The risk of flooding is on the rise in Delta cities, such as Ho Chi Minh City (HCMC) in Vietnam, with projections indicating further increases due to climate change and urbanization. Flood risk analyses, for which loss modeling is a key component, play a crucial role in decisions on flood risk management and urban development. Probabilistic multi-variable loss models are increasingly being used to improve loss estimation, as they describe loss processes better and inherently provide a quantification of uncertainties. However, such models are often based on input variables that are determined by expert judgment. Thus, we propose the first probabilistic multi-variable flood loss model designed for residential buildings in delta cities such as HCMC (BN-FLEMO∆). BN-FLEMO∆ is built upon new building-level empirical survey data. The model is developed with an automatic machine learning-based (ML) feature selection framework and a systematic learning process to determine the optimal structure of the Bayesian Network. Based on a methods comparison, we demonstrate the following key advantages of BN-FLEMO∆: 1. enhanced, empirically-based description of flood loss processes leading to improved accuracy in loss estimation; 2. provision of a probability distribution of losses and inherent quantification of modeling uncertainty; 3. network structure allows model application even when data for one or more input variables are missing, which is particularly valuable in data-scarce environments. We therefore expect that BN-FLEMO∆ will significantly improve risk analyses in HCMC and similar delta cities and support decision-makers in developing sustainable flood risk management strategies for these dynamic flood-prone regions.