Risk-Adjusted Performance of Random Forest Models in High-Frequency Trading

Despite the theoretical limitations posed by the Efficient Market Hypothesis, technical indicators remain widely used in high-frequency trading (HFT). However, their effectiveness at minute-level frequencies, where market microstructure effects dominate, remains underexplored. This study evaluates the role of traditional technical indicators in Random Forest Regression (RFR) models using minute-level SPY data across 13 distinct configurations.

Our analysis reveals a significant divergence between in-sample and out-of-sample performance. While models demonstrated strong in-sample performance (R²: 0.749–0.812), their predictive power collapsed in out-of-sample testing, often yielding negative R² values. Feature importance analysis shows that price-based features overwhelmingly drive model decisions, accounting for over 60% of importance, while widely used technical indicators such as RSI and Bollinger Bands contribute only 14–15%.

Although integrating technical indicators slightly improved risk-adjusted metrics—yielding Rachev ratios between 0.919 and 0.961—models incorporating these indicators still underperformed a simple buy-and-hold strategy, generating negative returns between -2.4% and -3.9%. These results suggest that traditional technical indicators may be more effective for risk management than for return prediction in HFT settings.

Our findings underscore the importance of adaptive feature selection and regime-specific modeling over reliance on conventional technical indicators. Moreover, the stark contrast between in-sample and out-of-sample performance highlights the necessity of rigorous out-of-sample validation in algorithmic trading research. This study contributes to ongoing discussions on the limitations of technical indicators in HFT and provides insights for both practitioners and researchers aiming to develop more robust predictive models in high-frequency market environments.