AI Default Prediction Specialist Interview Questions

A strong answer defines PD as the likelihood a borrower defaults within a time horizon and LGD as the percentage of exposure lost given default occurs, and notes their distinct modeling approaches.

Discuss class imbalance in default data and the need for AUC-ROC, precision-recall, KS statistic, and business-driven threshold selection.

Describe how WoE transforms categorical/continuous features into log-odds values that linearize relationships and enable scorecard point assignments.

Explain that temporal splits mimic real-world deployment (predicting future defaults from past data) and prevent data leakage from autocorrelation in credit cycles.

Examples include bank-transaction cash-flow data, utility/rent payment history, and digital footprint or device-level behavioral signals.

Explain PSI as a measure of distribution shift between training and scoring populations, with thresholds (typically <0.1 stable, >0.25 significant shift) triggering model review.

Discuss stratified sampling, SMOTE/ADASYN, class-weight adjustments, focal loss, and the importance of evaluating on the natural (imbalanced) test set.

Explain that constraints enforce economically logical relationships (e.g., higher debt-to-income → higher default probability) to satisfy regulatory explainability requirements.

TTC averages over the cycle for regulatory capital stability; PIT reflects current conditions for IFRS 9 staging and expected credit loss calculations.

Discuss using published macro data with appropriate lags, scenario-specific projections from central bank models, and separating macro effects from idiosyncratic risk.

Explain ranking features by absolute SHAP contribution and mapping the top negative contributors to human-readable reasons compliant with ECOA/FCRA requirements.

Describe running a new model (challenger) alongside the existing model (champion) in shadow mode, comparing performance on live data before full switchover.

Discuss imputation strategies (MICE, domain-specific defaults), missingness as an informative feature, and the use of public filing databases like EDGAR or Companies House.

KS measures the maximum distance between cumulative distributions of defaulters and non-defaulters; higher KS indicates better rank-ordering ability.

Feature stores ensure consistency between training and serving features, enable feature reuse across models, and provide versioning for audit trails.

Discuss shared-encoder architectures with task-specific heads, the benefit of shared representations for correlated targets, and calibration of probabilistic outputs.

Describe tokenizing transaction sequences, using positional encodings for time, self-attention over spending/payment patterns, and comparing to LSTM baselines.

Discuss surrogate explainers, SHAP-based global/local explanations, monotonic constraints, model tiering (simple for small loans, complex for large exposures), and governance documentation.

Cover scenario definition (GDP, inflation, unemployment paths), mapping macro factors to model features, generating stressed PDs, computing expected vs. unexpected losses, and portfolio-level aggregation.

Discuss input validation bounds, consistency checks across related features, anomaly detection on application data, and adversarial training techniques.

Explain the hazard function, proportional hazards assumption, baseline hazard estimation, and contrast with Random Survival Forests or DeepSurv which relax the proportionality assumption.

Describe streaming data ingestion (Kafka/Kinesis), sliding-window feature computation, drift-aware scoring with alerting thresholds, and integration with collections workflow systems.

Discuss domain adaptation, feature mapping across bureau schemas, GDPR constraints on feature use, transfer learning with fine-tuning on local data, and recalibration for base-rate differences.

Explain factor model decomposition, conditional PD given macro states, correlation structures (Gaussian copula, CreditMetrics-style), and separating portfolio diversification effects.

Discuss graph neural networks over borrower relationship graphs, propagation mechanisms, counterparty exposure aggregation, and backtesting against historical contagion events (e.g., Lehman cascade).

Walk through model explanation using SHAP, reviewing feature drivers, checking for data quality issues, documenting the decision, and establishing a formal override policy with escalation.

Discuss audit of current explainability coverage, implementing SHAP/LIME layers, creating a model documentation package, engaging model risk management, and potentially building a simpler challenger model.

Check for data pipeline issues, analyze PSI across features, examine macro environment shifts, evaluate whether the population composition has changed, and determine if retraining or recalibration is needed.

Discuss leveraging alternative data (mobile money, telco), building a thin-file model with limited features, using transfer learning from analogous markets, and designing a data-collection strategy to improve models over time.

Discuss the cost of false positives, the value of human review for borderline cases, fairness and disparate impact implications, and recommending a tiered decision framework with overrides.

Explain fair lending analysis (disparate impact testing, proxy discrimination risk), propose removing or regularizing the feature, discuss alternative geographic features with lower correlation to protected classes, and reference ECOA guidance.

Discuss retrieval-grounded generation, confidence scoring, human-in-the-loop verification for high-stakes extractions, and building automated cross-reference checks against the original document.

Separate model error from underwriting policy drift, check if the approved population shifted, analyze if promotional pricing attracted higher-risk segments, and assess whether macro conditions changed beyond the model's training data.

Discuss using broader corporate-default data with transfer learning, incorporating issuer-level ESG and environmental risk features, Bayesian approaches for data-sparse regimes, and scenario analysis for transition risks.

Advocate for a modular architecture where a foundation model augments but does not replace domain-specific models, highlight regulatory and explainability risks of monolithic systems, and propose a phased integration plan.

Describe Airflow DAGs for scheduled data pulls and feature computation, DVC for data and model versioning, MLflow for experiment tracking and model registry, and gated deployment with validation checks.

Discuss document chunking and embedding strategy, vector store selection (Pinecone, Weaviate), retrieval chain design, prompt engineering for accurate extraction, and evaluation of answer faithfulness.

Cover dataset preparation with labeled transcripts, choosing a base model (FinBERT or DeBERTa), fine-tuning with appropriate loss functions, evaluation on temporal holdout, and integration into a feature pipeline.

Explain online vs. offline feature stores, defining feature groups for borrower attributes and behavioral aggregates, ensuring training-serving consistency, and setting up access controls for sensitive financial data.

Discuss prompt templates that ingest SHAP feature-importance rankings, constrained generation for factual accuracy, human review workflows, and comparison with analyst-written reports for quality benchmarking.

Cover PSI tracking over time, AUC on delayed-label data, feature distribution monitors, prediction volume and latency, alert thresholds, and linking to incident response workflows.

Explain DVC's data versioning with remote storage, pipeline definition files, reproducible experiments, and how this satisfies regulatory requirements for model lineage and reproducibility.

Discuss Step Functions for orchestration, separate SageMaker endpoints for NLM and tabular models, API Gateway for unified access, latency optimization, and error handling for partial pipeline failures.

Describe unit tests for feature engineering, integration tests for data pipeline, performance gate checks (minimum AUC, maximum PSI), automated model promotion in MLflow, and rollback procedures.

Explain CatBoost's ordered target encoding that prevents target leakage, native categorical feature support reducing preprocessing, and how ordered boosting reduces overfitting on small default samples.

Look for evidence of professional courage, data-driven communication, understanding of model limitations, and a constructive resolution that balanced risk appetite with business needs.

Assess self-awareness, ability to identify root causes (data leakage, distribution shift), humility in acknowledging errors, and concrete improvements made to validation processes.

Evaluate communication skills, use of analogies and visual aids, ability to distill technical detail into actionable insights, and respect for the decision-maker's domain expertise.

Look for proactive detection, escalation process, impact assessment, remediation steps, and whether they implemented safeguards to prevent recurrence.

Evaluate learning habits, judgment in evaluating new techniques against production constraints, and ability to balance innovation with regulatory compliance and operational stability.