Abstract. Tropical cyclones (TCs) produce strong winds and heavy rains accompanied by consecutive events such as landslides and storm surges, resulting in losses of lives and livelihoods particularly in regions where socioeconomic vulnerability is high. To proactively mitigate the impacts of TCs, humanitarian actors implement anticipatory action. In this work, we build upon such an existing anticipatory action for the Philippines, which uses an impact-based forecasting model for housing damage based on XGBoost to release funding and trigger early action. We improve it in three ways. First, we perform a correlation and selection analysis, to understand if Philippines-specific features can be left out or replaced with features from open global data sources. Secondly, we transform the target variable (percentage of completely damaged houses) and not yet grid-based global features to a 0.1 degrees grid resolution by de-aggregation using Google Building Footprint data. Thirdly, we evaluate XGBoost regression models using different combinations of global and local features at both grid and municipality spatial level. We introduce a two-stage model to first predict if the damage is above 10 % and then use a regression model trained on either all or on only high damage data. All experiments use data from 39 typhoons that impacted the Philippines between 2006–2020. Due to the scarcity and skewness of the training data, specific attention is paid to data stratification, sampling and validation techniques. We demonstrate that employing only the global features does not significantly influence model performance. Despite excluding local data on physical vulnerability and storm surge susceptibility, the two-stage model improves upon the municipality-based model with local features. When applied for anticipatory action our two-stage model would show a higher True Positive rate, a lower False Negative rate and furthermore an improved False Positive rate, implying that fewer resources would be wasted in anticipatory action. We conclude that relying on globally available data sources and working at grid level holds the potential to render a machine learning-based impact model generalisable and transferable to locations outside of the Philippines impacted by TCs. Also, a grid-based model increases the resolution of the predictions, which may allow for a more targeted implementation of anticipatory action. However, it should be noted that an impact-based forecasting model can only be as good as the forecast skill of the TC forecast that goes into it. Future research will focus on replicating to and testing the approach in other TC-prone countries. Ultimately, a transferable model will facilitate the scaling up of anticipatory action for TCs.