Real-Time Multi-Step Time Series Forecasting Using a GA-Optimized VMD-RF Framework

In vital industries including finance, energy, transportation, and meteorology, time series forecasting is essential. Prediction accuracy is frequently hampered by issues including nonlinearity, multi-scale patterns, and significant noise in real-world data. In order to tackle these problems, this study suggests a new hybrid forecasting framework called GA-VMD-RF, which integrates a real-time decomposition mechanism for deployment-ready multi-step forecasting and combines Genetic Algorithm-optimized Variational Mode Decomposition (VMD) with Random Forest (RF) regression. The proposed model introduces a permutation entropy-guided GA to dynamically tune VMD parameters, enhancing decomposition quality and preserving modal predictability. Each decomposed sub-series is modeled independently using RF, capturing distinct temporal dynamics. Unlike traditional full-series decomposition methods that risk information leakage, our real-time strategy performs VMD solely on the training data and incrementally updates decomposition during inference, ensuring fair and practical evaluations. Furthermore, an Average Window Reconstruction (AWR) mechanism is employed to fuse predictions from overlapping sliding windows, improving robustness and reducing temporal drift. Extensive experiments on diverse datasets—including U.S. Treasury yields, Brent crude oil prices, and wind speed observations—demonstrate that GA-VMD-RF outperforms classical models (SVM, LSTM) and VMD-based baselines in both short- and long-horizon tasks. Comparative results show improvements of up to 51.9% in MAPE over single-model baselines and 8.1% over traditional VMD-RF setups.