Banking

Proving the limits of a forward-looking LGD model

A UK bank asked us to review and enhance the forward-looking component of the LGD model behind its credit cards ECL.

The Challenge

Too few internal observations exist to forecast recoveries directly, so the macroeconomic link must run through an industry proxy.

What We Provided

An estimable base, rather than a forecast built on too little

Internal price observations were too sparse to carry a forecast, so the link was fitted to a published industry loss measure aligned to the definition of default, which a reviewer can check independently.

A search wide enough that the answer is not an accident

We swept the whole two-variable macroeconomic candidate space, screening on variable category, driver correlation and scenario ordering before ranking on fit, so the conclusion rests on the space rather than on the first specification tried.

A finding that saved the bank a model it could not have defended

No two-variable specification survived the screen. The apparent fits were overfitted, with noisy projections, so we recorded that the data could not support one rather than shipping the best-scoring candidate.

A simpler model, and the evidence a validator needs

Unemployment was refitted on refreshed data with a sensitivity parameter for severe scenarios, the methodology simplified, and controls and written evidence assigned, so the model reaches validation with its case already made.

The Results

The bank knows that on this data no two-variable specification predicts industry loss, and that the fits which appear to do so are overfitted. It holds the simpler model, refitted with a stress sensitivity, and the rejected alternatives documented.

Receive updates directly in your inbox

Stay connected