Research & Analytics
Controlled experiments, statistical rigor, and predictive models validated on real data with measurable outcomes.
How I Work
Every project starts with a question and a dataset. I design controlled experiments, validate with proper statistical methods (walk-forward CV, permutation tests, Benjamini-Hochberg correction), and only trust results that survive out-of-sample testing. If it ships, it has numbers behind it.
Systematic analysis of NHL betting market inefficiencies. Designed 235+ controlled experiments across 6,560 games and 100K+ odds records, applying Benjamini-Hochberg correction and permutation testing to rigorously evaluate market efficiency.
Key Findings
Methodology
Walk-forward cross-validation on rolling 3-season windows. Each experiment tests a specific hypothesis about market behavior (e.g., 'home favorites are overpriced after back-to-back losses'). Statistical significance assessed via Benjamini-Hochberg correction for multiple comparisons and permutation testing with 10,000 shuffles per hypothesis.
Results
The final model uses 158 engineered features including 5 custom Elo rating systems (overall, home/away, recent form, goaltender, situational), an expected goals pipeline built from 100K+ shot-level records, and rolling team performance metrics. Walk-forward accuracy of 60.9% on 6,560 fully out-of-sample games across 5 NHL seasons, with a Brier score significantly below the bookmaker baseline.
Impact
Findings deployed as PuckCast (puckcast.ai), a live analytics platform generating real-time win probabilities and value betting signals for every NHL game. Companion developer API (puckapi.com) serves model predictions and historical data. Platform has served 5,500+ unique visitors with sub-second response times.
Independent Research
Large-scale NFL prediction system benchmarking 600+ model configurations across 35 optimization rounds on 770,000+ plays of play-by-play data, systematically isolating the strongest predictive signals for game outcomes.
Key Findings
Methodology
Systematic grid search over 600+ model configurations combining feature sets, algorithms (Random Forest, Gradient Boosting, Logistic Regression), hyperparameters, and training window sizes. Each configuration evaluated via walk-forward backtesting across 6 complete NFL seasons. Play-by-play data from nfl_data_py covering 770,000+ individual plays decomposed into game-level predictive features.
Results
Identified two statistically independent signal families with zero feature overlap, each achieving ~67% standalone accuracy. When both signals agree (high-confidence consensus), accuracy reaches 77.6%. Backtested against historical closing lines, this consensus generates +41.6% ROI, surviving robustness checks across all 6 test seasons individually.
Impact
Proprietary backtesting infrastructure enables rapid iteration on new feature hypotheses. The dual-signal architecture provides natural confidence calibration: predictions are only surfaced when both independent models agree, dramatically reducing false positive rate. Public frontend currently in development.