ABSTRACT Hybrid and ensemble machine learning models are increasingly used to predict monthly precipitation. This review synthesizes 85 studies (published between 2020 and 2025) using a unified R 2 -based framework. For each study, the hybrid model’s R 2 is contrasted with the best non-hybrid baseline, and results are summarized by hybrid class. Median R 2 is highest for parameter optimization and data preprocessing hybrids, followed by deep hybrids; component combination and post-processing configurations show more variable performance. A Kruskal–Wallis test on R 2 across classes indicates no statistically significant differences at α = 0.05 ( p = 0.63 ). The review clarifies complementarity with physics-based NWP (downscaling and post-processing), highlights recurrent pitfalls (leakage, representativeness, non-stationarity), and provides practical guidance for R 2 sound evaluation and reporting. While most reviewed studies operate at the station or grid-cell level, the scarcity of architectures that explicitly model spatial dependencies for monthly horizons is identified as a key gap.
Pérez et al. (Mon,) studied this question.