[Paper Review] A Vector Autoregression Prediction Model for COVID-19 Outbreak

Q Wang, Yongning Zhou+1 more|arXiv (Cornell University)|Feb 6, 2021

COVID-19 epidemiological studies27 references4 citations

TL;DR

This study proposes a multivariate Vector Autoregression (VAR) model to predict daily COVID-19 cases in the U.S. using endogenous and exogenous variables such as temperature, humidity, population density, and social trust. The model forecasts a continued rise in daily cases—projecting around 240,000 new cases by Christmas 2020—if no new control measures are implemented.

ABSTRACT

Since two people came down a county of north Seattle with positive COVID-19 (coronavirus-19) in 2019, the current total cases in the United States (U.S.) are over 12 million. Predicting the pandemic trend under effective variables is crucial to help find a way to control the epidemic. Based on available literature, we propose a validated Vector Autoregression (VAR) time series model to predict the positive COVID-19 cases. A real data prediction for U.S. is provided based on the U.S. coronavirus data. The key message from our study is that the situation of the pandemic will getting worse if there is no effective control.

Motivation & Objective

• To develop a robust, multivariate time series model for short-term forecasting of daily COVID-19 cases in the U.S.
• To incorporate both endogenous (e.g., confirmed cases, deaths, recoveries) and exogenous (e.g., weather, population density, civic engagement) variables to improve prediction accuracy.
• To evaluate model performance through internal validation and compare it with univariate models like ARIMA and compartmental models like SIR.
• To provide actionable insights for public health planning by forecasting epidemic trends under current conditions.
• To ensure model generalizability for future epidemic predictions by using a flexible, interpretable framework.

Proposed method

• Employ a multivariate Vector Autoregression (VAR) model to capture dynamic interdependencies among multiple time series variables.
• Select key predictor variables based on literature: confirmed cases, deaths, recoveries, temperature, precipitation, wind speed, humidity, population density, social trust, and civic engagement.
• Apply stationarity and co-integration tests to ensure model validity and consistent estimation via OLS.
• Use 95% prediction intervals to assess uncertainty and model coverage of real data.
• Conduct internal validation by comparing predicted values against actual daily case counts over multiple time points.
• Focus on daily new cases rather than cumulative cases to better reflect epidemic severity and response effectiveness.

Experimental results

Research questions

• RQ1 Can a multivariate VAR model outperform univariate models like ARIMA and compartmental models like SIR in forecasting daily COVID-19 case trends?
• RQ2 How well does the VAR model capture the dynamic interplay between epidemiological variables and environmental or social factors?
• RQ3 To what extent does the model’s prediction interval cover real-world daily case counts during periods of rapid increase or decline?
• RQ4 What is the forecasted trend in daily new cases over the next 30 days if no new public health interventions are introduced?
• RQ5 How does the model perform in predicting the resurgence of cases, such as the sharp rise observed in late September and November 2020?

Key findings

• The VAR model successfully predicted the rapid increase in daily cases starting from late August 2020, with a trend that closely followed real data.
• For the period from November 3 to November 24, 2020, the model’s predictions (86,890 to 152,604 daily cases) were within the 95% prediction interval, with real values slightly exceeding the upper bound on November 10 and 17 by ~10,000 cases.
• The model projected a daily case count of approximately 177,000 on December 1, 186,000 on December 8, 195,000 on December 15, and 208,000 on December 22, indicating a continued upward trend.
• The 95% prediction intervals consistently covered the real values, demonstrating strong model reliability and uncertainty quantification.
• The model correctly captured the second peak in late November, with a predicted peak of ~153,000 cases on November 24, close to the actual value of 167,012.
• The forecast for the 30-day period starting November 24, 2020, projected a daily case count of around 240,000 by Christmas, highlighting the risk of continued transmission without new control measures.