n a world where big data is everywhere, no one has big data relative to the economic cycle. Data volume needs to be thought of along two dimensions. (1) How many accounts / transactions / data fields do we have? (2) How much time history do we have? Few, if any, big data sets include history covering one economic cycle (back to 2005) or two economic cycles (back to 1998). Therefore, unstructured learning algorithms will be unable to distinguish between long-term macroeconomic drivers and point-in-time variations across accounts or transactions. This is the colinearity problem that is well known in consumer lending.

When the underlying system or process that is being observed is based upon observations versus age, vintage (origination time) and calendar time, Cox proportional hazards models can exhibit instabilities because of embedded assumptions.

Current expected credit loss procyclicality: it depends on the model

The new accounting standards of CECL for the US and IFRS 9 elsewhere require predictions of lifetime losses for loans. The use of roll rates, state transition and “vintage” models has been proposed and indeed are used by practitioners. The first two methods are relatively more accurate for predictions of up to one year, because they include lagged delinquency as a predictor, whereas “vintage” models are more accurate for predictions for longer peri- ods, but not short periods because they omit delinquency as a predictor variable. In this paper we propose the use of survival models that include lagged delinquency as a covariate and show, using a large sample of 30 year mortgages, that the proposed method is more accurate than any of the other three methods for both short-term and long-term predictions of the probability of delinquency. We experiment extensively to find the appropriate lagging structure for the delinquency term. The results provide a new method to make lifetime loss predictions, as required by CECL and IFRS 9 Stage 2.