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.

In addition to forecasts, we are also preparing scenarios with which to model long term vaccination and immunity.

Please take care when submitting the forecasts to be clear which target and horizon you are using: use this page and the formatting guide.

Targets

Targets

We are currently focused on two forecast targets:

• Weekly new COVID-19 cases
• Weekly new COVID-19 deaths

You can submit forecasts for only cases or deaths, or both 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 forecasts of health system burden indicators, or estimates of infections or reproduction numbers.

Truth data

We evaluate forecasts against John Hopkins University data, and we recommend using this dataset as the basis for forecasts.

• Data are available to download here.
  • Incident cases by country and day are stored in: time_series_covid19_confirmed_global.csv
  • Incident deaths by country and day are stored in: time_series_covid19_deaths_global.csv
• We also download the data into this repository, saved in data-truth.
• JHU also provide country metadata, including population counts and ISO-3 codes.

Note there are some differences between the format of the JHU data and what we require in a forecast.

• Spatial unit: Subnational data are available in some countries, but we require national-level forecasts.
• Time unit: 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.
• Values: JHU data may in some cases differ slightly from individual national datasets, and JHU data can be revised retrospectively. We still recommend using JHU data in forecasting so that we share a common dataset with which to compare across models.
• Location codes: JHU uses ISO-3 codes whereas we use ISO-2 codes. For conversion between the two, you can consult our template or, if using R, the countrycode package.

Horizon and frequency

Horizon

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.

Date format

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 typically require submission by the Monday 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)

Frequency

Forecasts can be submitted once per week only.

• Submission can be at any time between Thursday and Monday (23:59 GMT).
• We recommend submitting on a Monday: this is also the forecast_date.

Location

You can submit a forecast for a single country, or any combination of the 27 EU nations 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.