Testing set error.

This study constructs a multi-stage hybrid forecasting model using hog price time series data and its influencing factors to improve prediction accuracy. First, seven benchmark models including Prophet, ARIMA, and LSTM were applied to raw price series, where results demonstrated that deep learning models significantly outperformed traditional methods. Subsequently, STL decomposition decoupled the series into trend, seasonal, and residual components for component-specific modeling, achieving a 22.6% reduction in average MAE compared to raw data modeling. Further integration of Spearman correlation analysis and PCA dimensionality reduction created multidimensional feature sets, revealing substantial accuracy improvements: The BiLSTM model achieved an 83.6% cumulative MAE reduction from 1.65 (raw data) to 0.27 (STL-PCA), while traditional models like Prophet showed an 82.2% MAE decrease after feature engineering optimization. Finally, the Beluga Whale Optimization (BWO)-tuned STL-PCA-BWO-BiLSTM hybrid model delivered optimal performance on test sets (RMSE = 0.22, MAE = 0.16, MAPE = 0.99%, ), exhibiting 40.7% higher accuracy than unoptimized BiLSTM (MAE = 0.27). The research demonstrates that the synergy of temporal decomposition, feature dimensionality reduction, and intelligent optimization reduces hog price prediction errors by over 80%, with STL-PCA feature engineering contributing 67.4% of the improvement. This work establishes an innovative “decomposition-reconstruction-optimization” framework for agricultural economic time series forecasting.