Wavelet-Transformer Hybrid Networks for multiscale stock market price forecasting: an integrated approach to temporal feature learning

Stock market forecasting remains one of the most challenging tasks in financial analytics due to the inherent non-stationarity, high volatility, and complex multi-scale temporal patterns in price movements. Traditional deep learning approaches often struggle to capture information across different time scales simultaneously. This paper introduces a novel Wavelet-Transformer Hybrid Network (WTHN) that integrates Discrete Wavelet Transform (DWT) for multi-scale temporal decomposition with Transformer architecture for modeling long-range dependencies. The DWT decomposes stock price time series into approximation and detail coefficients at multiple resolution levels, revealing patterns that operate at different frequencies. These multi-scale features are then processed through a specialized Transformer encoder enhanced with positional encoding and multi-head self-attention mechanisms. We evaluate our approach on daily stock price data from three major technology companies: IBM, Apple, and Intel, spanning 2015-2024. Experimental results demonstrate that WTHN significantly outperforms traditional LSTM networks and vanilla Transformer models across multiple evaluation metrics. The proposed model achieves RMSE improvements of up to 23.7 percent compared to baseline approaches, demonstrating the effectiveness of combining wavelet-based multi-scale analysis with attention-based temporal modeling for financial time series prediction.