Automated Hyperparameter Search for Stock Price Forecasting Using GA-Optimized 1D-CNN-LSTM and Technical Indicators

This study utilizes the capabilities of a 1D-CNN for local feature extraction and LSTM for long-term temporal dependency modeling to forecast the LQ45 stock price index. A Genetic Algorithm (GA) is implemented to optimize the 1D-CNN-LSTM hyperparameters, and the inputs are evaluated using various technical indicators including RSI, EMA, and MACD. For validation, the proposed GA-1D-CNN-LSTM model was evaluated against Hill Climbing (HC) optimization and a baseline 1D-CNN-LSTM under a fair time comparison framework. The results show that the combination of Close and RSI inputs provides the best performance, achieving an RMSE of 8.994 and a MAPE of 0.767%. Furthermore, the Wilcoxon Signed-Rank Test confirms that the proposed approach is statistically significant compared to all baseline and alternative models, with p-values well below 0.05. While baseline models may achieve competitive training accuracy, they are highly prone to overfitting. In contrast, the GA-optimized architecture offers superior generalization, demonstrating that a well-optimized model with carefully selected leading indicators can effectively outperform larger but less efficient architectures.