Spatio-Temporal Hierarchical Feature Engineering for Forecasting of Urban Footfall

Patterns of footfall counts in urban environments show regularity at various spatial and temporal scales. In this work, we study a lightweight hierarchical approach in which forecasts use four lagged higher-level aggregates as predictors trained with simple CPU-only models. For a fair comparison, the baseline is expanded to use a horizon-matched lag window, so that the variants have access to the same maximum lookback in time. The study uses hourly pedestrian counts from 13 sensors on two shopping streets in Newcastle upon Tyne, aggregated across spatial and temporal levels. Combined spatial and temporal aggregate predictors reduced forecast error by adding information from higher aggregation levels without changing the base learner. The best-performing configuration was SHTH+CP, which combines spatial and temporal parent features with a spatio-temporal cross-parent, and yielded an average pooled 4.3% improvement in RMSE and 3.5% in MAE, with the largest gains at 12 h directional counts, where RMSE decreased by 6.7% and MAE by 11.4%.