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

Bernd Skiera, Lukas Jürgensmeier, Kevin Stowe, Iryna GurevycharXiv 2020 · arXiv

Abstract

Official COVID-19 case reporting in Germany, primarily managed by the Robert Koch Institute (RKI), suffered from significant structural issues during the early stages of the pandemic. These included substantial reporting delays and a pronounced "weekend effect," where case numbers were underreported by up to 50%. This paper evaluates the effectiveness of alternative data sources—specifically Johns Hopkins University (JHU) data, Google Search trends, and Twitter activity—as real-time proxies for infection rates. Using log-log regression models and lag analysis, the study finds that digital behavior serves as a powerful leading indicator, capable of predicting infection trends up to three days before they appear in official government statistics.

Introduction

In the early months of 2020, policymakers relied heavily on daily infection counts to make critical decisions regarding lockdowns and healthcare resource allocation. However, the "ground truth" provided by official sources like Germany’s Robert Koch Institute (RKI) was often flawed. Local health departments frequently failed to report data over weekends, leading to systematic dips in official numbers. Furthermore, the "final" count for any given day often took several days to settle due to reporting lags. This research explores how digital proxies can fill this information vacuum and provide more agile, real-time insights.

Methodology and Data Sources

The researchers employed a log-log regression model to compare official RKI data against three alternative streams:

  1. RKI (Official): Confirmed cases in Germany.
  2. JHU CSSE: A global dashboard aggregating various official and unofficial sources.
  3. Google Trends: Daily search volume for the term "corona" in Germany.
  4. Twitter: German-language tweets containing the term "corona."

To account for the systematic reporting errors in RKI data, the models included a "weekend dummy" variable. Additionally, the team performed a lag analysis, testing offsets between zero and ten days to determine the predictive lead time of digital behavior relative to official reports.

The Challenge of Reporting Lags

A primary obstacle in pandemic monitoring is the time it takes for a case to be officially recorded. A patient must develop symptoms, seek testing, and have that test processed and reported through multiple levels of bureaucracy.

Figure 3: Illustration of the reporting lag, showing how the share of cases for a specific period is added incrementally over several days.

As shown in Figure 3, the "final" number of cases for a specific date is not known immediately; instead, cases trickle in over the following days. This delay makes raw daily reports a lagging indicator of the actual epidemiological situation.

The Weekend Mirage and Data Discrepancies

The "weekend effect" creates a dangerous blind spot for real-time policy making. The study quantifies this effect, finding that RKI underreported weekend cases by between 38.2% and 49.6%. Even the widely used JHU dashboard, which attempted to mitigate these gaps, showed an average deviation of 79% from RKI’s eventually settled "final" numbers.

Figure 5: Comparison of daily new infections predicted by JHU versus official RKI data, with the bottom panel highlighting the percentage difference over time.

Figure 5 illustrates these discrepancies. While JHU data often tracked higher than RKI's initial reports, this was frequently due to JHU's attempts to "fill in" the weekend gaps that RKI missed. This makes JHU a more useful real-time tool, despite its deviations from the retrospective RKI ground truth.

Digital Proxies as Leading Indicators

The most significant finding of the study is the predictive power of Google Search and Twitter data. Because individuals often search for symptoms or discuss the virus online before seeking clinical testing, these digital "breadcrumbs" act as leading indicators.

  • Predictive Lead Time: Both Google Search and Twitter data best predict official numbers with a 3-day lag.
  • Model Fit: The JHU-based model achieved an $R^2$ of 0.892, while the Google Search model reached 0.807, suggesting that digital proxies capture the vast majority of variance in infection trends.

While JHU data was more accurate in predicting final counts (MAPE of 34.9% when adjusted for the weekend effect), Google (50.8%) and Twitter (59.8%) provided their best predictions three days in advance. In a fast-moving pandemic, this three-day head start is often more valuable than a more accurate retrospective count.

Conclusion and Future Directions

Digital proxies offer a vital supplement to official health statistics, providing agility where traditional reporting lacks. However, challenges remain, including the relative nature of Google Trends data and the difficulty of mining "walled garden" social platforms. The authors conclude that official bodies should provide more transparent update histories, and tech companies should share virus-related data more openly to aid public health efforts. Ultimately, the true number of infections—the "dark number"—may be more closely reflected by these digital proxies than by clinical tests alone.

Summary updated

BibTeX
@misc{skiera2020how,
  author = {Bernd Skiera and Lukas Jürgensmeier and Kevin Stowe and Iryna Gurevych},
  title = {{How to Best Predict the Daily Number of New Infections of Covid-19}},
  year = {2020},
  eprint = {2004.03937},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2004.03937},
}
Figure