Where uncertainty is high, structure becomes the advantage.
I am drawn to problems where the data is noisy, the structure is hidden, and the decision still matters. Markets are the cleanest version of that challenge: uncertainty, incentives, timing, feedback, and consequences compressed into a live system.
My default mode is to make ideas falsifiable: turn the question into data, the data into a model, the model into a test, and the test into a decision rule that can be challenged.
Quant systems built as instruments.
Each system is built around a market mechanism and an empirical question, then instrumented for deterministic replay, controlled experiments, implementation realism, and measurable failure modes.
Queue-Aware Market-Making & LOB Replay Engine
An event-level market simulator that reconstructs queue state and models passive fills under price-time priority, latency, cancellations, adverse selection, inventory pressure, fees, and stochastic order flow. Deterministic replay supports counterfactual quoting experiments against identical market trajectories.
- 412M book events
- 2.4M ev·s⁻¹ replay
- p99 1.8 µs
- 23 symbols × 180 sessions
Market questions resolved through evidence.
Working papers built around explicit market mechanisms, identifiable estimands, point-in-time data, out-of-sample validation, and implementation-aware economics. Each study asks not simply whether an effect exists, but what generates it, where it breaks, and whether its information survives translation into a trade.
Queue Position as a State Variable: Fill Hazard, Toxicity, and the Economics of Passive Liquidity
A competing-risk study of passive execution that estimates fill and cancellation hazards conditional on queue-ahead volume, order-flow imbalance, microprice, spread state, cancellations, and latency. Post-fill mark-outs decompose realised spread capture from adverse selection to estimate when apparently favourable fills are economically toxic.
- 2.9M order lifetimes
- fill C-index 0.74
- 10 s mark-out −0.42 bps
- net passive edge +0.18 bps
Conditional Alpha Under Regime Drift: Cross-Asset Residuals, Lead-Lag Structure, and Signal Half-Life
A walk-forward study of whether cross-asset predictability survives factor neutralisation, changing volatility and liquidity states, and repeated model selection. Horizon-specific information coefficients, decay curves, turnover-adjusted returns, and orthogonalised feature contributions distinguish persistent information from transient correlation.
- 1.7M name-days
- Δ IC +0.011 after neutralisation
- half-life 2.6 sessions
- net IR 0.94
Event-Time Volatility Surfaces: Arbitrage-Free Repricing Around Scheduled Information
A study of how scheduled information is embedded into implied volatility before and after events. Constrained surface calibration separates level, skew, curvature, and term-structure effects, while delta- and vega-aware attribution tests whether observed dislocations survive hedging error, spread, and surface re-marking.
- 18.6K events
- 74K event surfaces
- RMSE 0.61 vol pts
- ATM crush −6.8 vol pts
Execution as a Control Problem: Transient Impact, Fill Risk, and Adaptive Scheduling
A state-space study of execution under stochastic liquidity and decaying market impact. Adaptive policies condition participation on spread, depth, volatility, remaining inventory, and impact recovery, then compete against TWAP, VWAP, and POV under identical order and market paths.
- 41.2K matched paths
- impact half-life 4.7 min
- IS −2.4 bps vs TWAP
- −0.6 bps vs POV
Questions worth solving.
A set of market questions I keep returning to because the answer depends on mechanism, state, horizon, and implementation rather than a single backtest statistic.
When does latent liquidity dominate the visible book?
How much of short-horizon price formation can be explained by queue dynamics, cancellation behaviour, replenishment, and hidden liquidity rather than displayed depth alone?
When is lead-lag genuine information transmission rather than shared exposure?
Can delayed cross-asset responses be separated from common factors, asynchronous price discovery, stale prices, and correlated order flow strongly enough to produce incremental predictive information?
Does the volatility surface reveal information before spot incorporates it?
Do changes in skew, curvature, and term structure contain forward-looking information distinct from realised volatility, spot moves, positioning, and scheduled-event risk?
When does optimal execution require forecasting the decay of alpha itself?
How should execution adapt when market impact, liquidity recovery, fill probability, and signal half-life evolve jointly rather than independently?