Numerai Challenge

Competing in Numerai, a global ML tournament where predictions are used to trade a live hedge fund.

Numerai is a data science competition where participants build ML models to predict obfuscated stock market data. The best models are combined into a meta-model that trades Numerai's global equity hedge fund. Submissions are scored on CORR (correlation to target) and MMC (unique contribution beyond the crowd). Payout formula: 0.5×CORR + 2×MMC.

• Built a full ML training pipeline from scratch on 2.7 million stock observations across 20+ years of market history
• Trained and deployed four separate model variations, each optimized around a different edge
• Implemented era-boosting and feature neutralization to reduce regime overfitting
• Model reweights itself across different market regimes to maintain consistency over time
• Isolates signals that are genuinely predictive by removing exposure to known risk factors
• Weekly submission pipeline

• Top 10% of all participants globally
• Top 30% among the subset of professional competitors who stake their own money on model performance

• CORR — Spearman correlation of predictions to the 20-day forward return target
• MMC — contribution of my model's unique component beyond the meta-model
• Reputation — rolling 1-year average score, determines leaderboard rank

The biggest takeaway wasn't the ranking. In applied investing, emotional control is just as important as the signal itself. A model doesn't panic, second-guess, or chase losses. Building one from scratch made that very clear.

• Python, LightGBM, CatBoost, scikit-learn, pandas
• Machine learning, gradient boosting, ensemble methods
• Feature engineering, era-based cross-validation
• Numerai API, automated submission pipeline

• Applied ML in a live, competitive, real-money-adjacent environment
• No prior ML experience at the start, built everything from scratch
• Understanding of market prediction as a structured ML problem
• Discipline to compete consistently week over week