Deep Learning for Commodity Price Forecasting: LSTM, BiLSTM, GRU Comparison with Variable Window Sizes and Activation Functions

The purchasing power of a significant portion of Myanmar's population is directly impacted by commodity prices, making it critical to maintain price stability. Developing a robust prediction model is one approach to achieve this stability. This research focuses on forecasting the daily prices of salt, oil, and rice from 2007 to 2020. Advanced recurrent neural network models, including Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Bidirectional Long Short-Term Memory (BiLSTM), are utilized for this purpose. These models learn from observed data over time to predict future values in sequential data. The impact of different window sizes and activation functions on model performance, specifically comparing ReLU and sigmoid activations are investigated. The models are evaluated using key performance metrics: Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and R-squared (coefficient of determination). The proposed approach indicates that the GRU model with a sigmoid activation function consistently outperforms the LSTM and BiLSTM models, achieving the lowest RMSE and highest R-squared values, particularly with a window size of 30. While the BiLSTM model demonstrates strong performance with smaller window sizes and sigmoid activation, the LSTM model's performance is less consistent across different configurations. This experiment provides valuable insights for researchers and practitioners aiming to enhance commodity price prediction using advanced neural network models.