How CDC Uses Flu Forecasting
The timing and severity of influenza in the United States can vary widely from season to season. Flu forecasting aims to predict the characteristics of influenza seasons before disease activity occurs and is captured by the U.S. influenza surveillance system. Since the 2013-2014 flu season, the Influenza Division at CDC has worked with CDC’s Epidemic Prediction Initiative and external researchers to improve the science and usability of flu forecasting, including the coordination of flu forecasting challenges. Recently, these challenges have been extended to the state level.
This page includes information on current and past influenza forecasting efforts, including working with external researchers and the flu season characteristics being forecasted.
Every influenza season since 2013–2014, CDC’s Influenza Division has engaged with members of the scientific community on real-world influenza forecasting challenges known as FluSight. Originally, 11 teams participated, but interest in FluSight has grown in the six years since FluSight’s inception. We anticipate that more than 20 teams will participate during the 2019-2020 flu season.
Each challenge usually runs from late October through mid-May. Teams independently develop their forecasting approaches using a variety of methods and data sources, and teams submit forecasts to CDC on a weekly basis. At the conclusion of each challenge, CDC determines how accurate each team’s forecasts were by scoring forecasts against actual flu activity, announcing an overall winner.
FluSight brings together multiple forecasts and researchers, making real-time forecasts available in one central location and providing structured and standardized assessments of forecast accuracy. Recently, funding was awarded to Carnegie Mellon University and University at Massachusetts Amherst to further improve the accuracy and communication of influenza forecasts at the national, regional, and state level.
CDC and FluSight partners worked together to develop targets – or the outcome that a forecast is predicting (like season peak) – that would be meaningful to public health. Targets for the main FluSight Challenge at the national level are season onset, peak week, peak intensity, and short-term activity. These target definitions rely on the percent of visits to health-care providers that are for influenza-like illness from the CDC’s U.S. Outpatient ILI Surveillance Network (ILINet).
- Season onset: The first week when ILINet is at or above baseline and remains there for at least two more weeks
- Peak week: The week when ILINet is the highest for the whole season (it is possible to have more than one peak week in a given season if two weeks are “tied” for the highest value)
- Peak intensity: The highest value that ILINet reaches during the season
- Short-term ILI activity: The ILINet value one, two, three, and four weeks ahead of the date that they are available in FluView.
State-level forecasting efforts are recent, only having been available starting in the 2017-18 season. Some targets in these systems vary from the national-level.
At the end of every forecasting season, CDC, FluSight partners, and stakeholders gather to review forecasting approaches, discuss the accuracy of forecasts from the past season, identity overall challenges and successes and plans for future seasons, such as the additions of new forecasting targets. These meetings improve the usefulness of forecasting by providing the opportunity for collaboration among forecasters and public health officials.