Decomposition-reconstruction-optimization framework for hog price forecasting: Integrating STL, PCA, and...

Area of Science:

• Agricultural Economics
• Time Series Analysis
• Machine Learning

Background:

• Accurate hog price forecasting is crucial for agricultural economic stability.
• Traditional time series models often struggle with the complexity of agricultural market data.
• Deep learning models show promise but require further optimization for practical application.

Purpose of the Study:

• To develop a multi-stage hybrid forecasting model for hog price time series.
• To enhance prediction accuracy by integrating temporal decomposition, feature engineering, and intelligent optimization.
• To establish an innovative

Main Methods:

• Applied seven benchmark models (Prophet, ARIMA, LSTM) to raw hog price data.
• Utilized Seasonal-Trend decomposition using Loess (STL) for series decomposition.
• Implemented Principal Component Analysis (PCA) for dimensionality reduction and Spearman correlation for feature selection.
• Developed a hybrid model incorporating BiLSTM and optimized using Beluga Whale Optimization (BWO).

Main Results:

• Deep learning models outperformed traditional methods on raw data.
• STL decomposition reduced Mean Absolute Error (MAE) by 22.6%.
• Feature engineering (STL-PCA) reduced BiLSTM's MAE by 83.6% (from 1.65 to 0.27).
• The optimized STL-PCA-BWO-BiLSTM model achieved superior performance (RMSE=0.22, MAE=0.16, MAPE=0.99%).

Conclusions:

• The proposed hybrid model significantly reduces hog price prediction errors (over 80% reduction).
• STL-PCA feature engineering is a key contributor (67.4% of improvement).
• The "decomposition-reconstruction-optimization" framework offers a robust approach for agricultural economic time series forecasting.