A temporal fusion transformer for multi-step forecasting of Indonesia’s strategic food commodity prices

• We forecast the average daily prices of ten strategic Indonesian commodities. • We propose the application of Temporal Fusion Transformer (TFT), and compare its performance with Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), two types of recurrent neural networks. • As a baseline, we employ Naïve and ARIMA, a classical approach in time-series models. • The proposed model can aid policymakers in developing effective food distribution and pricing strategies. Accurate forecasting of food commodity prices is essential for mitigating economic instability and ensuring national food security. While deep learning has advanced time series forecasting, this study employs the Temporal Fusion Transformer (TFT) architecture for multi-step (30-day) forecasting of daily prices across ten strategic food commodities in Indonesia. The TFT’s performance is compared against benchmark models Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), ARIMA, and Naïve models. The results demonstrate TFT’s superior performance, particularly for commodities with high price variability. For these, the Mean Absolute Error (MAE) of the best-performing benchmark was higher than TFT’s by 285.6% for bird’s eye chili, 65.7% for shallots, 52.8% for red chili, 36.2% for chicken meat, and 29.9% for beef. This work validates the TFT as a powerful tool for developing data-driven policy and more effective price management in Indonesia.