Optimizing Rainfall Forecasting in Iraq: A Comparative Study of XGBoost, Nature-inspired Algorithms, and Tweedie Models

In fields such as agriculture and disaster management, it is essential to obtain accurate rainfall predictions. However, due to discontinuous precipitation and a large number of zero-valued data, modeling becomes difficult. In these complex situations, traditional statistical methods often cannot achieve highly accurate predictions. Using meteorological data from the Abu Ghraib station in Iraq (2011–2025), this article examines the performance of machine learning (ML) models against statistical frameworks. Therefore, we utilized an XGBoost model, nature-inspired optimization techniques (Whale Particle Optimization (WPO) and Hybrid Bio-inspired Optimization Algorithm (HBOA)) and the Tweedie model. The nature-inspired optimization techniques were utilized to refine the parameter tuning of the XGBoost algorithm separately. Specifically, it provides a comparative analysis of four models: XGBoost, WPO-enhanced XGBoost, HBOA-enhanced XGBoost, and Tweedie. The findings showed that the XGBoost model is clearly better than other models. Briefly, the XGBoost model achieved the highest coefficient of determination (R2=0.1483) and the lowest Root Mean Square Error (RMSE=3.3367). Although the WPO-enhanced XGBoost reduced the Mean Absolute Error (MAE) to 0.8026, it failed to outperform basic XGBoost. Additionally, the Tweedie model performed worst, with a negative R2 (-0.1116). Accordingly, the results demonstrated the superiority of ML models over statistical frameworks in dealing with complex situations (discontinuous features and a high frequency of zero-valued data). Nevertheless, further research is needed to enhance XGBoost with advanced methods, such as Bayesian optimization.