Optimized Hybrid CNN-LSTM Model for Predicting Transportation Sector Stock Prices Using Optimizer and Activation Function Tuning

Abstract

Stock price prediction is inherently complex due to nonlinear dynamics and high volatility, particularly in Indonesia’s transportation sector, which experienced significant inflationary pressure and extreme instability during and after the COVID-19 pandemic. These disruptions introduced structural breaks and regime shifts, intensifying non-stationary market behavior and increasing forecasting uncertainty. Such conditions create an urgent need for robust predictive information systems capable of supporting investment decision-making and risk management in highly volatile environments. However, standalone recurrent models such as Long Short-Term Memory (LSTM) often struggle to capture local micro-patterns and long-term dependencies. Moreover, prior studies have rarely implemented systematic hyperparameter optimization, resulting in inconsistent predictive performance across stocks with heterogeneous volatility. In contrast, Convolutional Neural Networks (CNN) extract local patterns and short-term nonlinear features, making them effective for modeling high-frequency fluctuations. This study proposes a systematically optimized hybrid CNN-LSTM model to forecast transportation sector stock prices using daily OHLC data from 2020-2025. The research framework follows the Cross Industry Standard Process for Data Mining (CRISP-DM) methodology, encompassing business understanding, data understanding, data preparation, modeling, evaluation, and deployment. Prior to modeling, preprocessing includes data cleaning, Min-Max normalization, and sliding window transformation to construct supervised learning sequences. CNN is employed to extract localized nonlinear features and reduce noise, while LSTM models long-term temporal dependencies. Model performance is evaluated using MAE, MSE, RMSE, MAPE, and R². Results show that the optimized CNN-LSTM model outperforms the baseline across all stocks. The highest R² of 0.9725 is obtained from one stock, indicating strong performance. In addition, the average R² improves from 0.8736 to 0.9483, an increase of 0.0747 (8.55%). The best results are achieved using ReLU with Adam and Nadam optimizers, demonstrating improved convergence and generalization. These findings highlight the effectiveness of optimized hybrid deep learning models for forecasting in nonlinear and non-stationary financial markets.