A Hybrid LSTM–Stacking–SMOTE Model for Weather-Aware Palm Oil Price Prediction Addressing Data Imbalance and Forecast Accuracy

Abstract

Accurate forecasting of palm oil prices is crucial for agribusiness decision-making due to high market volatility influenced by dynamic weather conditions. This study proposes a novel hybrid deep learning model combining Long Short-Term Memory (LSTM), Stacking Ensemble, and Synthetic Minority Over-sampling Technique (SMOTE) to improve predictive accuracy and handle class imbalance in price trend classification. The model was trained using a multivariate time-series dataset sourced from Kaggle, consisting of daily records of temperature, humidity, rainfall, and palm oil prices. A binary classification scheme was applied by labeling instances as either price increase (class 1) or price stable/decrease (class 0), based on a 0% price change threshold. Four experimental configurations were evaluated: standard LSTM, LSTM + SMOTE, LSTM + Stacking, and the proposed LSTM + SMOTE + Stacking. The proposed model outperformed all baselines, achieving the highest accuracy of 83.12%, an F1-score of 0.8466, MAE of 0.1688, RMSE of 0.4109, and a perfect recall of 1.0000, indicating excellent sensitivity to minority class trends. In contrast, the standard LSTM achieved only 77.32% accuracy and an F1-score of 0.7224, showing limited ability in handling imbalanced data. Visualization of loss curves and confusion matrices confirmed the model’s learning stability and classification effectiveness. This study contributes a novel integration of ensemble learning and oversampling in time-series commodity forecasting and demonstrates the effectiveness of this approach in capturing weather-driven price patterns, offering a robust framework for predictive analytics in agriculture.