Learning to Forecast and Forecasting to Learn from the COVID-19 Pandemic

Ajitesh Srivastava, Viktor K. PrasannaarXiv 2020 · arXiv

Introduction

Predicting the trajectory of a pandemic like COVID-19 is a significant challenge because the underlying transmission dynamics are non-stationary. Public health interventions, changes in human mobility, and social distancing measures mean that models trained on historical data can quickly become obsolete. Traditional epidemic models, such as SIR or SEIR, often rely on complex differential equations that are computationally intensive and slow to adapt to these sudden shifts. Researchers at USC developed a linearized, heterogeneous infection rate model designed to adapt to rapidly changing trends while providing actionable metrics for policy evaluation.

Methodology: The SI-kJα Model

The proposed Heterogeneous Susceptible-Infected (SI) model introduces several engineering innovations to improve adaptability and efficiency:

  1. Sub-state Modeling: To account for varying infectiousness over time, the model uses $k$ sub-states ($I_1, \dots, I_k$). This recognizes that an individual's transmission rate ($\beta$) depends on the duration of their infection.
  2. Mobility Integration: The model incorporates inter-region travel data ($F$) to capture how movement between countries or states initiates new outbreaks.
  3. Linearization and Weighted Least Squares (WLS): By transforming non-linear epidemic equations into a linear system, the authors can solve for parameters efficiently using WLS.
  4. The Forgetting Factor (α): A hyperparameter $\alpha \leq 1$ is introduced to prioritize recent data points. This allows the model to "forget" outdated trends caused by past policy regimes and focus on current transmission dynamics.
The SI-kJα model architecture featuring sub-states and mobility integration.

Performance and Results

The model, designated SI-kJα, was evaluated using global data from the JHU CSSE dataset and US state-level data from the New York Times. It significantly outperformed the Gen-SEIR baseline across multiple metrics. For US state-level predictions, the ensemble model achieved a Mean Absolute Percentage Error (MAPE) of 5.93%, compared to 14.31% for the baseline. Globally, the model maintained high accuracy, even in regions where traditional models struggled to converge.

Comparison of true vs. predicted cases across Australia, US, Iran, and Italy, demonstrating the model's high predictive accuracy.

Discussion: Forecasting to Learn

A key contribution of this work is the concept of "Forecasting to Learn." The authors argue that model parameters are valuable data sources themselves. By analyzing the transmission rate ($\beta$) before and after policy shifts, they developed the Contact Reduction Score (CRS), which measures the fractional change in transmission rates due to social distancing. This distinguishes between "effort" (contact reduction) and "outcome" (actual case reduction), providing a more nuanced view of regional pandemic responses.

While the model is highly effective for short-term forecasting, it currently focuses on reported cases. Future work aims to better integrate asymptomatic populations and increase granularity to the city and county levels, addressing the challenges of noisier data at smaller scales.

Summary updated

Abstract

Accurate forecasts of COVID-19 is central to resource management and building strategies to deal with the epidemic. We propose a heterogeneous infection rate model with human mobility for epidemic modeling, a preliminary version of which we have successfully used during DARPA Grand Challenge 2014. By linearizing the model and using weighted least squares, our model is able to quickly adapt to changing trends and provide extremely accurate predictions of confirmed cases at the level of countries and states of the United States. We show that during the earlier part of the epidemic, using travel data increases the predictions. Training the model to forecast also enables learning characteristics of the epidemic. In particular, we show that changes in model parameters over time can help us quantify how well a state or a country has responded to the epidemic. The variations in parameters also allow us to forecast different scenarios such as what would happen if we were to disregard social distancing suggestions.
BibTeX
@misc{srivastava2020learning,
  author = {Ajitesh Srivastava and Viktor K. Prasanna},
  title = {{Learning to Forecast and Forecasting to Learn from the COVID-19 Pandemic}},
  year = {2020},
  eprint = {2004.11372},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2004.11372},
}
Figure