AbstractNowadays, drinking water demand forecasting has become fundamental to efficiently manage water distribution systems. With the growth of accessible data and the increase of available computational power, the scientific community has been tackling the forecasting problem, opting often for a data-driven approach with considerable results. However, the most performing methodologies, like deep learning, rely on the quantity and quality of the available data. In real life, the demand data are usually affected by the missing data problem. This study proposes an analysis of the role of missing data imputation in the frame of a short-term forecasting process. A set of conventional imputation algorithms were considered and applied on three test cases. Afterward, the forecasting process was performed using three state-of-the-art deep neural network models. The results showed that a good quality imputation can significantly affect the forecasting results. In particular, the results highlighted significant variation in the accuracy of the forecasting models that had past observation as inputs. On the contrary, a forecasting model that used only static variables as input was not affected by the imputation process and may be a good choice whenever a good quality imputation is not possible.