Good Prediction, Wrong Latent Order: A Reporting-Kernel Misspecification Counterexample for Simulator-Based System Identification
Abstract
When dynamical models are fitted to delayed reports, underestimating the reporting delay can make the unaccounted delay appear as an additional latent timescale. We demonstrate this in a controlled epidemic-surveillance simulator where the true admissions-to-death response has one exponential timescale and deaths have a mean reporting delay of 7 days. Validation selects one timescale under the matched assumption, but fixing the mean at 3, 4, or 5 days selects two in all 60 configurations while error over three entirely held-out weekly-death trajectories remains below 0.9%. An idealized transfer calculation attributes this to compensation between reporting and latent dynamics. The pattern is asymmetric: a fixed 11-day assumption retains one timescale but raises error to 1.66% to 2.22%. Thus accurate held-out prediction does not certify latent order; comparison across plausible, prespecified reporting-delay settings can flag observation-model compensation before learned structure is interpreted mechanistically.