Timber price forecasting in Vietnam using graph neural network and LSTM method

Abstract

Timber price forecasting plays a crucial role in the strategic planning of the forestry sector, particularly in countries like Vietnam, where the market is highly volatile and influenced by complex factors such as supply-demand imbalances, environmental policies, and globalization. However, traditional forecasting models such as ARIMA, SARIMA, and VAR often fail to provide accurate predictions in non-linear and multi-dimensional market conditions. This study aims to develop a more effective forecasting approach by integrating Graph Neural Networks (GNNs) and Long Short-Term Memory (LSTM) models. The GNN component captures spatial dependencies among timber-producing regions, while the LSTM component models long-term temporal trends. Together, they address the limitations of conventional models and provide a comprehensive framework for price forecasting. Using data from 2000 to 2024 – including timber prices, harvesting volumes, climate data, and export records – the proposed GNN-LSTM model demonstrated superior predictive performance. It achieved a Mean Squared Error (MSE) of 0.0023, Root Mean Squared Error (RMSE) of 0.0480, and Mean Absolute Error (MAE) of 0.0340, along with an R² score of 0.94. These results confirm the model’s capability to capture complex spatial-temporal relationships and improve forecasting accuracy. The findings suggest that the GNN-LSTM hybrid model provides a valuable decision-support tool for stakeholders in Vietnam’s forestry industry, including forest managers, investors, and policy makers. It enables more informed decision-making, risk mitigation, and resource optimization. Additionally, the study highlights the potential of applying advanced machine learning techniques to data-constrained markets. Future research should focus on enhancing data availability and incorporating mechanisms to better capture the effects of external disruptions, such as global supply chain shocks.