12-month PD, lifetime PD and a macroeconomic model: redevelopment post-validation
A European bank asked us to redevelop the Stage 1 and Stage 2 PD models behind its IFRS 9 ECL, and the macroeconomic model driving them.
Defaults are scarce in a performing book and history is short, yet the rebuild had to land before year end.
What We Provided
A model the committee can follow, chosen rather than assumed
Random forest, XGBoost and neural network candidates were all trained on the bank's data as genuine options. Each overfitted and each was harder to explain, so logistic regression was retained on the evidence.
Features that rank on risk, not on the passage of time
Judgement dropped features that drift with prices, amortisation or seasoning, then information value, K-fold cross-validated incremental discrimination, a forward stepwise pass and variance inflation factors decided what stayed.
Transparency that costs nothing in performance
Every internal feature was encoded as Weight of Evidence, binned under monotonic constraints and fitted with logistic regression. A challenger built this way beat a proxy of the previous specification at every out-of-time snapshot.
A forward-looking view that does not depend on short history
The forward-looking component was built from scratch and kept outside the scorecard, on independently sourced write-off rates and macroeconomic projections, so the economics can be refreshed without reopening the scorecard.
Stage 2 provisions that reflect when defaults actually happen
A single flat PD across every Stage 2 account was replaced with a lifetime PD curve fitted to observed default timing, with marginal PDs from its first differences.
The Results
The asset and liability committee approved the new models, implementation into the bank's own code went through successfully, and full documentation went with it. The bank now runs its ECL engine on the new models itself.
Other Case Studies
Independent validation of a Pillar 2A concentration add-on
A European bank needed to know that the capital it holds against concentrated lending would stand up to scrutiny, so it asked us for an independent view of the model behind the number.
Redeveloping every IFRS 9 model across two portfolios
A UK lender needed every model in its IFRS 9 expected credit loss suite rebuilt, at the point when the team that had built them was no longer there.