Table 4 Glossary of terms commonly used in forecasting

Ensemble model

A model that incorporates two or more models into a single model.

Epidemic Prediction Initiative

A CDC initiative launched in 2014 that aims at improving the science and usability of epidemic forecasts by facilitating open forecasting projects with specific public health objectives.

FluSight Challenge

A multi-participant competition that began during the 2013–14 influenza season (then called the “Predict the Influenza Season Challenge”) to forecast the timing, intensity, and short-term trajectory of the influenza season.

Forecast

A quantitative, probabilistic statement about an unobserved event, outcome, or trend and its surrounding uncertainty, conditional on previously observed data.

Forecast accuracy

A measurement of how well the forecast matched the outcome once it has been observed. There are a number of ways forecast accuracy can be measured, but CDC uses the logarithmic score. For more information regarding logarithmic score, please see the definition below.

Forecast calibration

An indicator of reliability in assigning probabilities. For FluSight forecasts, calibration is evaluated by assessing how often forecasts were correct.

Forecast confidence

A characterization of the uncertainty in a forecast. The Epidemic Prediction Initiative requires that forecast confidence be expressed as a probability (e.g., a 0.2 probability or 20% chance that the peak week of the influenza season will be on week 2).

Hindcast

Forecast of past conditions, also known as “pastcast.” For example, due to delays in reporting and data accrual, the FluSight forecast for ILI outpatient visits “one week ahead” is actually a forecast for the previous calendar week.

ILI

Influenza-like illness, fever and either a cough or sore throat.

ILINet

US Outpatient Influenza-like Illness Surveillance Network; a surveillance system that accrues weekly data on the number of patients with ILI and the total number of patients seen in healthcare settings, reported by outpatient healthcare providers in the United States.

Logarithmic score

The logarithm of the probability assigned to the observed outcome averaged across various forecasts (e.g., weeks, targets, and geographic regions). Used to measure the accuracy of a forecast.

Nowcast

Forecast of current conditions. For example, due to delays in reporting and data accrual, the FluSight forecast for ILI outpatient visits “two weeks ahead” is actually a forecast for the current calendar week.

Onset

The start of sustained disease activity. As a seasonal target for FluSight forecasts, it is defined as the first week when the percentage of visits for ILI reported through ILINet reaches or exceeds the baseline value for three consecutive weeks. No onset is a possible outcome.

Peak intensity

The maximum weekly or monthly value that disease activity reaches. As a seasonal target for FluSight forecasts, it is defined as the highest numeric value that the weighted ILINet percentage reaches during a season.

Peak week

The week that disease activity reaches it maximum. As a seasonal target for FluSight forecasts, it is defined as the week during the influenza season when the weighted ILINet percentage is the highest. More than one peak week is a possible outcome.

Reliability

A measure of how well the forecasted probability of an event occurring matches the observed outcome. Reliability answers the question whether a forecast that assigns a probability of 0.2 observes the forecasted event 20% of the time. This is also known as forecast calibration.

Retrospective forecast

A forecast of a past event (e.g., past influenza or dengue seasons) using data only from time periods prior to the event.

Skill

The average confidence (or probability) that was assigned to the observed outcome.

Seasonal target

Forecasts for the overall influenza season characteristics. These forecasts currently include the onset week, peak week, and peak intensity.

Short-term target

Forecasts for the near-term trajectory of the influenza season. These forecasts currently include forecasts for influenza activity one, two, three, and four weeks ahead from the date of data publication.

Target

The outcome that a forecast is predicting.