AI Credit Risk Analyst Interview Questions

A great answer distinguishes that a credit score is a rank-ordering heuristic (e.g., FICO 300-850) while a PD model outputs a calibrated probability (e.g., 3.2% default likelihood) and explains why calibrated probabilities are more useful for pricing and capital allocation.

Discuss false positives (approving bad borrowers → financial loss) versus false negatives (rejecting good borrowers → opportunity cost) and how the business chooses the optimal threshold by balancing expected loss against revenue forgone.

Explain the performance window (e.g., 12 months post-origination) and observation window, and how 'bad' is typically defined as 90+ days past due, charge-off, or default within that window.

Cover PD (probability of default), LGD (loss given default), and EAD (exposure at default), and note that expected loss = PD × LGD × EAD.

Discuss missing values in bureau data, class imbalance (defaults are rare events), data leakage from using future information, and survivorship bias in historical portfolios.

Cover data exploration, target variable definition, train/validation/test split respecting time, feature engineering, model selection (logistic regression baseline then GBMs), hyperparameter tuning, evaluation (Gini, KS), and documentation.

Explain that WoE transforms categorical/continuous features into a log-odds scale, ensures monotonic relationships with the target, and produces interpretable coefficients in logistic regression - all critical for regulatory acceptance.

Discuss strategies like class weighting, SMOTE/ADASYN oversampling, undersampling, focal loss, and - importantly - why adjusting the decision threshold or using stratified sampling for training is often preferred over synthetic data generation.

Define PSI as a measure of shift between the expected (training) and actual (current) score distributions, describe the threshold conventions (PSI < 0.1 stable, 0.1-0.25 investigate, > 0.25 significant shift), and explain that rising PSI may indicate model drift requiring recalibration.

Discuss that scorecards (logistic regression with WoE) are highly interpretable and regulator-friendly but may sacrifice predictive power, while GBMs/neural nets capture nonlinear interactions better but require additional explainability layers (SHAP) for compliance.

Cover temporal leakage (using post-origination data to predict default), look-ahead bias in features, target leakage through collection activity variables, and the importance of strict train/test splits by origination date.

Explain grouping loans by origination month/quarter, tracking cumulative default rates over time, comparing vintages to identify underwriting drift, and how macroeconomic conditions affect different cohorts differently.

Explain that discrimination (AUC/KS) measures rank-ordering ability while calibration ensures predicted probabilities match observed default rates - critical for pricing, capital reserves (Basel IRB), and expected loss calculations.

Discuss using GDP growth, unemployment rate, housing price indices, and interest rates as scenario variables; explain through-the-cycle vs. point-in-time calibration; and describe how the Fed's CCAR/DFAST scenarios are applied.

Define KS as the maximum difference between cumulative distribution functions of good and bad borrowers across score thresholds, explain that higher KS (e.g., > 0.4) indicates better separation, and note it is a standard metric alongside Gini in credit risk.

Discuss model serialization (ONNX, PMML), containerized serving (Docker + Kubernetes), feature store architecture for pre-computed features, A/B traffic routing for champion/challenger, and automated audit logging for every decision.

Describe the Foundation IRB vs. Advanced IRB distinction, how banks estimate PD internally while regulators may set LGD/EAD (F-IRB) or banks estimate all three (A-IRB), and how the risk-weight function converts PD, LGD, EAD into capital requirements.

Discuss testing at multiple stages (application, approval, pricing, limit assignment), intersectional fairness (race × gender), counterfactual fairness methods, adverse action reason mapping, and documentation standards for regulatory examinations.

Cover monitoring metrics (PSI, KS decay, default rate drift), automated alerting thresholds, data quality gates using Great Expectations, triggered retraining with the latest data, shadow scoring before promotion, and rollback mechanisms.

Explain SHAP-based feature importance rankings, SHAP interaction values to detect multicollinearity and nonlinear interactions, using SHAP dependence plots to verify monotonicity constraints expected by credit policy, and comparing SHAP profiles across demographic groups.

Discuss model documentation (SR 11-7 compliance), SHAP/LIME explanations, surrogate model validation (approximating the neural net with a logistic regression to show alignment), sensitivity analysis, benchmarking against interpretable alternatives, and independent validation unit (IVU) review.

Cover regulatory concerns (ECOA adverse action requirements, explaining non-traditional factors), proxy bias risk when no traditional credit history exists, information value decay for behavioral features, and the ethical implications of surveillance-based lending.

Discuss segmenting portfolio by product, geography, risk grade; modeling PD as a function of macro variables (logistic or Merton-style); estimating LGD conditional on collateral and economic conditions; aggregating expected and unexpected losses; and running Monte Carlo simulations for tail risk.

Cover offline store (batch-computed features from data warehouse) vs. online store (low-latency key-value lookup), point-in-time correctness to prevent leakage, feature versioning and lineage, and integration with tools like Feast, Tecton, or AWS SageMaker Feature Store.

Explain TTC ratings are stable across economic cycles (conservative for capital) while PIT ratings reflect current conditions (useful for pricing and provisioning); discuss migration matrices, rating philosophy, and how regulators view procyclicality.

Systematically check: (1) data pipeline issues (missing features, schema changes), (2) population shift (new applicant segments), (3) macroeconomic deterioration, (4) definition drift in the target variable, (5) upstream data provider changes, and (6) model score distribution comparison using PSI.

