Evaluating impact-based forecasting models for tropical cyclone anticipatory action

Impact‐based Forecast (IBF) is increasingly adopted for Anticipatory Action in disaster risk management, yet systematic comparison of the diverse models in use remains limited. To address this gap, we evaluated two existing IBF models that were developed by humanitarian agencies for tropical cyclones in the Philippines and Bangladesh: a statistical machine learning model and an elementary damage curve model. These represent contrasting approaches, systematically characterised with a model card framework across indicators, including data, hazard-impact thresholds, modelling, and decision rules. We used Typhoon Kammuri (2019, the Philippines) as a case study. Both models showed low event-specific accuracy. The statistical model triggered action 81 h before landfall, detecting only 3 % of affected municipalities with a 75 % False Alarm Ratio (FAR). The elementary model adapted for the Philippines context would have triggered 72 h before landfall with a 17 % Probability of Detection (POD) and 40 % FAR. The tropical cyclone forecast uncertainty, particularly high for Kammuri, propagated into the IBF in terms of location and timing. The study also illustrated the influence of parameters such as lead time, trigger threshold, and forecast uncertainty buffer on the model performances, through an interactive portal showcasing the usefulness of such tools in understanding the interplay between indicators. These findings underscore the need for transparent, interpretable IBF frameworks that explicitly communicate uncertainties and trade-offs. The choice between complex and simpler models should be tailored to local data and operational requirements, rather than assuming one approach is generally superior. It requires a statistical analysis of many events.