hard · FRM Part 2 Current Issues
A bank's model validation team is performing 'effective challenge' under SR 11-7 on a deep neural network used for unsecured-loan underwriting. To explain individual predictions, the team computes SHAP (SHapley Additive exPlanations) values, which are constructed so that the attributions for a given loan application always sum exactly to the difference between the model's predicted score and its baseline (average) output. The team notices that two inputs, 'years at current job' and 'years at current address,' are highly correlated (ρ ≈ 0.85) across the applicant population. Across a sample of otherwise-similar declined applications, the SHAP attribution splits the combined negative effect of short job/address tenure very differently from applicant to applicant — sometimes assigning nearly all of the penalty to 'years at current job,' sometimes to 'years at current address' — even though the model's actual predicted score barely changes between these applicants.
What is the most defensible interpretation of this observation for purposes of model risk validation and effective challenge?
- The instability means the neural network is fundamentally unreliable, and the model must be replaced with a simpler algorithm regardless of its predictive performance, since stable attribution is a prerequisite for any AI model used in credit decisions.
- Because SHAP values sum to the model's output by construction, splitting credit between correlated features is a known artifact, not an error; validators should treat single-feature attributions as unreliable and instead assess correlated features jointly.
- The team should drop one of the two correlated features from the model so SHAP assigns all of the shared effect to the one remaining feature, which the team argues restores stable, interpretable attributions suitable for effective challenge.
- Because SHAP is a theoretically rigorous, model-agnostic method, any instability in its attributions must originate from a bug in the software implementation, so the team should escalate this to the vendor as a technology defect rather than a modeling limitation.
Sign up free to see the explanation and track your rank →
More FRM Part 2 Current Issues practice
- According to the BCBS standard for cryptoassets, a bank's to… — What is this limit?
- What primary balance (as a % of GDP) is required to stabilize the debt-to-GDP ratio?
- In monitoring model stability, which metric is commonly used to quantify the shift in dist
- SVB's management removed interest rate hedges in 2022. According to the Economic Value of
- What does a SHAP value of zero for a specific feature in a credit scoring model imply?
- What is the bank's primary ongoing obligation to regulators regarding this new model?
- A bank provides a 'NAV facility' to a private equity fund. If the fund's underlying assets
- If the underlying loans are quarterly-marked and illiquid, what is the most significant ri