LSTM-Based Forecasting Model for Carrot Price Prediction Across Indonesian Provinces

Abstract

Price volatility in food markets poses a significant challenge for Indonesia due to its impact on food security, inflation stability, and economic well-being. Carrots, as a horticultural commodity, are particularly vulnerable to price fluctuations because of their perishable nature, dependence on seasonal harvests, and sensitivity to climatic variations. Accurate forecasting is therefore important to support agricultural price monitoring and supply chain planning. However, conventional statistical approaches often struggle to capture the non-linear patterns and long-term temporal dependencies inherent in volatile commodity markets. This study develops and evaluates a Long Short-Term Memory (LSTM) model for forecasting monthly carrot prices across 28 provinces in Indonesia using a single pooled modelling framework trained on combined provincial price data. Historical price data were preprocessed through data cleaning, interpolation, normalization, and sequence generation using a 12-month sliding window. A sequential stacked LSTM architecture with dropout regularization was implemented to improve model generalization. The evaluation results produced a Mean Absolute Error (MAE) of Rp 1,376.90, a Root Mean Square Error (RMSE) of Rp 1,742.74, and a Mean Absolute Percentage Error (MAPE) of 17.76%, indicating good predictive performance for a highly volatile commodity market. Nevertheless, provincial-level analysis revealed substantial variation in forecasting accuracy, with several provinces achieving excellent results while Bengkulu, Banten, and South Sulawesi exhibited considerably higher prediction errors. These findings indicate that the proposed framework has potential for agricultural price forecasting and monitoring; however, further validation involving comparative baseline models, additional external variables, and region-specific modelling approaches is necessary before practical implementation.