Abstract
Price volatility in agricultural markets is influenced by seasonality, supply-demand fluctuations, policy changes, and climate. These factors significantly impact agricultural production and the broader macroeconomy. Traditional time series models, limited by linear assumptions, often fail to capture the nonlinear nature of price fluctuations. To address this limitation, we propose an integrated forecasting model that combines TCN and XGBoost to improve the accuracy of agricultural price volatility predictions. TCN captures both short-term and long-term dependencies using convolutional operations, while XGBoost enhances its ability to model nonlinear relationships. The model uses 65,750 historical data points from rice, wheat, and corn, with a sliding window technique to construct time series features. Experimental results demonstrate that the TCN-XGBoost model significantly outperforms traditional models such as ARIMA (RMSE = 0.36, MAPE = 8.9%) and LSTM (RMSE = 0.34, MAPE = 8.1%). It also outperforms other hybrid models, such as Transformer-XGBoost (RMSE = 0.23) and CNN-XGBoost (RMSE = 0.29). Specifically, the TCN-XGBoost model achieves an RMSE of 0.26 and a MAPE of 5.3%, underscoring its superior performance. Moreover, the model shows robust performance across various market conditions, particularly during significant price fluctuations. During dramatic price movements, the RMSE is 0.28 and the MAPE is 6.1%, effectively capturing both trends and magnitudes of price changes. By leveraging TCN's strength in temporal feature extraction and XGBoost's capability to model complex nonlinear relationships, the TCN-XGBoost integrated model offers an efficient and robust solution for forecasting agricultural prices. This model has broad applicability, particularly in agricultural market decision-making and risk management.