20262026, Python, Market Microstructure, Price Discovery, Econometrics, VAR, VECM, Granger Causality, Hasbrouck, Reproducibility
Coinbase vs Kraken Market Microstructure and Econometric Price Discovery
resultAcross 9 econometrically usable paired sessions, Coinbase showed stronger short-horizon price leadership over Kraken: BH-significant Granger evidence appeared in 6/9 sessions Coinbase → Kraken versus 1/9 in the reverse direction.
- accepted sessions
- 10
- econometrically usable
- 9
- BH-significant Granger CB→KR
- 6/9
- BH-significant Granger KR→CB
- 1/9
A confirmatory empirical study of BTC/USD price discovery across Coinbase and Kraken using synchronized market data, VAR/VECM models, Granger causality, predictive regressions, impulse responses, Gonzalo–Granger component shares, and Hasbrouck information-share bounds. The evidence favors stronger Coinbase short-horizon leadership, while long-run price discovery remains heterogeneous across sessions.
status: complete release: v1.0.0 accepted sessions: 10 econometrically usable: 9 confidence: strong short-horizon evidence; heterogeneous long-run evidence
2026C++, Market Microstructure, Limit Order Book, ITCH Replay, Market Making, Avellaneda-Stoikov, Queue Diagnostics, Reproducibility
Matching Engine, ITCH Replay, Market-Making Diagnostics, and Artifact Validation
result3.7M events/sec benchmark, 12,423 QQQ ITCH messages, 30-seed strategy statistics, 10-seed fill-rate diagnostics, and validator-backed artifacts.
A deterministic C++ market microstructure simulator with price-time priority matching, Nasdaq ITCH replay, naive vs Avellaneda-Stoikov market-making experiments, queue-position diagnostics, ten-seed fill-rate mechanism tests, and artifact validation.
status: complete CTest: 69/69 passed pytest diagnostics: 21 passed ruff: clean CI: green validator: passed
2026Python, Options, Volatility, SABR, Heston, Numerical Methods, Model Diagnostics
SABR/Heston Calibration, Robustness, and Failure Diagnostics
result101 tests, 206 row recommended universe, SABR median RMSE 0.0190 to 0.0077, Heston global RMSE 0.1174.
Built a reproducible Python options research engine for chain cleaning, forward extraction, OTM IV surface construction, static-arbitrage diagnostics, SABR robustness testing, Heston calibration, and same-universe model comparison. The main result was diagnostic: filtering reduced SABR median RMSE from 0.0190 to 0.0077, while global/per-expiry Heston underfit SABR despite passing synthetic recovery checks.
status: complete tests: 101 passing raw rows: 1,169 clean rows: 957 OTM rows: 595 confidence: high on reproducibility and diagnostics, limited by one yfinance snapshot
2025Python, algorithmic trading
result#194 algo, top 1.4% of 18,800 teams.
the hard part was adapting strategy across rounds under sparse feedback.
status: complete artifact: public leaderboard rank: #194 algo
2026honors thesis, Bayesian inference, regime detection, HMM, CUSUM, CPPI
resultBayesian regime detection, HMM plus CUSUM, CPPI drawdown control, and S&P 500 crisis validation.
the hard part was accepting that a well trained Bayesian agent can still stay wrong too long.
status: complete artifact: thesis PDF confidence: high
2026Python, Bayesian Inference, LLM Agents, Adaptive Oversight, Experimental Design, Reproducibility
Budget-Constrained Oversight for Hierarchical LLM Workflows
result144 confirmatory and held-out trajectories across 36 matched quartets, three constraint domains, and 1,945 passing tests.
- trajectories
- 144
- matched quartets
- 36
- constraint domains
- 3
- passing tests
- 1,945
A reproducible research framework for deciding when limited oversight should inspect, escalate, or intervene in hierarchical LLM workflows. I built rule-based and Bayesian monitors, strategic attackers, objective violation scorers, typed constraint inheritance, provider-cost gates, and deterministic artifact validation.
v1.0.0 released artifact: public repository and technical report confidence: high in implementation and reproducibility infrastructure, external validity limited to synthetic tasks and one provider model
2025PyTorch, neural SDEs, rough volatility
resultResearch prototype with repo; no public performance number claimed.
the hard part was making path structure matter without overclaiming the metric.
status: research prototype artifact: GitHub repo confidence: medium
2026Python, Kalshi API, DuckDB, scikit learn, XGBoost
result122 tests, 85% coverage, documented API limits.
the hard part was proving which historical data did not exist.
status: framework complete tests: 122 passing coverage: 85% confidence: high on engineering, documented data limits
2026Python, NumPy, SciPy
resultOut of sample Sharpe near 0.7 after fees.
the hard part was reporting the modest number after fixing the inflated one.
status: complete result: out of sample scope: single asset and one pair confidence: medium