Rainfall is an important climate variable with high variability in tropical regions such as Indonesia, thus requiring accurate forecasting methods. This study aims to evaluate the performance of a machine learning-based hybrid model, analyse the effectiveness of bias correction on reanalysis data (ERA5 and NASA POWER), and assess the model’s ability to represent variability and extreme events. The data used consists of an integration of BMKG observations and reanalysis data for the 2022–2025 period in the Yogyakarta region. The methods employed include a two-stage modelling approach (classification and regression) using the Random Forest and Gradient Boosting algorithms, with evaluation based on RMSE, MAE, and R². The results show a bimodal rainfall pattern, with peaks in February (10.60 mm/day) and November (11.72 mm/day), and a dry period from June to September, with a minimum in September (0.59 mm/day). Rainfall anomalies indicate significant annual variability, with negative values in 2022–2023 and positive values in 2024–2025. NASA POWER data tends to be overestimated, particularly in September, with a difference of 11.62 mm/day. The hybrid model reduced bias and improved data fit (RMSE 12.01 mm; MAE 6.85 mm), but yielded a negative R² value, indicating limitations in representing variability and extreme events.