Applying machine learning for enhanced signal prediction and risk management in multi-asset futures portfolios

This study examines an innovative volatility-based multi-asset strategy that integrates forecasts based on machine learning models, dynamic risk budgeting, and adaptive filter and signal generation. The aim of the study was not only to confirm the practical feasibility and performance of the developed approach, but also to systematically evaluate its limitations and potential extensions. The strategy was tested using extensive empirical analyses of twelve liquid futures from the asset classes equities, commodities, bonds, and foreign exchange. A central component of the work was the use of a XGBoost model to forecast future volatility levels. The results of the forecast quality were promising: High 𝑅2 values of over 0.80 were achieved, particularly for equity and bond futures, accompanied by low forecast errors (RMSE) in the range of a few basis points. These results confirm the high precision and relevance of the model forecasts for asset allocation and risk management. In addition to the pure forecast results, particular attention was paid to the effectiveness of the implemented Bollinger-based filter mechanism, which isolates significant market phases and makes them systematically usable. Statistical tests, particularly the Kolmogorov-Smirnov test, confirmed significant differences between the return distributions in different volatility regimes. However, the filter logic also revealed a critical sensitivity to changes in market regimes, which requires regular recalibration of the parameters and significantly limits the robustness of the approach in less volatile, sideways markets. The study also identified fundamental challenges regarding the automated and reliable detection of market regime changes. The paper concludes with detailed recommendations for future research directions. It suggests focusing on regime-dependent parameter calibration. It also recommends analyzing the influence of the rebalancing pace on portfolio decisions to further optimize the strategy in practical trading applications. Another promising area of research concerns the use of leveraged products to improve the return profile. In summary, this study provides a comprehensive methodological and practical basis for the development and implementation of robust quantitative trading strategies. At the same time, it highlights the challenges and limitations of the current approach and offers sound perspectives and starting points for further research and optimization in the field of quantitative asset management.