A Comprehensive Analysis of Resilient Multivariate Forecasting Models for Steel Plate Price Prediction

The global steel industry, holding paramount economic significance, is characterized by the inherent volatility of steel prices. Leveraging the reliable weekly steel plate price data from the Commodity Research Unit (CRU), this research employs sophisticated machine learning algorithms to forecast plate prices. The dataset spans from July 27, 2011, to July 5, 2023, encompassing six key predictive factors. Notably, total inventory levels exhibit the highest correlation (0.88) with plate prices, with the finished goods inventory value of heavy machinery emerging as the most influential factor. A comprehensive training regimen is undertaken for machine learning models, incorporating Prophet, XGBoost, LSTM, and GRU. Time Series Cross-Validation is implemented to maintain the temporal order of the data, and a Bayesian optimization function is employed for hyperparameter tuning. XGBoost emerges as the top-performing model, yielding the lowest Mean Squared Error (MSE) of 332.25 and Mean Absolute Error (MAE) of 14.55. Demonstrating superior predictive accuracy, with a Mean Absolute Percentage Error (MAPE) of 0.94% and a Root Mean Squared Error (RMSE) score of 18.06, XGBoost establishes itself as the most effective model in steel plate price forecasting. This outcome underscores the efficacy of advanced machine learning methodologies in navigating the complexities of steel market dynamics for enhanced predictive insights.