Wavelet-Driven Multi-Model Ensemble: A Synthesis Box for Time Series Forecasting
Rui Tang, Minglei Lyu, Yuwen Zheng · 2024
This paper presents a wavelet-driven multi-model ensemble framework for time series forecasting, aimed at addressing the limitations of traditional and machine learning-based methods in handling non-stationary data, periodic components, and stochastic variations. The framework begins with wavelet decomposition, separating the original data into trend, periodic, and stochastic components. The periodic component is analyzed using Fast Fourier Transform (FFT) and predicted via an AutoRegressive Integrated Moving Average model, followed by reconstruction through inverse Fourier transform. For the trend and stochastic components, feature engineering techniques are employed to train an Adaboost model, enabling iterative forecasting with dynamically updated feature sets. Experimental results demonstrate the framework's effectiveness, achieving a SMAPE of 0.8%, thereby verifying the model's robustness in handling diverse temporal dynamics. The results highlight the robustness of the model over different temporal dynamics, and the proposed method will greatly improve the reliability of the model in diverse application scenarios.