Improving the Timeliness of Data on Influenza-like Illnesses using Google Search Data

Abstract

Google Flu Trends reports the current flu activity in the US based on search activity indicators. A recent letter to Nature reported how the model was obtained from historical search records. Using internet search activity to improve short-term forecasts is an exciting new development. The letter to Nature shows that it does indeed contain useful information. Recently, however, there has been a dramatic increase in flu activity in the US — an episode that was missed entirely by the Google Flu Trends model. It is shown how the Google Flu Trend forecasts can be robustified, and how the model can be improved. The objective of the former is to limit the duration of a forecast failure. The latter shows how the forecasts error can be reduced significantly. It turns out that a dynamic model with calendar effects has similar forecast performance as the robustified Google Trends model. Therefore, two further models are built that use Google Trends data. These improve on the model with calendar effects. The pooled model is better still, so for the periods considered, search data can indeed help with nowcasting.