Short-term crowd forecasting for SAIL2025

Abstract

Every five years, the SAIL event in the center of Amsterdam attracts millions of visitors. Managing large crowds in relatively small areas requires real-time insight into the traffic situation – particularly pedestrian flows – and forecasts thereof. To provide reliable short-term crowd forecasts without historical data, we developed an adaptive hybrid AI model that learns exclusively from live sensor, weather, and event data during the SAIL event. Our hybrid architecture integrates a periodically retrained LightGBM model with a Kalman Filter for real-time error correction and Conformalized Quantile Regression yields calibrated 90 % prediction intervals. The resulting system produces robust 4-hour forecasts and offers a generalizable solution to manage crowds at other large-scale events, particularly those infrequent events with limited historical data.