[Paper Review] Fast and Accurate Forecasting of COVID-19 Deaths Using the SIkJ$α$ Model

Ajitesh Srivastava, Tianjian Xu+1 more|arXiv (Cornell University)|Jul 10, 2020

COVID-19 epidemiological studiesMathematics15 references18 citations

TL;DR

This paper proposes the SIkJα model extension for fast, accurate forecasting of COVID-19 deaths by transforming complex epidemic dynamics into a system of linear equations, enabling efficient parameter learning via linear regression. It achieves superior performance—outperforming seven CDC-used models in root mean squared error during most of the evaluation period—while maintaining sub-second execution times across all U.S. states, counties, and 184 countries on modest hardware.

ABSTRACT

Forecasting the effect of COVID-19 is essential to design policies that may prepare us to handle the pandemic. Many methods have already been proposed, particularly, to forecast reported cases and deaths at country-level and state-level. Many of these methods are based on traditional epidemiological model which rely on simulations or Bayesian inference to simultaneously learn many parameters at a time. This makes them prone to over-fitting and slow execution. We propose an extension to our model SIkJ$α$ to forecast deaths and show that it can consider the effect of many complexities of the epidemic process and yet be simplified to a few parameters that are learned using fast linear regressions. We also present an evaluation of our method against seven approaches currently being used by the CDC, based on their two weeks forecast at various times during the pandemic. We demonstrate that our method achieves better root mean squared error compared to these seven approaches during majority of the evaluation period. Further, on a 2 core desktop machine, our approach takes only 3.18s to tune hyper-parameters, learn parameters and generate 100 days of forecasts of reported cases and deaths for all the states in the US. The total execution time for 184 countries is 11.83s and for all the US counties ($>$ 3000) is 101.03s.

Motivation & Objective

• To develop a fast, scalable, and accurate method for forecasting COVID-19 deaths at national, state, and county levels.
• To address limitations of traditional epidemiological models that rely on slow numerical integration or Bayesian inference, which are prone to overfitting and high computational cost.
• To incorporate complex epidemic dynamics—such as reporting delays, asymptomatic transmission, immunity, and mobility—into a simplified, linearizable framework.
• To enable rapid policy-relevant forecasting by reducing training and inference time to under 12 seconds for all 184 countries and under 102 seconds for all U.S. counties.
• To provide a robust alternative to existing CDC forecasting models with improved predictive accuracy and open-source reproducibility.

Proposed method

• The SIkJα model extends the SIR framework by introducing k sub-states of infection, each with a distinct transmission rate βi, to model time-varying infectiousness over a period of kJ days.
• The model incorporates reporting delay via a fixed lag λ, where reported cases R_t are generated from infected individuals I_{t−λ} with probability γ.
• It models mobility effects through a flow matrix F(q,p), which transfers infection from region q to p based on travel patterns and regional transmission rates.
• The model accounts for non-carriers (immune or isolated) via a fraction ρ of the population excluded from susceptible and infected compartments.
• The core innovation is transforming the nonlinear system into two sequential linear regression problems: one for infection dynamics and one for death reporting, enabling fast, stable, and scalable learning.
• Parameter learning is achieved through linear regression on cumulative infection and death data, avoiding iterative optimization and reducing overfitting risk.

Experimental results

Research questions

• RQ1 Can a simplified, linearized model capture complex epidemic dynamics such as reporting delays, asymptomatic spread, and mobility while maintaining high forecasting accuracy?
• RQ2 Does replacing traditional differential equation solvers or Bayesian inference with linear regression improve both speed and accuracy in death forecasting?
• RQ3 How does the SIkJα model compare to seven CDC-used forecasting methods in terms of root mean squared error across U.S. states during the pandemic?
• RQ4 To what extent do data inconsistencies and reporting conflicts affect model performance, and how does the model handle such noise?
• RQ5 Can the model efficiently scale to forecast deaths across all U.S. counties and 184 countries with minimal computational overhead?

Key findings

• The SIkJα model outperformed all seven CDC-used forecasting methods in terms of root mean squared error during the majority of the evaluation period (May 10, 2020 – June 28, 2020).
• On a 2-core desktop machine, the model completed hyperparameter tuning, parameter learning, and 100-day forecasts for all U.S. states in just 3.18 seconds.
• The total execution time for forecasting all 184 countries was 11.83 seconds, and for all U.S. counties (>3000), it was 101.03 seconds.
• The model demonstrated robustness to data inconsistencies, with relative errors among methods converging when evaluated on states with no reporting conflicts.
• Weekly RMSE evaluations showed consistent performance, with only minor dips in accuracy compared to top baselines like YYG_ParamSearch and UCLA_SuEIR in some forecast windows.
• The model enables real-time computation of dynamic reproduction numbers and model-based case fatality rates, which are updated weekly and publicly available via an interactive web portal.