A new forecasting model remembers what past flu seasons looked like instead of treating every week like the first one.
The approach comes from an arXiv preprint, "TERN: A Delta-rule Memory with a Seasonal Reference and Online Adaptation for Epidemic Forecasting" (arXiv:2609.18407), posted September 17, 2026. TERN pairs a delta-rule fast-weight memory - one that decays channel-wise and erases along a learned address, gated by local epidemic-phase signals - with an explicit reference to prior seasons and the ability to keep adapting as new data arrives. On three Cola-GNN influenza benchmarks, the authors report TERN beat both epidemic graph models and general-purpose forecasters, and matched or exceeded simple seasonal baselines. A controlled ablation isolated the memory component and confirmed it, not just the seasonal reference, was doing real work.
Weekly flu counts steer vaccine distribution and public-health alerts, but forecasters have always faced a thin-data problem: each region only offers a handful of past seasons, and models that treat all history the same either ignore useful prior-year shape or keep leaning on associations that go stale once a wave peaks. TERN's gating mechanism is a more surgical fix than the usual answer of bigger models or longer context windows - it tries to decide, week by week, what's still relevant.
It's one preprint validated on three benchmarks, not a system already running in a health department. Whether gated memory generalizes past Cola-GNN's datasets, and whether public-health agencies would trust a model's judgment about what to forget, are separate questions this paper doesn't answer.