Discuss using bureau data as a proxy, transfer learning from similar product models, rule-based underwriting as a starting point, conservative initial cutoffs, rapid A/B testing to build labeled data quickly, and psychometric/applicant-level alternative data.

Describe conducting a formal fair lending analysis (disparate impact ratio, marginal effect analysis), improving adverse action reason specificity using SHAP-based feature attribution, ensuring reason codes map to actionable borrower-controllable factors, and documenting the remediation plan.

Cover data quality assessment, feature engineering from transaction streams (cash flow volatility, income stability, expense patterns), information value analysis, marginal model improvement testing, regulatory review of new data usage, and phased rollout with champion/challenger.

Discuss tightening credit cutoffs for marginal segments, reducing exposure limits, increasing pricing for higher-risk cohorts, accelerating stress testing scenario updates, communicating revised loss forecasts to the CFO/board, and monitoring early delinquency trends daily.

Address model comparability (regulatory requires parallel run), feature mapping differences, team skill gaps, operational integration (SAS batch vs. Python API), documentation of new model for SR 11-7 compliance, and maintaining the old model as a fallback during the transition period.

Discuss pre-computed features in an online feature store, lightweight model inference (ONNX-optimized gradient-boosted model), soft-pull bureau data, real-time fraud signals, and a decision engine that combines ML scores with policy rules for instant approve/decline/refer decisions.

Discuss investigating root cause (thin credit files, different feature distributions), evaluating whether a separate sub-model or age-specific calibration helps, ensuring any age-based treatment doesn't violate ECOA, considering alternative data for young applicants, and monitoring for proxy discrimination.

Describe examining feature importance rankings and SHAP values for correlated proxies, running residual analysis regressing model scores against protected attributes, conducting disparate impact tests across geographies, and establishing a prohibited variable policy with regular compliance audits.

Discuss mobile money transaction data, psychometric testing (e.g., Lenddo, First Access approaches), social network analysis, airtime usage patterns, device metadata, ensemble approaches combining multiple weak signals, and the importance of local domain experts for feature validation.

Describe a pipeline that takes model score, SHAP feature contributions, borrower financials, and policy rules as structured inputs, constructs a prompt template with context, uses an OpenAI or Anthropic LLM via LangChain to generate a narrative memo, and implements output parsing for structured sections (summary, risk factors, recommendation).

Explain using LayoutLM or Donut for document understanding, OCR preprocessing with Tesseract, fine-tuning on labeled financial documents, extracting key fields (income, expenses, account balances), and integrating the extraction pipeline into the credit feature engineering workflow.

Cover logging model parameters (hyperparameters, feature sets), metrics (AUC, KS, Gini per subgroup), artifacts (serialized model, SHAP plots), model registry stages (staging → production → archived), and integration with CI/CD pipelines for automated model promotion.

Describe defining functions for querying borrower data, running model predictions, retrieving policy rules, and generating adverse action reasons; building a conversation flow where underwriters ask natural language questions and receive structured responses with evidence; and implementing guardrails to prevent hallucinated risk assessments.

Discuss staging raw loan and bureau data with dbt source definitions, building intermediate models for feature computation (rolling averages, behavioral aggregates), creating mart-layer tables for model training and monitoring dashboards, and using dbt tests for data quality (not_null, accepted_range, unique).

Explain computing SHAP values for each prediction, ranking the top contributing features, mapping raw feature names to borrower-friendly reason codes (e.g., 'high_utilization' → 'Your credit card balances are high relative to limits'), and filtering for controllable borrower factors only.

Cover DAG structure: extract yesterday's scored applications, compute PSI and KS against baseline, check data completeness with Great Expectations, compare default rates by segment, generate alerts via Slack/email if thresholds are breached, and push metrics to a monitoring dashboard.

Describe creating a SageMaker endpoint with production variants for champion (80% traffic) and challenger (20% traffic) models, configuring auto-scaling policies based on invocation metrics, using SageMaker Model Monitor for drift detection, and logging all inference requests for audit.

Explain integrating fairness metrics (disparate impact ratio, equal opportunity difference, predictive parity) into the CI/CD pipeline, using tools like Aequitas or Fairlearn, setting automated pass/fail gates based on thresholds, and generating a fairness report artifact that model risk management reviews.

Describe using a structured template with sections (model purpose, methodology, data, performance, limitations), feeding model artifacts and test results as context to an LLM via LangChain, generating draft documentation, and implementing a human-in-the-loop review workflow where analysts verify and approve generated content.

A strong answer demonstrates courage to escalate, explains the specific risk clearly to non-technical stakeholders, offers an alternative solution that meets both business and compliance goals, and shows the outcome of the decision.

Look for honest self-reflection, identification of the root cause (data drift, leakage, overfitting), concrete remediation steps, and a systemic change they implemented to prevent recurrence (e.g., stricter validation protocols, monitoring improvements).

Strong answers show the ability to simplify without losing accuracy, use visual aids (lift curves, segment comparisons), frame risks in business terms (expected loss in dollars, revenue impact), and tailor depth of technical detail to the audience.

Look for structured decision-making under uncertainty, awareness of the tradeoff between speed and accuracy, use of heuristics or rule-based fallbacks when ML isn't available, and a post-mortem to improve the process.

A great answer includes specific sources (papers, conferences like NeurIPS Credit Risk Workshop, regulatory publications, industry newsletters), hands-on experimentation with new tools, and evidence of applying new knowledge to their work.