4 Main Pillars of a Winning Strategy [COMPETTITION-5]

1). Code Structure & Technical Rules (Rigid) Mandatory De-meaning: The output signal must have a mean of zero ($\sum P_j = 0$) at every timestamp. Code Isolation: All feature engineering logic and helpers must be inside a Predictor subclass. CPU Limitations: The model must be lightweight. Maximum training time is 4 minutes and maximum prediction time is 60 seconds without GPU. 2). Signal Geometry & Novelty (Prize Pool Key) The maximum prize ($50,000) is only achieved if the ecosystem produces many unique quality signals. Your signal will be completely discarded (prize = 0) if it has a correlation$\ge 0.5$ with a higher Sharpe signal already in the Quality Set. You must achieve a minimum angular distance of 60° from existing Signal Cities using non-linear derived features or different time horizons. 3). Data Character & Evaluation (Walk-Forward) Data is obfuscated per epoch. The asset identity (ticker) is random and changes between periods. The model should not memorize the behavior of a specific stock, but rather should extract macro relationships between cross-sectional features (cross-sectional patterns). 4). Strict Overfitting & Leakage Constraints The use of Signal Cities coordinates as input features is strictly prohibited (data leakage). Signals with historical Sharpe values ​​that are “too perfect” or unnatural will be immediately flagged as data manipulation (reverse-engineering) and disqualified.