Global Chicken Price Forecasting Using Neural Network and Hybrid Approach

Abstract

Chicken is one of the fastest-growing and most widely consumed animal protein sources globally, making price stability important for producers, consumers, processors, retailers, and policymakers. However, feeding costs, disease outbreaks, seasonal demand, supply-chain disruptions, and market uncertainty contribute to nonlinear and volatile price movements. Although several previous poultry price studies have used traditional linear models such as ARIMA and SARIMA, there is limited evidence on the comparative performance of time-series, machine-learning, and hybrid forecasting models for global monthly chicken meat prices. Therefore, this study aimed to forecast global chicken prices by identifying the most accurate model using comparative forecast evaluation metrics. Monthly chicken price data, expressed in USD/kg, were obtained from the World Bank Commodity Price Data “Pink Sheet” for January 1960 to February 2026. The study fitted Auto-ARIMA, ETS, NNAR, Theta, STL+ETS, TBATS, equal-weight averaging, cross-validation error-weighted averaging, and 14 hybrid models. Forecast performance was evaluated using ME, RMSE, MAE, MPE, and MAPE. Results showed substantial historical price variation, ranging from USD 0.30/kg to USD 2.72/kg, with a mean of USD 1.16/kg. Among the 22 fitted models, NNAR performed best, producing the lowest RMSE, MAE, and MAPE values of 0.0189, 0.0122, and 1.014%, respectively. Among hybrid models, NNAR × STL was the most accurate, indicating the value of combining nonlinear learning with seasonal decomposition. The 30-month NNAR forecast suggested relatively stable prices with moderate fluctuations through August 2028. These findings support the use of nonlinear forecasting tools for poultry market planning, feed management, risk reduction, trade monitoring, and food affordability policy.