henhaobio
Sign in

bioRxivEpidemiology

Dynamic forecasting of Zika epidemics using Google Trends

Yue Teng, Dehua Bi, Guigang Xie, Yuan Jin, Baihan Lin, Dan Feng

PDFDOIPreprintData
In-depth full-text digest available · Jump to digest ↓

Abstract

We developed a dynamic forecasting model for Zika virus (ZIKV), based on real-time online search data from Google Trends (GTs). It was designed to provide Zika virus disease (ZVD) surveillance for Health Departments with early warning, and predictions of numbers of infection cases, which would allow them sufficient time to implement interventions. We used correlation data from ZIKV epidemics and Zika-related online search in GTs between 12 February and 25 August 2016 to construct an autoregressive integrated moving average (ARIMA) model (0, 1, 3) for the dynamic estimation of ZIKV outbreaks. The online search data acted as an external regressor in the forecasting model, and was used with the historical ZVD epidemic data to improve the quality of the predictions of disease outbreaks. Our results showed a strong correlation between Zika-related GTs and the cumulative numbers of reported cases, both confirmed and suspected (both p<0.001; Pearson Product-Moment Correlation analysis). The predictive cumulative numbers of confirmed and suspected cases increased steadily to reach 148,510 (95% CI: 126,826-170,195) and 602,721 (95% CI: 582,753-622,689), respectively, in 21 October 2016. Integer-valued autoregression provides a useful base predictive model for ZVD cases. This is enhanced by the incorporation of GTs data, confirming the prognostic utility of search query based surveillance. This accessible and flexible dynamic forecast model could be used in the monitoring of ZVD to provide advanced warning of future ZIKV outbreaks.

In-depth summary

Claude Opus 5.5 · xhigh effort · 5 Nov 2026 · from the full text of version 1 (22 Sept 2016)

Incorporating Zika-related Google Trends search volumes as an external regressor in an ARIMA(0, 1, 3) model enables eight-week dynamic forecasting of cumulative Zika virus cases across the Americas.

Traditional government and healthcare surveillance of Zika virus disease is resource-intensive and delayed during rapid outbreaks. Near real-time Google Trends query volumes correlate with reported case trajectories and can serve as an external predictor for short-term epidemic forecasting.

Key results

  1. ShownCumulative Zika-related Google Trends search volume correlates strongly with cumulative confirmed and suspected Zika cases in the Americas.

    Over the 29-week period from 12 February to 25 August 2016 (reaching 111,333 confirmed and 466,815 suspected cases), cumulative Google Trends search volume showed Pearson correlations of 0.78 (p < 0.001) with confirmed cases and 0.99 (p < 0.001) with suspected cases. · Abstract; Correlations between data on ZIKV outbreaks and…; Figure legends

  2. ShownAn ARIMA(0, 1, 3) model using Google Trends as an external regressor accurately brackets held-out confirmed cases but slightly overpredicts suspected cases.

    Trained on 75% of the time series (12 February to 28 July 2016) and tested on the remaining 25% (5 August to 25 August 2016), the model's 25 August 2016 95% confidence interval for confirmed cases (93,540 to 127,400) contained the actual count (111,333), whereas actual suspected cases (466,815) fell slightly below the predicted 95% interval (476,077 to 525,811). · Statistical machine learning and reconstructed…; Reconstructed ARIMA model; Figure legends

  3. ShownBaseline ARIMA nowcasting projected steady weekly Zika search volumes through September and October 2016.

    An in-sample ARIMA(0, 1, 3) model fitted to the 29-week Google Trends series predicted approximately 350 new searches per week (95% CI: 228-475) over the next 8 weeks, reaching cumulative search volumes of 10,996 (95% CI: 10,532-11,459) on 22 September 2016 and 12,398 (95% CI: 11,447-13,349) on 21 October 2016. · Dynamic nowcasting of the Google Trends data; Figure legends; Tables

  4. ShownDynamic forecasting projected substantial continued growth in cumulative confirmed and suspected Zika cases in the Americas through 21 October 2016.

    Using the forecasted cumulative Google Trends series as an external regressor in the ARIMA(0, 1, 3) model fitted to 12 February-25 August 2016 data projected cumulative confirmed cases to reach 128,826 (95% CI: 113,420-144,231) by 22 September and 148,510 (95% CI: 126,826-170,195) by 21 October 2016, alongside 602,721 suspected cases (95% CI: 582,753-622,689). · Abstract; Forecasting the ZIKV outbreaks in Americas…; Conclusions; Tables

Methods, data and reproducibility

Retrieved weekly Zika Google Trends search volumes and PAHO confirmed and suspected Zika case counts in the Americas across 29 weeks (12 Feb to 25 Aug 2016). Computed Pearson correlations in Python 2.7 (Scipy), validated an ARIMA(0, 1, 3) model in R 2.14 using a 75%/25% split, and forecasted cumulative cases 8 weeks ahead (2 Sep-21 Oct 2016) using ARIMA-predicted Google Trends volumes as an external regressor.

  • Google Trends weekly and cumulative search volume logs for Zika in the Americas (12 February to 25 August 2016) · human · n = 29 weeks
  • Pan American Health Organization (PAHO) reported weekly and cumulative confirmed and suspected Zika virus disease cases in the Americas · human, Zika virus (ZIKV) · n = 111,333 confirmed and 466,815 suspected cases across 29 weeks
Methods & tools
Google Trends, ARIMA (0, 1, 3), Python 2.7, Scipy, R (version 2.14)
Reagents & environment
Python 2.7 with Scipy for Pearson correlation analysis; R version 2.14 for ARIMA(0, 1, 3) time-series modeling.
Reproducibility gaps
No analysis code is provided, although the 29-week historical series and 8-week forecasts are listed in supplementary Tables S1-S3.
Limitations (3)
  • Effective local surveillance with Google Trends depends on internet access and must be interpreted within the local geographic context of ZIKV transmissibility. · stated by the authors
  • Correlation and ARIMA(0, 1, 3) analyses were performed on cumulative rather than weekly incident counts, which inflates correlations between monotonically non-decreasing series and still missed the 25 August 2016 suspected-case count (466,815 vs 95% CI: 476,077-525,811). · noted in this summary
  • The forecasting evaluation does not compare out-of-sample error against a baseline ARIMA model without the Google Trends external regressor. · noted in this summary

Read this if you explore digital epidemiology or time-series forecasting of arboviral outbreaks using web search query volumes.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISZhang '16Ceia-Hasse '22Perkins '16Roth '19Gwon '18Bewick '16
70% semantic matchEpidemiology
Qian Zhang, Kaiyuan Sun, Matteo Chinazzi+11 more·bioRxiv·Jul 29, 2016