<![CDATA[This study analysed fifteen years of monthly rainfall and flood records from Java’s six provinces to determine when disaster managers should act. We measured how long periods of elevated rainfall risk tend to persist and used machine learning to forecast how many flood events to expect each month. The results show that rainfall volatility in Java is persistent: a volatility shock loses half its effect over approximately 7.4 months. In practical terms, the October–November wet season onset is an important window for disaster preparedness: agencies that assess rainfall conditions during these two months can use that assessment to gauge whether the coming season may warrant high-alert mobilisation before the worst floods arrive, though this study does not test that window against other possible timings. Monthly flood forecasting performed reasonably well across all models, but adding the volatility estimate to the prediction models did not significantly improve forecast accuracy. This happens because the GARCH-derived variance is structurally collinear with the recent rainfall lags already used as predictors, so it adds little once recent rainfall is already represented in the model. This clarifies the roles of the two tools. Volatility monitoring answers when to prepare, while machine learning forecasts answer how many events to expect. We recommend Indonesian disaster management agencies adopt a two-step framework—assess rainfall volatility in October–November to set mobilisation levels, then update monthly flood forecasts throughout the wet season to guide operational response.]]>
