The importance of data noise reduction–wavelet transformation in stock price forecasting with Bidirectional Long Short-Term Memory network

With the rapid development of implemented artificial intelligence, various deep learning models are proposed. Unlike conventional models, Bi-directional Long-Short-Term Memory (BiLSTM)’s architecture allows for the capture of both forward and backward dependencies in data, making it particularly effective in detecting trends and seasonality. This study explores the application of BiLSTM networks for stock price forecasting, focusing on multivariate time-series data. In this work, a novel approach integrates Wavelet transformation with BiLSTM to decompose stock price data into trend and seasonal components, improving model interpretability and forecast accuracy. The model is further enhanced by incorporating daily stock price deviations as an additional feature during training. The effectiveness of this approach is demonstrated by comparing the performance of BiLSTM with other deep learning models, including LSTM, Gated Recurrent Unit, and Recurrent Neural Networks, across two stock datasets. Experimental results show substantial gains: up to 250× reduction in Mean Squared Error (MSE), 20× in Mean Absolute Error (MAE), and 15× in Mean Absolute Percentage Error (MAPE) after applying the wavelet transform; relative to LSTM, the proposed model achieves 4× lower MSE, 2× lower MAE, and 2.5× lower MAPE, validating the robustness of the proposed method for capturing both trend and seasonality in financial time-series data. The results verify the importance of applications of artificial intelligence.