Period-Aware Inductive Bias Versus Scale on Influenza-like Illness Forecasting

YongKyung Oh, Alex Bui

Keywords: Time Series Forecasting, Influenza-like Illness, Periodic Non-Stationarity, Inductive Bias

Abstract: Forecasting influenza-like illness (ILI) is important for public-health planning, but weekly surveillance series combine short-lag persistence with annual recurrence and season-to-season shifts in peak timing and amplitude. Foundation models (FMs) and large language model (LLM)-adapted forecasters are increasingly used for time-series forecasting, yet on the standard ILI benchmark they do not show a clear average-error advantage over strong compact baselines. We therefore propose EPIC (Efficient Periodic Inductive Convolution), a compact attention-free forecaster that combines temporal convolution, period projection, and latent horizon projection. Under the same benchmark setting, EPIC achieves lower four-horizon average error than published compact and LLM-based comparators, and ablations show that the temporal and periodic branches contribute complementary information. This advantage is concentrated in the average and shorter-to-medium horizons, while the longest horizon remains a limitation. On this weekly ILI benchmark, a small period-aware inductive bias is more effective than additional model scale.

Submission Number: 177