A Hybrid ARIMA–LSTM Model Optimized with a Tree-Structured Parzen Estimator for Automatic Hyperparameter Tuning to Forecast Indonesia's Chili and Climate Data

Red chili peppers are a strategic agricultural commodity in Indonesia, yet their prices and production are highly volatile due to climate variability and supply chain disruptions. Accurate forecasting is therefore essential to support agricultural planning and price stabilization policies. This study proposes an automated forecasting framework based on a hybrid Autoregressive Integrated Moving Average–Long Short-Term Memory (ARIMA–LSTM) computational modeling approach that integrates statistical and deep learning methods while incorporating weather variables, namely rainfall and sunshine duration. The main novelties of this research are the application of Tree-Structured Parzen Estimator (TPE) for efficient automated hyperparameter optimization, improving upon conventional Grid Search (GS), and the construction of a hybrid model capable of capturing both linear and nonlinear temporal patterns. The framework is evaluated using time-series data from North Sumatra, Central Java, and Bali covering the period from March 2021 to December 2023. The results show that the proposed hybrid ARIMA–LSTM model significantly outperforms standalone and benchmark models. For retail price forecasting, the hybrid model achieves a Mean Absolute Percentage Error (MAPE) of 1.45%, compared to 42.74% for ARIMA, 32.41% for Seasonal ARIMA (SARIMA), and 31.73% for Temporal Convolutional Network (TCN), indicating a substantial reduction in forecasting error. These findings confirm the effectiveness of combining TPE-based optimization with hybrid modeling in capturing complex agricultural and climatic dynamics. Overall, the proposed framework provides a scalable and data-driven forecasting tool to support farmers in production planning and assist policymakers in designing more effective price stabilization and agricultural resource management strategies under climate variability.