Studying structural breaks across markets and models, and building systems to test what survives them.
Inventory-aware market-making (C++20), equity futures
microstructure, and systematic long/short equity,
grounded in quantitative model validation.
IIT Roorkee → Wells Fargo → NC State, where I'm
completing a Master of Financial Mathematics (MFM).
- Crypto perp futures
- ES calendar spreads
- Russell 2000 momentum
- Credit model diagnostics
- AS market-making
- XGBoost L/S reclassifier
- Mean reversion (z-score)
- C++20 / Python
- Regime-aware parameter adaptation
- Honest failure analysis
- Real-tape fill replay
- Walk-forward validation
Inventory-aware market-maker on BTCUSD perpetual. MLE-calibrated fill intensity from real-tape replay (not Poisson simulation). Vol-adaptive stationary horizon τ with hard inventory bounds (±5 lots). 185 walk-forward windows.
- No queue-position model: any fill at a touched price assumed ours; optimistic at scale
- ~49% stale-quote fills at vol_budget=0.5: PnL is a lower bound (price-improved in real LOB)
- No maker rebates modelled: reported PnL is a lower bound if rebate schedule applies
- Funding rate f = 0: perpetual carry not wired into reservation price
Bimodality-motivated 4-quadrant XGBoost reclassifier on 12-1 momentum (Russell 2000, $10M notional, monthly rebalance). 10-member ensemble, 36-month rolling window, 19 cross-sectionally ranked features. Primary alpha: GL quadrant (bottom-decile reversal, +1.67%/mo), orthogonal to the momentum factor.
- XGB probability spread (GW−GL) = +0.0012, near-random; model is a volatility-regime filter, not a stock picker (vol_12m dominant feature, importance 0.1132)
- Market state gate halves exposure in DOWN regimes (IWM 12m < 0); alpha concentrated in UP-market states, returns are regime-dependent (Cooper et al. 2004)
- Momentum crash risk (Daniel-Moskowitz 2016): crashes cluster in high-vol post-bear periods; market state gate mitigates but does not eliminate sharp reversal exposure
Fair Value deviation-based mean-reversion on ES calendar spread (SOFR + div yield + expiry differential). Databento MBP-10 nanosecond data. Primary entry at |z| > 2.5; HC add-on doubles position when |z_fill| > 3.0. European session is a structural alpha source across all 4 roll windows.
- Roll window dependency: signal computed per expiry cycle
- Cost sensitive: $8.04/lot (exchange + NFA + broker) already baked in
- European session edge may compress as roll approaches expiry
Ranked 42nd globally in Phase 1 (~18k teams); achieved final rank of ~900 (top 5%) across all participants. Multi-product market with synthetic derivatives, ETF arbitrage, and manual trading rounds.
- Phase 1 → Phase 2 rank drop (42 → ~900): manual trading rounds penalized automated-only strategies
- Later rounds introduced cross-product dependencies not captured in Phase 1 models
- Ranked 42nd globally in Phase 1 (~18,000 teams)
- Final rank ~900, top 5% overall
- Stochastic Calculus for Finance · Options & Derivatives Pricing
- Monte Carlo Methods · Statistical Learning · Statistical Inference
- Linear Models & Regression · Probability & Stochastics for Finance
- Diagnosed regime-conditional signal decay in Cards ($48B) credit models using KS and PSI drift metrics
- Identified feature-level drift predicting scoring model instability ahead of formal model review cycles
- Quantified model sensitivity to macro and behavioral regime shifts across multi-million-record datasets
- Manager's Spotlight Award