside of this is that regulators may be unwilling to sign-off on the use of a GBT model and this sign-off is necessary to protect against claims of discrimination. In summary, Ed delivered a very interesting, informative and energetic talk that enlightened us with an understanding of the challenges faced in the world of credit scoring.