AI Market Risk Analyst Interview Questions

A strong answer explains the confidence-level interpretation, then covers historical simulation, variance-covariance (parametric), and Monte Carlo methods with their trade-offs.

Define each clearly with a concrete example; market risk covers price movements, credit risk covers counterparty default, operational risk covers internal process failures.

Cover the concept of applying extreme but plausible scenarios to portfolios, distinguish historical vs. hypothetical scenarios, and mention regulatory mandates like CCAR.

Mention libraries like yfinance or Bloomberg API, handling missing data, adjusting for splits/dividends, and aligning time-series across multiple assets.

Parametric assumes a distribution (e.g., normal); non-parametric uses historical data directly. Discuss when each is appropriate and the fat-tail problem.

Discuss using ML to dynamically weight historical scenarios, incorporate regime detection, or use kernel density estimation with learned bandwidth parameters.

Cover text preprocessing, section extraction (MDA, Risk Factors), NER for entities, sentiment scoring with FinBERT, and change detection year-over-year.

Explain CVaR as the average loss beyond VaR, its coherence property (subadditivity), and mention Basel III/FRTB's shift to ES as the primary risk measure.

Discuss hidden Markov models, regime-switching GARCH, change-point detection algorithms, and adaptive retraining windows for ML models.

Outline the agent architecture: tools for data retrieval, computation, and summarization; memory for context; how it chains quantitative outputs into narrative text.

Cover data drift (PSI, KS tests), concept drift, performance degradation tracking, and automated alerting thresholds with retraining triggers.

Explain systematic vs. idiosyncratic risk, common factor models (Fama-French, Barra), risk attribution by factor, and how ML can enhance traditional factor discovery.

GARCH is parametric with interpretable parameters; LSTM captures non-linear patterns but is a black box. Discuss data requirements, interpretability, and when to prefer each.

Discuss signal extraction pipelines, feature engineering, latency considerations, backtesting for alpha decay, and ensuring data provenance for compliance.

Cover tail dependence modeling, multivariate dependency beyond correlation, Gaussian vs. t-copulas, and their role in portfolio credit and market risk aggregation.

Discuss streaming architecture (Kafka), incremental VaR computation, GPU-accelerated repricing, LLM-powered anomaly narratives, and tiered alerting for risk breaches.

Cover grounding via RAG, structured output parsing, confidence calibration, human-in-the-loop review, and citation of sources in generated risk reports.

Discuss the shift to ES, P&L attribution tests, expected shortfall with liquidity horizons, non-modellable risk factor (NMRF) treatment, and desk-level model approval.

Explain parallel model architectures (parametric + ML), performance comparison dashboards, SHAP-based explainability parity, and documentation for SR 11-7 compliance.

Discuss attention mechanisms for cross-asset dependency learning, training on historical crisis episodes, tail-specific loss functions, and comparison with DCC-GARCH baselines.

Cover proxy bias in sector/geography risk scoring, disparate impact testing, feature importance audits for protected characteristics, and governance review processes.

Discuss walk-forward validation, expanding vs. rolling windows, crisis-period segmentation, Kupiec and Christoffersen tests, and stress-scenario backtests beyond historical data.

Cover chunking strategy for regulatory text, vector store selection (Pinecone, Weaviate), embedding model fine-tuning on financial domain, retrieval ranking, and citation verification.

Discuss prediction intervals, Bayesian approaches, ensemble disagreement metrics, confidence band visualization, and translating statistical uncertainty into business language.

Cover the interpretability-performance trade-off, SHAP/LIME for post-hoc explanations, intrinsically interpretable models, regulatory expectations, and tiered model governance.

Describe the sequence: real-time data ingestion, automated repricing, stress scenario injection, AI-driven impact summary generation, escalation to risk committee with LLM-narrated briefing.

Discuss hypothesis generation (liquidity stress, positioning unwind), drilling into factor loadings, cross-referencing alternative data sentiment, documenting findings, and recommending hedging actions.

Walk through model audit steps: feature importance review, scenario decomposition, benchmarking against alternative models, sensitivity analysis, and transparent documentation of assumptions.

Cover model documentation (methodology, data, assumptions, limitations), validation backtest results, change logs, explainability reports, and mapping to SR 11-7 or equivalent guidance.

Discuss hallucination detection via source verification, structured output validation, fact-checking layers, confidence scoring, human review gates, and continuous evaluation of LLM accuracy.

Address unique crypto risks (24/7 markets, liquidity fragmentation, exchange counterparty risk), alternative data sources (on-chain analytics, social sentiment), model recalibration, and regulatory landscape.

Cover breach analysis (systematic vs. idiosyncratic), backtest failure investigation (Kupiec test), model recalibration, risk limit review with portfolio managers, and escalation documentation.

Discuss SHAP/LIME integration for real-time inference, feature attribution dashboards, switching to intrinsically interpretable models where feasible, and building an explanation API layer.

Explain cross-validating sentiment with price action, checking for data pipeline lag, investigating if sentiment leads price, presenting both views to stakeholders, and designing an integrated signal framework.

Cover tiered automation (simple strategies auto-approved, complex ones flagged for human review), model validation requirements, kill-switch mechanisms, audit logging, and regular strategy re-evaluation.

Detail the agent design: document loader, text splitter, tool definitions (NER, sentiment, summary), memory for context, output parser for structured memo, and evaluation loop.

Cover data preparation, annotation strategy, training config (learning rate, epochs, batch size), evaluation metrics (F1, AUC), deployment via SageMaker, and A/B testing against the base model.

Walk through: experiment tracking (MLflow), containerization (Docker), CI/CD (GitHub Actions), SageMaker endpoint deployment, monitoring (CloudWatch, SageMaker Model Monitor), and rollback strategy.

Cover PDF parsing, chunking strategy (semantic vs. fixed-size), embedding model choice, vector store (Pinecone/Weaviate), retrieval ranking, prompt template design, and answer quality evaluation.

Explain defining JSON schemas for risk events, entity extraction, and sentiment scores; crafting system prompts; handling parse errors; validating outputs against schemas; and chaining functions for multi-step analysis.

Discuss feature definitions (technical indicators, macro factors, sentiment scores), Feast or Tecton for feature serving, point-in-time correctness, online vs. offline stores, and feature versioning.

Cover Kafka ingestion, windowed computations, online learning models (e.g., River library), autoencoder-based anomaly detection, alerting thresholds, and false positive management.

Describe generating global and local SHAP explanations, visualizing feature importance (summary plots, force plots), translating into plain-language narratives, and addressing 'why now' questions.

Discuss model diversity (parametric, ML, simulation), stacking/blending approaches, performance metrics (quantile loss, backtest coverage), dynamic weighting based on regime, and production serving.

Cover template libraries, few-shot example curation, output format standardization, version control for prompts, evaluation metrics for prompt quality, and a/b testing different prompt strategies.

Look for analytical rigor, persistence in investigation, clear communication of the risk, and a concrete impact on decisions or losses avoided.

Assess ability to translate statistical concepts into business language, use visualizations effectively, and balance honesty about limitations with confidence in the analysis.

Evaluate debugging methodology, intellectual humility, willingness to question the model, and whether they escalated appropriately rather than dismissing or blindly trusting the output.

Look for structured learning habits: following specific journals/arXiv, attending conferences (NeurIPS, RiskMinds), engaging in professional communities, and continuous project experimentation.

Assess professional courage, evidence-based argumentation, understanding of risk appetite frameworks, and ability to maintain relationships while standing firm on risk principles.