-
Notifications
You must be signed in to change notification settings - Fork 94
Targets and horizons
There is no obligation to submit forecasts for all suggested targets or horizons, and it is up to you to decide which you are comfortable forecasting with your model.
Please take care when submitting the forecasts to be clear which target and horizon you are using: use this page and the formatting guide.
We are currently focused on three forecast targets:
- Weekly new COVID-19 cases
- Weekly new COVID-19 hospitalisations
- Weekly new COVID-19 deaths
You can submit forecasts for only cases or hospitalisations or deaths, or multiple targets.
We only use incident forecasts: the count of new cases or deaths per week. We do not use cumulative (running total) forecasts. This varies from other forecast hubs.
In future we might add other targets such as estimates of infections or reproduction numbers.
We evaluate forecasts against Johns Hopkins University data (for cases and deaths) and data collated by ECDC from national health authorities (for hospitalisations), and we recommend using these datasets as the basis for forecasts.
Data are available in the data-truth directory.
Note there are some differences between the format of the source data sets and what we require in a forecast.
- Spatial unit: Subnational data are available in some countries, but we require national-level forecasts.
- Time unit: Source data are reported daily, but we require weekly forecasts.
- Counts: JHU data are cumulative, but we require a forecast of the new (incident) count over the time horizon of each target. We derive incident data by taking weekly differences of the cumulative data.
- Values: JHU data may in some cases differ slightly from individual national datasets, and JHU data can be revised retrospectively. We still recommend using the data provided for forecasting so that we share a common dataset with which to compare across models.
We do not use or evaluate against these data, but the following might be useful for modelling targets:
Data | Description | Source | Link |
---|---|---|---|
Vaccination | Number of vaccine doses distributed by manufacturers, number of first, second and unspecified doses administered | ECDC | Data on COVID-19 vaccination in the EU/EEA |
Variants of concern | Volume of COVID-19 sequencing, the number and percentage distribution of VOC for each country, week and variant submitted since 2020-W40 | ECDC | Data on SARS-CoV-2 variants in the EU/EEA |
Testing | Weekly testing rate and weekly test positivity | ECDC | Data on testing for COVID-19 by week and country |
For forecasts, we focus horizons of between 1 and 4 weeks ahead. Teams can submit forecasts for any combination of horizons up to 20 weeks ahead but caution that teams should only submit forecasts for horizons that they are happy to be compared to truth data, i.e taken as true predictions. We suggest that teams focus on the period up to 4 weeks ahead.
Forecast horizons should use the Epidemiological Week (EW) format, defined by the US CDC. Each week starts on Sunday and ends on Saturday. For example:
- A 1 week ahead case forecast submitted on Monday 4th January 2021 should cover the seven days including Sunday 3rd January to Saturday 9th January. In this example, the forecast file would include:
forecast_date = "2021-01-04"
target = "1 wk ahead inc case"
target_end_date = "2021-01-09"
- A 4 week ahead case forecast submitted on Monday 4th January would include the period Sunday 24th January to Saturday 30th January 2021. In this example:
forecast_date = "2021-01-04"
target = "4 wk ahead inc case"
target_end_date = "2021-01-30"
Note that weeks start on a Sunday, but we require submission by the Monday evening of that week.
Date conversion helpers include:
- Packages in R: lubridate, MMWRweek
- Packages in Python: pymmwr, epiweeks
- We provide a csv template with the start and end dates for one-week-ahead forecasts at the respective forecast dates (which we assume to be Mondays, in line with the weekly submission deadlines)
Forecasts should be submitted once per week only.
- Submission can be at any time/day of the week.
- We recommend submitting on Monday: if the forecast date is on a Saturday, Sunday or Monday then the “1 week ahead forecast” is for the week ending the Saturday following. If the forecast date is any other day of the week, then the “1 week ahead forecast” is for the week ending the second Saturday following. For example, for a forecast made on 30 July, 31 July, 1 August or 2 August, the “1 week ahead forecast” is for the week ending 7 August, and for a forecast date of 3-9 August the “1 week ahead forecast” is for the week ending 14 August. Whilst submitting on a weekday other than Monday therefore puts teams at a disadvantage in terms of the information available, we do accept forecasts on any other dates, and we will monitor whether the submission date makes any difference in the quality of forecasts.
You can submit a forecast for a single country, or any combination of the target countries (EU, EFTA and the UK). We use ISO-2 codes to identify countries.
Currently we do not accept sub-national forecasts. Sub-national country datasets can be incompatible with each over and over time, making comparisons between forecasts unclear.
- Preparing to submit
- Forecasting
- Submitting
- Other
- Forecast evaluation and ensemble building
- Editing the website
- Creation of an equally weighted ensemble forecast
- Merging weekly submissions