A noticeable increase in drought frequency and severity has been observed across the globe due to climate change, which attracted scientists in development of drought prediction models for mitigation of impacts. Droughts are usually monitored using drought indices (DIs), most of which are probabilistic and therefore, highly stochastic and non-linear. The current research investigated the capability of different versions of relatively well-explored machine learning (ML) models including random forest (RF), minimum probability machine regression (MPMR), M5 Tree (M5tree), extreme learning machine (ELM) and online sequential-ELM (OSELM) in predicting the most widely used DI known as standardized precipitation index (SPI) at multiple month horizons (i.e., 1, 3, 6 and 12). Models were developed using monthly rainfall data for the period of 1949-2013 at four meteorological stations namely, Barisal, Bogra, Faridpur and Mymensingh, each representing a geographical region of Bangladesh which frequently experiences droughts. The model inputs were decided based on correlation statistics and the prediction capability was evaluated using several statistical metrics including mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), correlation coefficient (R), Willmott's Index of agreement (WI), Nash Sutcliffe efficiency (NSE), and Legates and McCabe Index (LM). The results revealed that the proposed models are reliable and robust in predicting droughts in the region. Comparison of the models revealed ELM as the best model in forecasting droughts with minimal RMSE in the range of 0.07-0.85, 0.08-0.76, 0.062-0.80 and 0.042-0.605 for Barisal, Bogra, Faridpur and Mymensingh, respectively for all the SPI scales except one-month SPI for which the RF showed the best performance with minimal RMSE of 0.57, 0.45, 0.59 and 0.42, respectively.