COVID-19: How to Best Predict the Daily Number of New Infections

Article of Bernd Skiera, Lukas Jürgensmeier, Kevin Stowe and Iryna Gurevych

Bernd Skiera (Goethe University), Lukas Jürgensmeier (Goethe University), Kevin Stowe (TU Darmstadt) and Iryna Gurevych (TU Darmstadt) have just completed their first “corona paper” (https://bit.ly/34ir6t5). They compare the ability of several data sources, in particular Johns Hopkins University (JHU), Google search data and Twitter data, to predict the official number of new infections of Covid-19 and examine the need to complement the official numbers with additional predictions.

Using Germany as an illustration, their most important findings are:

1) The widely popular predictions from Johns Hopkins University (JHU) deviate for Germany on average by 79% from the official numbers.

2) Using a simple regression to adjust the prediction from Johns Hopkins University (JHU) reduces the prediction error to 35%.

3) Google search and Twitter data predict three days ahead of time and predict better than the unadjusted predictions of John Hopkins University.

4) The official numbers of Germany suffer from an underreporting on weekends in the area of more than 40%.

The main conclusion of the authors is that there is a strong need for complementing the official numbers in Germany (and, probably, in other countries as well) with other predictions, such as those that build upon Google search and Twitter data.

Download the result here.