A PD Model Has Deteriorated. Should You Recalibrate or Rebuild?

Recalibration and redevelopment repair different things

Recalibration preserves the model’s underlying score or ranking relationship while revising the mapping from that output to probability of default. It is plausible when ranking remains sufficiently useful, the population is relevant, structural relationships are stable enough, and the observed calibration shift can be estimated from representative outcomes.

Redevelopment revisits the model’s predictive architecture: variables, transformations, segmentation, functional form, interactions, estimation sample or modelling method. It becomes more plausible when relative ordering weakens, relationships no longer generalise, important populations sit outside the model’s support, or repeated adjustments cannot restore stable performance.

This is a continuum, not a clean binary. A concentrated segment may need a new submodel while the wider portfolio needs only recalibration. A decision-policy correction may improve outcomes without changing the model. A data defect must be repaired before either intervention is assessed. PD model monitoring detects and decomposes these signals; intervention begins only after that diagnostic work.

Four states organise the first decision

Discrimination and calibration answer different questions. Discrimination asks whether higher-risk borrowers still rank above lower-risk borrowers. Calibration asks whether predicted PD levels correspond to realised risk. Their combination provides a compact starting map.

State 3 is easily missed. Aggregate predicted and observed default rates can align even while risk ordering deteriorates inside the portfolio. Offsetting errors may make the calibration total look acceptable. A level correction cannot restore lost separation, and acceptable O/E does not prove that account-level probabilities remain decision-useful. The distinction is developed further in PD ranking and calibration research.

Two similar symptoms, two different interventions

Consider a fictional consumer-credit model evaluated on mature, comparable twelve-month outcome windows. The figures are synthetic and demonstrate reasoning rather than universal action thresholds.

In Scenario A, AUC changes only modestly while observed risk rises far above predicted risk. The ranking relationship appears broadly intact; absolute risk has shifted. If the movement is persistent, outcomes are representative and calibration error is not hiding segment failure, recalibration is a credible candidate. It is not yet the decision.

Scenario B has almost the same aggregate calibration error, but AUC falls materially. A recalibration might improve portfolio-level O/E while preserving a weakened ordering underneath. The investigation now belongs at relationship, variable and segment level: rank ordering by band, stability of coefficients or feature effects, population support, missingness and policy selection. Redevelopment becomes materially more plausible.

Vintage evidence also matters. If deterioration is concentrated in recent originations after a cut-off or channel change, same-age vintage analysis can separate a new-book effect from broad model failure. Immature cohorts or temporary macro pressure can make an immediate intervention premature.

Recalibrate or rebuild? Follow the cause, not the alarm

If deterioration is segment-specific, a portfolio-wide change may dilute the real problem. Investigate eligibility, overrides, feature coverage and a segment treatment before rebuilding everything. If a new cut-off, pricing strategy or acquisition campaign changed the booked population, assess the strategy and selection effect before blaming model methodology. If upstream data changed, restore implementation integrity before estimating any new parameters.

Governance escalation is not reserved for redevelopment. A material recalibration changes risk estimates and may affect approvals, pricing, provisions or capital. The required validation, approval and change classification depend on model use, materiality and the applicable regime. No universal statistical boundary can make that decision safely.

Recalibration is not a repair button

It becomes one when speed is mistaken for diagnosis. Recalibrating before analysing population movement can fit the model to a changed mix it was never intended to represent. Recalibrating on too few mature outcomes can turn noise into a production parameter. Correcting only the portfolio average can leave serious underprediction in one segment offset by overprediction in another.

Repeated recalibration is a particularly valuable warning. If each adjustment loses relevance quickly, the baseline may not merely be moving; the relationships or operating environment may be unstable. Continuing to adjust the output scale can accumulate methodological patches while postponing the structural decision.

Nor should temporary macro movement be treated mechanically. The model’s rating philosophy, observation horizon and intended use determine how economic conditions should appear in estimates. Diagnosis should test whether the deviation is persistent, systematic and decision-material, not simply whether one period is adverse.

Choose the smallest intervention that repairs the actual failure

The model did not deteriorate in the abstract. Its data, population relevance, ranking, probability mapping, segment behaviour or operating environment deteriorated in a particular way. The defensible response is the smallest governed intervention that addresses that cause without concealing residual risk.

The recurring sequence is monitor → detect → diagnose → classify → prepare investigation → human decision. Calculation, historical comparison, exception prioritisation and evidence assembly can be systematically augmented. Model-change decisions, materiality judgements and accountability remain human-governed.

Methodological boundary

The framework is original Entimema reasoning grounded in general model-risk principles. The 2026 US interagency model-risk guidance explicitly positions adjustment, recalibration and redevelopment as outcomes informed by validation; the EBA’s PD calibration guidance requires analysis of why realised and estimated default rates deviate; and the Basel Framework requires ongoing review of performance, stability, model relationships and outcomes. Institution-specific standards remain governed by portfolio, use and regulatory context.