Financial Time Series Forecasting Based on Sliding Window–Variational Mode Decomposition and Deep Learning

To address the time-series characteristics of financial data, this paper proposes a data preprocessing method based on Sliding Window–Variational Mode Decomposition (SW-VMD). The method decomposes and reconstructs stock index closing prices and return time series, transforming nonlinear and nonstationary sequences into linear and stationary data. The processed data is then used as input for a Long Short-Term Memory (LSTM) neural network to predict future stock index closing prices and returns. Empirical analysis adopts trend accuracy as the evaluation metric to reflect the model's capability in forecasting the upward or downward trends of the next day's closing price and return. Results indicate that, compared to models without data decomposition, the LSTM model enhanced with SW-VMD shows significant improvements in trend prediction accuracy.