Performance Evaluation of Univariate and Climate Augmented Multivariate Models for Onion Price Forecasting in Punjab
Published by: Admin
Authors: Pradipkumar Adhale, Kamal Vatta and Priya Brata Bhoi
Abstract
Onion prices in India are among the most volatile of any agricultural commodity, with monthly wholesale fluctuations routinely exceeding 20% and repeatedly prompting government intervention. This study benchmarks thirteen forecasting models six univariate, six climate-augmented multivariate, and a hybrid ensemble on monthly Punjab onion prices from January 2004 to December 2025 (n = 264), with 2025 held out for testing. The price series exhibits a coefficient of variation of 66.2%, an annualised log-return standard deviation of 85.8%, and a maximum drawdown by 86.6%. All six multivariate models outperform every univariate model, including modern approaches such as TBATS and N-BEATS, yielding a 65.7% aggregate reduction in MAPE attributable to climate augmentation. XGBoost achieves the lowest MAPE of 13.20%, followed by Gradient Boosting (14.92%) and SARIMAX (16.09%). Granger causality analysis identifies rainfall as the dominant climate driver (p = 0.041). Residual diagnostics reveal persistent ARCH effects even after climate inclusion. The results establish that climate-augmented multivariate frameworks are not incrementally useful but structurally necessary for forecasting highly volatile perishable commodity prices.