The 10-Day Window: Intervention Timing, Parameter Uncertainty, and the Collapse of Epidemic Forecast Reliability

Epidemic forecasting models are routinely used to guide public health decisions, yet the temporal boundary beyond which parameter uncertainty renders forecasts operationally useless remains poorly characterised. We introduce a formal, computable predictability horizon—defined as the first day on which the interquartile range of an ensemble of SEIR trajectories exceeds 20% of the ensemble median cumulative incidence—and demonstrate that this horizon is a stable, reproducible quantity jointly determined by the width of parameter uncertainty and the timing of nonpharmaceutical interventions (NPIs). We simulated 1,000-member ensembles of a standard SEIR model using a literature-calibrated parameter space for a generic respiratory pathogen (R0 ≈ 2.1–4.9, consistent with pre-variant SARS-CoV-2). Under nominal uncertainty, the predictability horizon is 13 days. Doubling the parameter envelope collapses it to 7–8 days; halving it extends it to 23 days. A 40% reduction in transmission rate (β), analogous to moderate NPI packages, extends the horizon by up to 6 days when applied at Day 0, with diminishing returns through Day 5, and provides no measurable gain when applied after Day 10. A complementary σ-reduction intervention (vaccination-like) produces a smaller absolute horizon gain through a mechanistically distinct pathway. A phase diagram mapping intervention timing × parameter uncertainty to predictability horizon length is presented as a decision-support contribution currently absent from the epidemiological literature.