<p>Reliable drought prediction is essential for effective water resource management, especially in drought-prone regions. Machine learning (ML) models have gained traction in recent years for predicting drought. However, standalone models often need help to capture drought indices’ complex, nonlinear behaviour despite yielding generally acceptable results. To address these challenges, signal decomposition algorithms have been increasingly applied as preprocessing tools in combination with standalone models to improve prediction performance. However, relying on a single decomposition method can limit the model’s accuracy. This study introduces a novel ‘integration-prediction’ model that integrates multiple decomposition algorithms for enhanced predictive capability. The present study employs 60 experimental model variations across four districts in South Bihar, namely Nalanda, Nawada, Jamui, and Gaya, using long-term monthly time series data from 1975 to 2020. Each district’s drought prediction utilized 15 different models: three standalone models (SVR, RF, XGBoost), nine hybrid decomposition models (EMD-SVR, EEMD-SVR, VMD-SVR, EMD-RF, EEMD-RF, VMD-RF, EMD-XGBoost, EEMD-XGBoost, VMD-XGBoost), and three integration models (INT-SVR, INT-RF, INT-XGBoost). The Standardized Precipitation Index (SPI) was analyzed at two-time scales—SPI-6 for short-term drought assessment and SPI-12 for long-term prediction. Model performance was evaluated using key metrics such as NSE, R², RMSE, and MAE. The results show that hybrid decomposition models significantly improved prediction accuracy compared to standalone models, and integration models provided even more significant enhancement. The VMD-SVR model consistently demonstrated superior performance across all districts, establishing it as the most reliable model in this study.</p>