AI Market Microstructure Analyst Interview Questions

Describe the order-driven matching mechanism, bid-ask sides, price-time priority, and contrast with dealer/market-maker models like NASDAQ's old structure.

Cover adverse selection, inventory holding costs, order processing costs, and how spread relates to volatility, liquidity, and information asymmetry.

Explain Volume-Synchronized Probability of Informed Trading, its construction from volume buckets, and its use as a real-time toxicity/flow toxicity indicator.

Discuss how market orders signal impatience and potentially informed trading, while limit orders provide liquidity but face adverse selection risk.

Define temporary vs. permanent price impact, explain how large orders move prices, and why minimizing impact is a core objective of algorithmic execution.

Describe the single-period model with informed trader, noise trader, and market maker; explain lambda as price impact. Discuss limitations: no HFT, single asset, static game.

Define order imbalance as (bid volume - ask volume) / total volume over a rolling window; discuss optimal window length as a tradeoff between signal-to-noise and latency.

Adverse selection is losses to informed traders; inventory risk is directional exposure from holding positions. Discuss how bid-ask spread is compensation for both.

Effective spread measures cost at trade time; realized spread measures P&L after the market reverts. The gap captures the adverse selection component.

Explain self-exciting nature: one order increases probability of subsequent orders; captures clustering and contagion in order flow that Poisson cannot.

Tradeoff between market impact (fast execution) and volatility risk (slow execution); quadratic utility minimization yields optimal trajectory; discuss temporary vs. permanent impact parameters.

Cover order imbalance at multiple levels, queue position features, bid-ask spread, trade flow toxicity, volume-weighted distance from mid, micro-price, and recent trade flow momentum.

Describe the CNN layers extracting spatial LOB features, LSTM/Inception layers capturing temporal dependencies, and the encoder-decoder structure; explain why it outperforms raw LSTM on LOB data.

Queue position determines probability of being filled before the price moves; models like the symmetric zero-intelligence model quantify fill probabilities; critical for passive execution algorithms.

Roll (1984) estimates spread from serial covariance of price changes; negative serial covariance implies spread. Limitations: fails with discrete prices, direction changes, and high-frequency noise.

Describe feature engineering (order-to-cancel ratios, time-in-book distributions, distance-from-mid of cancelled orders), online learning or sliding window models, and the tradeoff between false positives and detection speed.

Crypto has 24/7 trading, fragmented liquidity across venues, wash trading concerns, and different informed trader dynamics; models need to account for cross-exchange arbitrage, on-chain data as additional signals, and regime changes around major events.

Discuss concept drift detection, adaptive recalibration (online learning, exponential decay weighting), regime-switching models, ensemble approaches, and the importance of monitoring model performance metrics in production.

OW model assumes exponential decay of impact (mean-reverting transient impact), separating transient and permanent components; better for modeling multi-day execution; discuss equilibrium implications and parameter estimation.

Self-attention over LOB snapshots at different time steps; expect attention to focus on periods of high order imbalance, trade bursts, and spread widening; discuss positional encoding challenges for irregularly spaced financial timestamps.

Cover VPIN, bulk volume classification, and ML approaches; discuss features (trade size, timing, price reversion); address production challenges: latency requirements, label generation, class imbalance, and the regulatory implications of misclassifying flow.

Discuss spread compression, displayed depth changes, price improvement, midpoint execution; build a counterfactual model using natural experiments; discuss difference-in-differences methodology and venue-level analysis.

Micro-price weights the bid and ask by the probability of execution at each level, incorporating queue sizes; derive the intuition from the Glosten-Milgrom model; discuss extensions to multi-level micro-prices.

Discuss state space (remaining inventory, time, LOB features), action space (trade rate), reward function (implementation shortfall with impact penalty); cover model-based vs. model-free approaches; discuss simulation environment design and domain randomization.

Discuss how HFT market makers have reduced spreads but face different adverse selection dynamics (latency arbitrage, speed races); cover the empirical evidence on HFT market making profitability and its implications for market resilience during stress events.

Systematic diagnosis: check data quality, venue-level fill rate analysis, compare against alternative venues, analyze order book depth changes, check for regime changes in volatility or volume, investigate exchange-side changes (matching engine upgrades, fee changes), and assess whether your model's market impact parameters need recalibration.

Engineer features from closing auction dynamics, overnight news/sentiment (using LLM on news feed), options-implied volatility, recent intraday spread patterns, volume profile, earnings calendar; build a classification model with proper cross-validation respecting temporal ordering.

Analyze fill rates, adverse selection (post-trade price reversion), midpoint execution rates, latency characteristics, information leakage (compare price movement before vs. after fills), toxic flow proportion, and compare against lit venue benchmarks.

Immediate: trigger risk limits, switch to conservative execution mode, assess position exposure; analysis: examine order cancellation rates, cross-venue liquidity dynamics, identify whether it's fundamental or technical, analyze whether your impact models are still valid in the stressed regime, and prepare a post-event microstructure reconstruction.

