The Promise of Time-Series Foundation Models for Agricultural Forecasting: Evidence from Marketing Year Average Prices
Le Wang, Boyuan Zhang
TL;DR
This paper evaluates 17 forecasting approaches across four model classes on USDA ERS data (1997–2025) to forecast monthly prices and the policy-relevant Marketing Year Average (MYA). It finds that zero-shot time-series foundation models, particularly Time-MoE, outperform traditional econometric, machine learning, and deep learning models trained from scratch, achieving about 45% MAE improvement overall and >50% for corn and soybeans. The results challenge the prior view that simpler models suffice in agricultural forecasting and demonstrate that large pre-trained TSFMs can generalize well in data-scarce domains. The study discusses policy implications for Farm Bill programs and highlights why foundation models may offer a scalable path for high-stakes economic forecasting, with caveats about explainability and data leakage concerns. $MYA = \sum_{t=1}^{12} P_t \times w_t$ is used to illustrate the MYA aggregation, linking monthly forecasts to policy-relevant outcomes.
Abstract
Forecasting agricultural markets remains a core challenge in business analytics, where nonlinear dynamics, structural breaks, and sparse data have historically limited the gains from increasingly complex econometric and machine learning models. As a result, a long-standing belief in the literature is that simple time-series methods often outperform more advanced alternatives. This paper provides the first systematic evidence that this belief no longer holds in the modern era of time-series foundation models (TSFMs). Using USDA ERS data from 1997-2025, we evaluate 17 forecasting approaches across four model classes, assessing monthly forecasting performance and benchmarking against Market Year Average (MYA) price predictions. This period spans multiple agricultural cycles, major policy changes, and major market disruptions, with substantial cross-commodity price volatility. Focusing on five state-of-the-art TSFMs, we show that zero-shot foundation models (with only historical prices and without any additional covariates) consistently outperform traditional time-series methods, machine learning models, and deep learning architectures trained from scratch. Among them, Time-MoE delivers the largest accuracy gains, improving forecasts by 45% (MAE) overall and by more than 50% for corn and soybeans relative to USDA benchmarks. These results point to a paradigm shift in agricultural forecasting: while earlier generations of advanced models struggled to surpass simple benchmarks, modern pre-trained foundation models achieve substantial and robust improvements, offering a scalable and powerful new framework for highstakes predictive analytics.