The GT-Score: A Robust Objective Function for Reducing Overfitting in Data-Driven Trading Strategies

Battle Count: 76

Directly addresses one of the most critical and pervasive problems in quantitative trading: backtest overfitting during strategy optimization. The GT-Score provides a practical, implementable composite objective function that can be directly integrated into existing backtesting and parameter-search pipelines. The empirical evaluation across 50 stocks, 15 years, multiple market regimes, and 14,000+ optimization trials provides substantial evidence. The 98% improvement in generalization ratio is a significant practical finding. However, the paper focuses on simple technical strategies rather than ML-based approaches, and the parametric Z-score assumption is a notable limitation for real-world financial data. The work is highly relevant for any quant developing or evaluating trading strategies.

The GT-Score formula itself is straightforward: a multiplicative composite of four components (μ, ln(z), r², σd) with a piecewise edge-case handler. The core calculation requires only basic statistics (mean, standard deviation, R-squared, downside deviation, Z-score). However, the full experimental framework (walk-forward validation with embargo periods, Monte Carlo with multiple seeds, random search optimization, multi-asset evaluation) adds moderate complexity. The author provides complete Python code, significantly lowering the barrier to implementation. No deep learning frameworks or complex model architectures are required.

Full Python source code for GT-Score implementation and backtesting framework provided as Supplementary Material. Processed result files for Monte Carlo and walk-forward experiments included. GitHub repository available (https://github.com/shep-analytics/gt_score). Data sourced from publicly available yFinance API. All experimental parameters (random seeds 42–56, 25 evaluations per trial, 9 walk-forward splits, 50 stocks, 3 strategies) are explicitly documented. The paper is published under CC BY license.