Machine Learning Approaches for Predicting Seasonal Stock Trends

Abstract

The financial market is vital for economic growth yet it often experiences volatility, particularly in Indonesia’s transportation sector where stock prices are strongly affected by seasonal fluctuations. Conventional forecasting methods often neglect these recurring patterns, lowering predictive accuracy. This study assesses the capability of Machine Learning algorithms to capture seasonality in stock price prediction, using PT Garuda Indonesia (Persero) Tbk (GIAA.JK)’s monthly data from August 2019 to May 2025, retrieved from Yahoo Finance. Four models–Linear Regression, Extreme Gradient Boosting (XGBoost), Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM)–were trained and tested, with performance evaluated using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). Hyperparameter tuning was applied to XGBoost, LSTM, and GRU, while statistical validation employed the Kruskal-Wallis test. Results showed that the tuned GRU outperformed other models, achieving MAE of 5.90, RMSE of 7.33, and MAPE of 9.67%, demonstrating ‘excellent’ accuracy in modelling both short-term and seasonal dynamics. These findings highlight the superiority of GRU in modelling both short-term fluctuations and long-term seasonal dependencies in stock prices. The results contribute practical insights for investors and emphasize the importance of integrating seasonality in predictive models for volatile sectors