Forecasting characteristics.

Data range 2012–2015. (a) Single Source: Forecast accuracy for each individual source (Weather, HealthMap, Twitter, Google Flu Trends, and Google Search Trends). No particular source is the best for all countries. (b) Multiple Sources: Percent increase in forecast accuracies while combining multiple sources at model level and at data level over best single source forecasts [1]. Model level gives better overall performance. (c) Ablation test: Percent reduction in forecast accuracies while removing one source at a time from the fused model. Removing a source can lead to better performance for some countries [1]. (d) Segregation Test: Percent increase in forecast accuracies for US ILI data considering forecasts made individually for ILI cases by age and by subtype over forecasts on unsegregated data. Segregated methods show better accuracy. (e) Instability Correction: Percent increase in forecast accuracies for different countries after correction over uncorrected forecasts. Significant improvement can be seen for countries like Argentina and Paraguay [1]. ILI, Influenza-like Illnesses.