You are asked to validate a bank's retail PD scorecard. Its Gini has fallen, its population stability index is high and observed defaults run well above predicted. Interpret each result and decide whether to recalibrate, redevelop or keep it with an overlay.

Varenth's scorecard still ranks borrowers, but the population has moved and the level of risk is badly understated. Gini of 48% is weaker but usable; a PSI of 0.28 means today's applicants differ materially from the development sample; and observed defaults are 1.64 times predicted. Recalibrate now, as an overlay, so provisions and capital reflect 4.1%, and schedule a redevelopment, because the shift in population will keep eroding the ranking.

Think of a bathroom scale. It can put three people in the right order of weight and still read five kilos light for all of them; or it can read correctly for adults and badly for children, because it was never built for them. A PD model is validated on three separate questions: does it rank borrowers, is it being used on the population it was built for, and is its level of risk right. These are discrimination, stability and calibration, and each can fail while the others pass.

Discrimination: a Gini of 48% still ranks borrowers usefully, but a 23% relative fall in four years says the model is ageing. Stability: the population stability indexA measure of how far the score distribution of current applicants has moved from the development sample; larger values mean a bigger shift. of 0.28 comes mostly from applicants crowding into the lowest score band, 37% of today's book against 20% at development. Calibration is the sharpest failure: on 50,000 accounts the model expected 1,250 defaults and 2,050 arrived, about 23 standard deviations away, far beyond chance.

Match the fix to the failure. Recalibration fixes the level without touching the ranking, which is exactly what the calibration failure needs, and it can be done in weeks. Scaling predicted PDs so the portfolio averages 4.1%, about 1.64 times, brings provisions and capital into line now. But recalibration cannot fix a population shift. Applicants the model was not built on are a growing share of the book, and the falling Gini is what that looks like as it happens.

So the recommendation has three parts: an immediate recalibration, documented as a model overlay with an owner and an expiry date; a redevelopment on recent data that includes the new applicant segments; and, until then, a review of the approval cut-off, since a cut-off set on understated PDs is approving loans the bank would otherwise decline. The limitation: one year of observed defaults may partly reflect a weak economy rather than the model, so check whether defaults rose across other models too before blaming this one alone.