Reliable anticipation of crowd build-up is a practical requirement for contemporary smart-city governance, particularly where transport systems, event programming, and public-safety planning intersect. This study presents a machine-learning framework for forecasting two operationally important outcomes in Emilia-Romagna, Italy: the class of event-related increases in local people presence, and short-horizon future density within census-defined sub-areas (ACEs). The analysis integrates automatically harvested public-event information, daily weather covariates, and anonymised mobile-network observations sampled every 15 minutes. Event data were collected for August–September 2019 and August–September 2020, yielding 5,123 retrieved records before preprocessing; 3,133 valid event-day instances were used for the classification task. For density forecasting, the base learning matrix comprised 58,696 hourly windows derived from 506 ACEs observed across 116 days.
A comparative evaluation of tree ensembles, kernel methods, and neural networks was conducted in a supervised-learning setting. For event-impact classification, Gradient Boosted Decision Trees achieved the strongest overall performance, with a peak accuracy of 0.86 on the best-performing feature configuration. For density forecasting, short-horizon custom accuracy was highest for boosted trees and multilayer perceptrons, whereas Support Vector Regression produced the lowest error profile and the most gradual deterioration as the forecasting horizon extended to 24 hours. The analysis also shows that a simple binary event-presence flag does not improve density forecasting, and that neighbouring-ACE variables contribute far less than initially expected. Framed for urban development and smart-city applications, the study demonstrates how structured event intelligence and mobility analytics can support transport coordination, public-space management, and policy-support systems for safer and more efficient cities.