Explainable Machine Learning for Wood Market Forecasting Using News Analytics and Temporal Causality

Abstract

The ability to forecast accurate dynamics of the wood market is vital for the management of the supply chain, production planning and sustainable forestry decision making. Most of the conventional techniques for forecasting are based on numerical data from the past and do not take into account the predictive capabilities of unstructured text data available from online news sources. In this research, an explainable framework based on news data and machine learning is proposed to predict wood sales on weekly basis, which incorporates natural language processing, Uniform Manifold Approximation and Projection (UMAP), Granger causality analysis, LightGBM and SHapley Additive exPlanations (SHAP) algorithms. News articles related to wood were gathered and preprocessed to create the semantic embeddings, followed by UMAP dimensionality reduction and clustering into semantic clusters. Using granger causality analysis, statistically significant news derived features were identified and then used in conjunction with historical sales variables to train forecasting models: Random Forest and LightGBM. Based on the experimental results, the proposed LightGBM model had an MSE of 6.54, RMSE of 2.56, MAE of 1.50, and an R² value of 0.99, which demonstrated the model showed the excellent predictive accuracy and generalization ability. Moreover, the SHAP-based explainability showed the most influential historical and semantic features that influenced forecasting performance, offering more transparency and interpretation of the decisions to be made. Overall, the proposed framework shows the potential of combining semantic news analytics with explainable machine learning to improve the accuracy and interpretability of wood market forecasting, thereby providing a valuable tool for researchers, policymakers, and stakeholders in the forestry and wood products sector.