Abstract - The care team at Zurich's municipal homeless shelter faces an increasing challenge as its users show substantial mental health and sociodemographic differences. Consequently, a more specialized workforce is required to provide adequate care. Given the shortage of skilled labour, accurate capacity and resource planning become more essential. Simultaneously, freely available machine learning libraries such as sci-kit learn or TensorFlow offer sophisticated algorithms that can be used for demand forecasting. With comparably few lines of code, these repositories can be implemented to make predictions based on historical time- series data. This study aims to build a proof-of-concept for a shelter demand forecasting model in collaboration with Zurich's municipal social housing department Wohnen+Obdach (W+O), and the city's largest private shelter operator Sozialwerk Pfarrer Sieber (SPS). Therefore, the effect size of external influencing variables on the shelter's demand is examined to determine explanatory variables for the forecasting model. As part of the exploratory data analysis, the shelter client population, including over 3'000 individuals, is analysed and segmented based on sociodemographic characteristics and shelter usage behaviours described in the literature. The findings are then used to feed, train, validate and evaluate different algorithms' fit for the purpose of forecasting. The models range from simple decision trees to multi- layer neural networks. Recurring weekly, monthly, and annual demand patterns were found in nine years of daily shelter visitor records between 2012 and 2020. On average, a 10% decrease in usage was found during weekend nights. In addition, across multiple seasons, a substantial increase was seen after sudden temperature drops during fall. In contrast, a considerable decrease was found during Zurich's coldest winter months, as many shelter clients migrate to the private operator once their gates open in mid- November. Besides this seasonal migration pattern, the analysis also showed that the long-term relapse rate of housing reintegration programs must be considered when predicting nightly shelter stays. The practical experience gained from this project and the validation of the model's performance further advise how a shelter demand forecasting model could be implemented in a productive environment. Therefore, the advantages of multivariate and univariate approaches are discussed. These insights can be elaborated on for further research and development.