Stock market prediction using IALSTM-BIGRU architecture and improved chi-square based selective feature set

Stock market prediction is challenging due to complex temporal dependencies and nonlinear variations in stock price data. This work proposes an IA-LSTM-BiGRU-based Stock Market Prediction (ILBSMP) approach comprising preprocessing, feature extraction, feature selection, and hybrid prediction. DMI, EMA, MACD, RSI, SMA, WWS, and DEMA features are extracted. An improved Chi-square feature selection method integrated with Shannon entropy selects informative features based on statistical dependency and information uncertainty. A novel Improved Attention-based LSTM (IA-LSTM) and Bidirectional GRU (Bi-GRU) hybrid model captures temporal dependencies in both directions. The IA-LSTM incorporates attention, clipping, and Leaky Bipolar activation mechanisms to improve feature learning and gradient stability. An improved Lehmer mean-based score-level fusion with tanh normalization combines prediction scores. Experimental results achieve MSE of 0.517, MAE of 0.583, MSLE of 0.005, MAPE of 0.045, and RMSE of 0.719, demonstrating the effectiveness of the proposed approach.