<![CDATA[Cooking oil is a critical staple commodity in Indonesia, where price fluctuations significantly impact household purchasing power and regional economic stability, especially in East Java. These fluctuations stem from complex, nonlinear interactions between crude palm oil prices, supply chain conditions, and market mechanisms. This study fills the gap in existing forecasting models, which often fail to address regional price dynamics and lack interpretability. We develop a short-term forecasting model using the Temporal Fusion Transformer (TFT), a deep learning architecture tailored for multi-horizon time series forecasting, to predict packaged and bulk cooking oil prices in East Java. Daily price data for packaged and bulk cooking oil, along with national palm oil prices, were sourced from official government records covering April 21, 2022, to October 2, 2024. The dataset was preprocessed with missing value interpolation, normalization, and transformation into a supervised multivariate time series format. The TFT model was trained using a 30-day historical window, with a seven-day forecasting horizon, optimized via quantile loss to generate probabilistic forecasts. Model performance was assessed using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and quantile loss. Results show that the TFT model achieves high accuracy, low validation errors, and provides reliable uncertainty estimates. Short-term forecasts suggest stable price trends, with greater uncertainty for packaged cooking oil than bulk. This research demonstrates the TFT's potential for short-term forecasting, policy support, and its broader application to regional price monitoring.]]>
