Three sessions.
One vertical.

Settled: FICO-based credit scoring models trained on 2010–2019 data have demonstrated performance degradation in 2020–2024 across multiple published validation studies (NY Fed, OFR). Contested: the magnitude of degradation — studies range from marginal AUC reduction to 15+ point Gini decline. Unknown: whether performance degradation reflects macroeconomic regime change (ZIRP to high-rate environment) or pandemic-specific data anomalies that will self-correct.

Temporal sampling problem: the model was trained in a prolonged low-rate, low-default environment. Consumer behaviour in high-rate environments (debt service ratio management, strategic default decisions) is governed by different factors than low-rate behaviour. A model trained on 2010–2019 data has never seen a consumer population making decisions under 7%+ mortgage rates and 22%+ revolving credit rates simultaneously.

“The validation studies you're citing measure model performance on the existing approved portfolio. They cannot measure performance on the counterfactual declined population. If the model has been systematically declining creditworthy borrowers in the new rate environment — which a regime change argument implies — your actual loss rates look better than the model deserves credit for. The model may be generating phantom accuracy by denying credit at scale.”

Confidence: Contested on model validity. The phantom accuracy objection is methodologically serious and unresolved. Recommendation: commission a reject inference study using a random sample of declined applications approved for monitoring purposes before next regulatory model validation.

Settled: forward curves systematically overestimate the persistence of rate levels at the peak of hiking cycles — documented across 1994, 2000, 2006, and 2019 cycles. Contested: whether the current cycle (2022–present) will follow the same pattern given fiscal dominance dynamics not present in prior cycles. Unknown: the degree to which AI productivity gains will affect neutral rate estimates and therefore the magnitude of the eventual pivot.

Forecasting accuracy calibration: across the four documented peak cycles, 18-month forward curves overestimated the terminal rate by an average of 180 basis points. Applied to current pricing: if the pattern holds, the market is pricing in 180bps too much persistence. The Contrarian should be asked specifically whether fiscal dominance changes this calibration.

“The 180bps historical overshoot is a useful prior but the fiscal dominance argument is material. In all four prior cycles, fiscal deficits were smaller at cycle peak (<3% GDP). Current deficit is 6.5% GDP. The academic literature on fiscal dominance suggests this level of deficit monetisation reduces the Fed's effective independence on the way down — pivots happen later and are shallower. The historical calibration should be discounted by at least 50%.”

Calibration: historical prior suggests 18-month forward curve overstates by ~180bps. Fiscal dominance discount suggested by Contrarian: 50%. Probability-weighted adjustment: ~90bps. Current positioning implication: duration extension is warranted but less aggressive than pure historical base rates suggest. Present as a range (60–120bps adjustment) given model uncertainty.

Settled: SEC and FINRA have both issued exam priorities letters (2024, 2025) explicitly flagging AI-assisted surveillance as an examination focus. Contested: whether the exam focus will result in formal findings or serve primarily as an industry signal. Unknown: what specific documentation and model governance standards examiners will apply — no formal rulemaking has been completed.

Regulatory signal analysis: in the two prior cases where SEC/FINRA issued joint exam priority letters for specific technology categories (algorithmic trading, 2019; digital assets, 2021), formal findings followed in 60–70% of examinations within 18 months of the priority letter publication. Current programme has three documentation gaps relative to the guidance in the 2024 exam priority letter.

“The 60–70% finding rate from prior tech exam cycles is the right base rate, but it applies to the entire examination population. Our firm has a clean 5-year exam history and a documented model governance framework. Firms with clean histories and documented governance have been found at approximately 20–30% rates in analogous cycles. The base rate needs to be conditioned on our firm-specific profile.”

Three pre-exam actions: (1) Close the three documentation gaps identified in the 2024 priority letter within 60 days — these are the lowest-hanging fruit and the most likely exam focus. (2) Commission a mock examination specifically on AI surveillance governance. (3) Brief the Board Audit Committee on the exam probability before the next scheduled exam notification.