Forecasting volatility by using variational mode decomposition and machine learning models

This study proposes a novel hybrid data science framework that integrates Variational Mode Decomposition (VMD), a powerful signal processing technique, with advanced machine learning models—including XGBoost, Random Forest, and Support Vector Regression (SVR)—to forecast complex, non-stationary time series. We demonstrate the framework's efficacy on the canonical problem of predicting the S\&P 500's realized volatility. Our empirical analysis shows that preprocessing the volatile and noisy data with VMD before applying machine learning drastically enhances predictive accuracy by disentangling the source signal into simpler, more stationary sub-components. The results indicate that all VMD-enhanced models substantially outperform their standalone ML equivalents and traditional econometric benchmarks. Among the hybrids, VMD-SVR emerges as the top performer, achieving the lowest Root Mean Square Error and Mean Absolute Deviation. Notably, these performance gains are most significant during high-volatility periods like the 2008 financial crisis and the 2020 pandemic, highlighting the framework's robustness and practical utility for decision-making under uncertainty. Although VMD-based models show a slight tendency to over-predict, this is outweighed by a remarkable reduction in overall forecast error. The study confirms that combining signal decomposition with machine learning creates a powerful and generalizable template for predictive modelling of complex time series across domains, with immediate applications in financial risk management.