Evaluate: discrimination (AUC, precision-recall on toxic flow events), calibration (predicted probabilities vs. observed), latency (inference time vs. signal decay), robustness (performance across regimes and instruments), backtest impact (does acting on signals improve execution quality), and monitoring plan for degradation detection.

Estimate market impact using calibrated models, analyze historical intraday volume and spread profiles for the stock, model the tradeoff between urgency and impact, design a parent-child order strategy, incorporate real-time adaptation based on fill quality, and set up TCA monitoring against implementation shortfall benchmark.

Analyze auction mechanism design (frequency, order types allowed), model the expected change in spread and fill rates, compare with international precedent (LSE periodic auctions), build simulation using historical order flow, and estimate impact on your firm's execution quality metrics.

Analyze the effective depth (volume available within X bps of mid), compute fill probability as a function of order size, distinguish displayed vs. actual liquidity using fill rate analysis, account for sub-penny hidden orders, and build a liquidity score that weights depth by reliability.

Options have fragmented, less liquid order books; the underlying drives most of the microstructure dynamics; model cross-asset order flow; address the challenge of managing delta hedging costs as a microstructure problem; discuss the role of volatility surface dynamics and the complexity of multi-leg order books.

Feature selection for low computational cost (order imbalance, trade-to-quote ratio, spread), lightweight model (gradient boosted trees or distilled neural network), pre-computed feature pipelines, co-located inference, and careful evaluation of the latency-accuracy tradeoff; discuss whether a simple heuristic could outperform a complex model in this regime.

Describe the RAG pipeline: ingest SEC PDFs, chunk and embed with a domain-specific embedding model, build a vector store, and create an agent that can answer questions, compare proposed rules, and extract impact-relevant sections; discuss handling legal/regulatory nuance.

Cover data pipeline (S3 + feature store), training jobs with hyperparameter tuning on SageMaker, experiment tracking with MLflow, model registry and staging, A/B deployment with shadow mode, monitoring for data drift and prediction quality, and rollback procedures.

Fine-tune a financial BERT model (FinBERT) on labeled financial sentiment data, generate real-time sentiment scores, engineer sentiment features (rolling average, sentiment change rate, source-weighted aggregation), and integrate as additional input features to your LOB prediction model; discuss latency requirements for news ingestion.

Describe topic design (raw quotes, trades, LOB snapshots as separate topics), Flink jobs for feature computation (order imbalance, trade flow, spread metrics), windowing strategies, exactly-once processing guarantees, and downstream consumers for model inference and monitoring dashboards.

Describe MLflow experiment structure (organizing by signal/strategy), logging of backtest metrics alongside ML metrics, model versioning tied to specific data snapshots and feature sets, reproducibility through config files and data versioning (DVC), and team collaboration patterns.

Describe the LOB simulator (agent-based or learned), realistic order fill simulation accounting for queue position, market impact feedback loop, stochastic volume and volatility processes, and the importance of domain randomization; discuss using OpenAI Gym-style interfaces and custom reward shaping.

Discuss online vs. offline feature store separation, streaming feature computation in Flink/Kafka Streams, pre-aggregated features at multiple time horizons, feature freshness requirements, consistency guarantees between training and serving, and tools like Feast or Tecton.

Discuss productivity gains for boilerplate (data loading, plotting, backtest scaffolding), pitfalls in financial logic (off-by-one errors in time alignment, look-ahead bias in backtests), importance of human review for domain-critical code, and the workflow of Copilot suggestions + careful unit testing.

Describe project structure with shared data modules and asset-specific model configs, Lightning's callback system for custom financial metrics, W&B sweeps for hyperparameter optimization across assets, artifact management for dataset versions, and dashboard design for comparing model performance across markets.

Describe parameterized notebooks for systematic experiments, papermill for running notebooks with different parameter sets (different assets, date ranges, model configs), version control with nbstripout to keep diffs clean, and automated report generation for weekly research reviews.

Look for awareness of common pitfalls (look-ahead bias, survivorship bias, market impact assumptions, overfitting to historical regimes), a structured debugging approach, and concrete changes to their research process as a result.

Look for systematic habits: following specific journals/conferences (Journal of Financial Markets, arXiv q-fin), engaging with practitioner communities, hands-on experimentation with new tools, reading exchange regulatory filings, and a balance between breadth and depth.

Look for ability to abstract away mathematical details, use intuitive analogies and visualizations, connect findings to business impact (cost savings, risk reduction), and adapt communication style to the audience.

Look for intellectual honesty about uncertainty, a framework for decision-making under uncertainty (confidence intervals, scenario analysis), willingness to be conservative when stakes are high, and the ability to clearly articulate assumptions and their implications.

Look for systematic data validation habits, healthy skepticism of 'clean' data, attention to detail in reconciliation, and the discipline to investigate anomalies rather than dismissing them. Bonus for examples involving exchange feed issues or corporate actions.