Dynamic-Parameter SIRD Forecasting: ReproducibleTime-Varying Rate Estimation for COVID-19

Technical Summary: Dynamic-Parameter SIRD Forecasting

Problem Statement
Classical compartmental epidemic models, such as SIRD (Susceptible-Infected-Recovered-Dead), typically rely on the assumption of constant transmission, recovery, and mortality rates. This static approach limits their ability to accurately track real-world dynamics, including policy shifts, behavioral adaptations, pathogen evolution, and multi-wave outbreaks. Conversely, purely statistical time-series models can track short-horizon signals but often fail to enforce the explicit conservation laws and mechanistic constraints inherent to disease transmission. The paper addresses the need for a hybrid approach that recovers time-varying parameters directly from surveillance data while maintaining the interpretability and structural integrity of mechanistic models.

Methodology
The study proposes a closed-form, two-stage hybrid framework implemented in the open-source Python library dynasir. The methodology avoids iterative high-dimensional optimization, instead utilizing algebraic inversion and multivariate time-series modeling.

1. Inverse Problem Formulation:
   The framework formalizes population transitions using a discrete-time SIRD structure. Rather than optimizing parameters to fit the curve, the method recovers time-varying transition rates—transmission $\alpha(t)$, recovery $\beta(t)$, and mortality $\gamma(t)$—directly from observed compartment increments (new cases, new deaths, and active infections) using closed-form algebraic equations.

   • $\alpha(t)$ is derived from the change in cumulative cases scaled by the mixing denominator $S(t) + I(t)$.
   • $\beta(t)$ and $\gamma(t)$ are derived from the ratios of new recoveries and new deaths to the current infected population $I(t)$.
   • To mitigate noise amplification inherent in numerical differentiation, empirical trajectories are filtered using Savitzky-Golay smoothing (window $w=7$ days) and mapped to logit space to ensure numerical stability and $(0, 1)$ bounds.

2. Two-Stage Hybrid Architecture:

   • Stage 1 (Assimilation): The system recovers the latent rate series $\{\alpha(t), \beta(t), \gamma(t)\}$ via algebraic inversion, applies smoothing, and performs the logit transformation.
   • Stage 2 (Forecasting): A Vector Autoregression (VAR) model is fitted to the logit-transformed rate series. Lag selection is data-driven using the Akaike Information Criterion (AIC). The model forecasts the latent rates for a 12-step horizon, applies the inverse-logit transform, and integrates the forward SIRD equations to generate compartment forecasts.
   • Complexity: The estimator maintains $O(N)$ computational complexity, bypassing the iterative loops required by variational or Bayesian assimilation methods.

3. Validation Strategy:
   The framework was evaluated on a global COVID-19 aggregate dataset (2020–2026) using an 80/20 temporal split. Performance was measured on daily incident counts (new cases and new deaths) rather than cumulative totals to provide a more stringent accuracy signal. Uncertainty was characterized via Monte Carlo sampling ($N=1000$) propagating VAR residual covariance through the non-linear SIRD equations.

Key Results
In a 12-step held-out benchmark, the dynamic SIRD model (dynasir) demonstrated superior point-forecasting accuracy compared to several baselines:

• New Cases: Achieved a Mean Absolute Percentage Error (MAPE) of 39.3%, outperforming Naive Persistence (100.0%), Static SIRD (360.6%), and ARIMA Incidence (631.7%). This represents a 60.7% reduction in error relative to the strongest non-dynamic reference.
• New Deaths: Achieved a MAPE of 29.7%, outperforming Naive Persistence (100.0%), Static SIRD (101.6%), and ARIMA Incidence (79.6%). This represents a 62.7% reduction in error relative to the strongest non-dynamic reference.
• Sensitivity Analysis: The optimal configuration was found to be a smoothing window of 3 days and a VAR lag of 12, yielding incident MAPEs of 16.7% (cases) and 18.5% (deaths) under specific grid settings.

However, the paper notes a significant limitation in probabilistic calibration: nominal 95% prediction intervals achieved only 33.3% coverage for cases and 41.7% for deaths. Consequently, the intervals are reported as diagnostic outputs rather than reliable probabilistic forecasts.

Significance and Claims
The paper positions the dynasir framework not as a definitive operational tool for probabilistic risk assessment or hospital capacity planning, but as a reproducible point-forecasting baseline and an educational prototype.

• Interpretability: The method successfully preserves mathematical conservation laws and generates interpretable trajectories for the time-varying basic reproduction number $R_0(t)$, capturing multi-wave variations that static baselines miss.
• Efficiency: By utilizing closed-form inversion, the approach achieves linear computational complexity, making it scalable for large datasets without the computational burden of iterative optimization.
• Scope of Claims: The authors explicitly limit generalization claims to the tested design (global aggregate series). They state that broader applicability requires further validation on regional, sub-national, and non-COVID surveillance series. The current evidence supports the method's utility for retrospective surveillance audits and short-horizon point forecasting, pending improvements in interval calibration and cross-pathogen validation.