Comparative Analysis of Machine Learning and Time-Series Models for Monthly Rainfall Forecasting: A Case Study of Mumbai

Rainfall estimation for a given month plays a crucial role in water resource management, agricultural planning, and flood mitigation. However, accurately predicting rainfall remains challenging due to the nonlinear and dynamic nature of hydrometeorological processes. This study presents a comprehensive comparative analysis of machine learning, deep learning, and classical time-series models for monthly rainfall forecasting in Mumbai. The proposed framework integrates six machine learning regression models, including Linear Regression (LR), Random Forest (RF), Gradient Boosting (GB), Extreme Gradient Boosting (XGBoost), K-Nearest Neighbors (KNN), and Support Vector Regression (SVR), along with a deep learning model (Multilayer Perceptron (MLP)) and classical statistical approaches (Seasonal Autoregressive Integrated Moving Average (SARIMA)) and (Error, Trend, Seasonal (ETS)).The analysis is conducted using 21 years of monthly data (January 2000 to December 2020), with meteorological predictors such as Specific Humidity, Relative Humidity, and Temperature, while precipitation is considered as the target variable. To preserve temporal dependency and avoid data leakage, a chronological 80–20 train–test split is adopted, and consistent preprocessing is applied across all models. Model performance is evaluated using multiple error-based metrics, including Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and coefficient of determination (R²), along with hydrological performance indicators such as Nash–Sutcliffe Efficiency (NSE) and Kling–Gupta Efficiency (KGE). The results demonstrate that no single model consistently outperforms others across all evaluation criteria. ETS and SVR achieve the lowest error values in terms of RMSE, while MLP and SVR exhibit the highest hydrological efficiency (NSE and KGE). In contrast, XGBoost achieves the lowest MAPE, indicating superior relative error performance. Ensemble methods, particularly Voting and Stacking, demonstrate stable and competitive performance across multiple metrics but do not consistently dominate individual models. Statistical validation using paired t-tests and Wilcoxon signed-rank tests confirms that performance differences among models are statistically significant at the 5% level. Additionally, SHAP (Shapley Additive Explanations) analysis identifies Specific Humidity as the most influential predictor of rainfall, enhancing model interpretability. The findings highlight the importance of multi-metric evaluation and demonstrate that both machine learning and classical time-series models can provide reliable rainfall predictions. The study emphasizes that model selection should be guided by the specific evaluation criteria and application context, supporting a robust and flexible approach to hydrometeorological forecasting.