Asymmetric Peak-Aware Loss for Peak-Critical Time Series Forecasting

arXiv cs.AI · 6 hours ago

Researchers introduced Asymmetric Peak-Aware Loss (APAL), a training objective for time-series forecasting that penalizes under-predictions more heavily and weights peak regions higher, designed to improve prediction of demand spikes in applications like crowd forecasting. APAL improved Top-10% tail accuracy and F1 scores for peak detection across five forecasting models tested on pedestrian demand datasets while introducing a trade-off with overall mean absolute error. The approach enables time-series forecasters to prioritize accuracy on extreme values and peak timing, which matters more operationally than aggregate error metrics in applications where missed demand spikes carry high costs.