73. Award Talk: Machine Learning for Economic Decision Making in Texas Stocker Cattle Operations.

Abstract Stocker operations occupy a pivotal position in the U.S. beef value chain, purchasing weaned calves and developing them on pasture to feeder-ready weights prior to feedlot placement. In Texas, which accounts for approximately 14% of total U.S. cattle and calves, stocker enterprises simultaneously absorb price risk on both the buy and sell side of every production cycle. The U.S. cattle industry contributed approximately $101.1 billion to total U.S. agricultural receipts in 2023, with feeder and stocker transactions representing a core share of that value. Auction prices for stocker and feeder calves at benchmark Southern Plains markets have exhibited dramatic cyclical volatility over the past two decades, ranging from a 25-year low near $103/cwt in 2003 to an all-time record exceeding $462/cwt in late 2025, a nearly 4.5-fold swing driven by interacting cycles of drought-induced herd liquidation, corn and fuel cost shocks, and shifting domestic and export beef demand. Yet buy-sell decisions across stocker enterprises remain largely driven by market intuition rather than data-driven price forecasts. This study addresses that gap by developing and benchmarking a suite of machine learning and time-series forecasting models for stocker calf auction prices in Texas, with explicit integration of both supply-side cost drivers and demand-side market signals. Six forecasting models were evaluated: three time-series models (ARIMA, SARIMA, SARIMAX) and three multivariate machine learning models (RF, AdaBoost, SVR). The target variable was the weekly average stocker calf auction price ($/Cwt) obtained from USDA AMS auction reports spanning 2017-2026. Six exogenous predictors were compiled across two categories. Demand-side proxies included feeder cattle futures prices and the beef retail/wholesale price spread, reflecting buyer willingness-to-pay and downstream consumer demand pull on stocker prices. Supply-side and macroeconomic drivers included corn prices, natural gas prices, the U.S.-Mexico exchange rate, the Consumer Price Index. All series underwent preprocessing including stationarity testing, seasonal decomposition, and feature engineering prior to model fitting. All models were evaluated under a walk-forward validation framework, this approach replicates real-world deployment conditions and preserves temporal ordering. Model performance was assessed using MAE, RMSE, and R2. Walk-forward validation results demonstrated that machine learning models outperformed time-series approaches across all error metrics. Among time-series models, SARIMAX yielded the strongest performance (R2=0.84), with the inclusion of demand-side predictors improving substantially over the SARIMA and ARIMA baselines. AdaBoost achieved the highest overall accuracy (R = 0.93), outperforming RF and SVR, suggesting that the nonlinear interactions among demand-side and supply-cost predictors exceed the capacity of linear time-series structures to capture. Overall, these results underscore the value of integrating demand-side market signals with supply-cost drivers under a walk-forward framework for stocker cattle price risk management in Texas. Future research should explore weight-class segmentation across Texas production zones and incorporation of drought indices as additional supply-demand proxies. For image description, please refer to the figure legend and surrounding text.