Observation-Aware Calibration of Agent-Based Epidemic Simulations from Public Aggregate Reports: A COVID-19 City-Report Benchmark Study

Abstract

City reports record confirmation, recovery or discharge, and death by publication date; agent-based simulations track latent disease events. We calibrated Covasim through an observation layer that converts cumulative reports to daily increments and aligns them with simulation outputs. A two-parameter Optuna-TPE search estimated transmission and initial infections. Shenzhen was reproduced under fixed stochastic settings, followed by independent recalibration in 12 non-Hubei first-wave cities. The Shenzhen fit reproduced the reported peak within one day (RMSE 7.2072, MAE 5.5848) but underestimated total reported cases by 20.66%. Gompertz fitted the same short curve more closely (RMSE 3.6874). Across the city panel, median RMSE was 3.3049 and peak-normalized RMSE was 0.1933. The framework links report-curve calibration to stochastic, auxiliary, and scenario outputs within a historical COVID-19 benchmark. Policy evaluation lies outside this benchmark.