Accurate prediction of monsoon rainfall remains a persistent challenge due to the intricate interplay among meteorological conditions, oceanic influences, and broader climatic patterns. While significant effort has been devoted to improving model architectures, the effect of input feature composition on prediction accuracy has received relatively less attention. This study addresses that gap by conducting an extensive empirical evaluation of 2,047 feature group combinations, systematically derived from 11 curated sets of climate-related variables. Using a hyperparameter-tuned XGBoost model, each configuration was evaluated independently to assess the predictive contribution of domains such as lagged climate indices, cyclical temporal encodings, and event-based indicators. The results show that model performance improved significantly from an R2 of 0.5801 (RMSE: 10.8999 mm, MAE: 4.9009 mm) using only meteorological features to an R2 of 0.7606 (RMSE: 8.2297 mm, MAE: 3.8752 mm) when combined with oceanic and climatic inputs, particularly lagged MJO and ENSO indices and temporal signals. These insights reinforce the value of domain-informed feature fusion and provide a replicable approach to enhancing monsoon prediction models through thoughtful feature group design and empirical validation.