Benchmarking autoregressive machine learning models for daily rainfall forecasting as a baseline for hydrological applications

Abstract

Abstract Accurate daily precipitation forecasting is a prerequisite for efficient reservoir operations, irrigation scheduling, and urban flood management, yet inherent predictability limits pose a significant challenge. This study quantifies these limits by evaluating a suite of seven machine learning models for one-day-ahead precipitation forecasting across 34 stations in New Jersey (NJ), USA (2015–2024). Using only antecedent precipitation and temporal features, models were rigorously assessed. Support Vector Regression (SVR) achieved the lowest average MAE (0.178 mm), a statistically significant advantage that did not translate into predictive skill, while Random Forest (RF) performed best across the highest number of individual stations. All seven models, including SVR, failed to explain daily rainfall variance (near-zero or negative R²) in this setting. This lack of predictive skill poses a significant risk for hydrological decision-making, such as flood control and reservoir regulation, where reliable forecasts are mandatory. This result establishes a regional performance benchmark, demonstrating that models reliant on historical data alone fundamentally struggle to capture the underlying atmospheric dynamics in this setting. Model performance showed significant spatiotemporal heterogeneity, with degraded accuracy during summer, tentatively attributed to the difficulty of predicting localized convective events. Within the scope of this study, these results indicate that for operational decision-support systems in this region, advancements likely require integrating dynamic atmospheric predictors. It provides a baseline for hydrologists, cautioning against the deployment of purely autoregressive machine learning models in critical water management scenarios without adequate validation of their skill.