GPF-EVMoLE: An ETS-Driven Variable Selection and Mixture-of-Experts Framework for Multi-Step Garlic Price Forecasting

Predicting garlic prices is difficult because the crop behaves as both an agricultural commodity and a speculative asset. Unlike staple grains, which follow more predictable seasonal supply cycles, garlic can be stored for over a year, its production is geographically concentrated, and its demand remains inelastic. This industry structure makes it susceptible to speculative hoarding, where even minor harvest deficits may trigger sharp price spikes. A typical example is the “Suan Ni Hen” (crazy garlic) phenomenon in the Chinese market: during the 2009–2010 and 2016 periods, speculative capital repeatedly exploited expectations of harvest reduction to engage in large-scale hoarding. According to data released by China’s National Development and Reform Commission (NDRC) at the end of October 2016, national wholesale garlic prices surged by 90% year-on-year, with purchase prices in some major producing areas doubling or multiplying within a short period. Such short-term price bubbles, together with severe volatility and abrupt regime shifts, can make standard forecasting models unreliable in this uncertain environment. Existing methods, ranging from traditional seasonal algorithms to deep learning networks, often overlook the need to decouple the local trend-weekly-seasonal baseline from the dynamic effects of multi-source external signals. This paper proposes GPF-EVMoLE, a compositional multi-step forecasting framework built on an explicit division of labor. The framework first extracts an interpretable local trend and weekly-seasonal baseline through an ETS decomposition module. Two specialized components then process the residual signal: a temporal fusion Transformer-style variable selection network (VSN) uses multi-source external features to identify informative macroeconomic and environmental signals at each forecasting step, while a Mixture of Linear Experts (MoLE) models phase-wise regime shifts within the residual series. Together, these modules adaptively integrate heterogeneous information. This study evaluates the framework on a custom daily evaluation dataset containing 17,685 records across six major producing regions in three provinces. At 7-day and 14-day forecasting horizons, GPF-EVMoLE consistently outperforms eight representative statistical, machine learning, and deep learning baselines across MAE, RMSE, and MAPE metrics. Ablation studies verify the necessity of each component, showing that structural separation of the forecasting tasks helps overcome the limitations of monolithic models and provides an accurate and interpretable solution for complex agricultural markets.