In Table 1 we summarise the comparison of our model and Google Flu Trends with the actual sentinel reports. The Swedish sentinel model based on web queries predicted the sentinel numbers better when delayed reporting was taken into account, no matter what performance indicator was used. This makes sense because we trained the models on complete sentinel data. In other words we have, by training the models on data including late reports, obtained a system which better mimics the values we will get after a while, once the data have been back-populated.
The MAD value of 0.15 can be compared with the change from 1.11 percentage points to 1.36 percentage points between week 45 and 46 in 2009 [21,22], during the height of the pandemic. Thus, the average deviation paralleled the weekly change during the most intense pandemic period.
Although the difference is small, the correlation coefficient indicates that our model performed better than Google Flu Trends for Sweden, with a correlation coefficient of 0.90 versus 0.87. Since Google Flu Trends only provides relative intensity indicators and not absolute estimates of reported influenza, R-squared, RMSEP and MAD could not be calculated for their data.