Streamflow predictions in ungauged basins (PUB) has been geared towards data-driven methods, including the use of machine learning methods such as random forest (RF). Such methods are applied in PUB regionalization or the transfer of a streamflow model from gauged to ungauged (sub) basins or watersheds after grouping watersheds on similarity rules. Regionalized streamflow models are needed for tropical-mountainous regions like Luzon, Philippines - where gauged data is limited - but demands on streamflow modelling for water resources accounting and management are high. In 21 watersheds in Luzon, we “regionalize” RF streamflow models after grouping watersheds based on: a principal components analysis (PCA-clustered), by major river basins (basin-clustered), whole study area scale (one-clustered). Another method without watershed grouping (watershed-level) was also included and inter-compared. Among the four methods, goodness-of-fit evaluations revealed that PCA-clustered method was higher by at most 0.35 coefficient of determination R2 and 0.31 nash-sutcliffe efficiency NSE, and the least bias in 8 of 12 monthly flows. These are attributed to the added-value of homogeneous watershed grouping, reflected by higher importance (to RF models), of static covariates from open and high-resolution data. Normalized errors from monthly streamflow showed a clear bias (least with the PCA-clustered metohd), linked to season and water management practices in the study area. Ungauged watersheds in the Philippines can effectively use streamflow models from gauged watersheds if they belong to the same cluster.Key PointsUngauged sub-basins (watersheds) can effectively use RF streamflow models from gauged watersheds if both belong to the same cluster.Biophysical data from high-resolution open data are valuable: as basis for watershed clustering and as streamflow predictors to RF models.The predicted streamflow from the RF models reflects seasonal deviations in streamflow caused by natural and man-made water regulation