Overcoming Overfitting: A New Approach to Ml Generalization

Abstract

Overfitting remains a critical challenge in machine learning, particularly for systems intended to generate reliable predictions in real-world contexts. Traditional approaches—such as regularization, cross-validation, and ensemble methods—can mitigate overfitting to varying degrees but may not consistently eradicate spurious correlations. This dissertation introduces a novel framework, the GT-Score, which combines performance and stability metrics into a single objective function, thereby discouraging the model from simply memorizing transient patterns.

Using historical stock market data, we systematically compare the GT-Score against widely used loss and objective functions such as a simple maximization function, the Sharpe Ratio, and regularized regression. To ensure that it is generalizable across multiple domains of machine-learning we employ three distinct optimization methods: random search, Bayesian optimization, and genetic algorithms. This allows rigorous testing of how different hyperparameter-tuning paradigms interact with various loss metrics. Across all methods, the GT-Score demonstrates a marked advantage in aligning training outcomes with validation performance, indicating stronger resilience to overfitting.

Results highlight that although classical regularization does improve model generalization, it is often outperformed by the GT-Score in scenarios where spurious correlations abound. The architecture of the GT-Score naturally penalizes excessive volatility, small sample trades, and lack of consistency over multiple periods, making it particularly suitable for dynamic and potentially noisy domains like financial markets. While this research focuses on a supervised learning approach within a trading context, the proposed methodology can be adapted to broader machine-learning settings.

These findings lay groundwork for integrating the GT-Score into various machine-learning pipelines. Future work will investigate its scalability for neural networks, its adaptability to other high-stakes areas such as healthcare, and robust comparisons with advanced Bayesian or meta-learning frameworks to further validate its capacity to overcome overfitting.