C3S seasonal description of hindcasts for DGOV discussion

For the seasonal forecast, as we are pretty far from the starting time of the forecast (initial conditions), it's too long to maintain the "memory" of the initial conditions of the atmosphere, so we cannot provide numerical values for the prediction (seasonal forecasts are not weather forecasts). The seasonal forecast provides mainly anomalies concerning the climatology in general, particularly the "model's climate". In order to produce the "climate of the model", we need to re-run the numerical weather prediction model with exactly the same configuration as the real-time forecast in the past. In other words, to re-run the same forecast system used to produce the real-time forecast for several starting points in the past in the same way as a forecast would be run (with only knowledge of the starting point), for the same length of time as an equivalent forecast. In that way, the systematic deviation or bias of the system can be corrected by estimating the systematic errors of the system. In other words, the set of re-forecasts (hindcast) is used to estimate the model's bias in the different variables.

For example 

These sets of retrospective forecasts (re-forecasts or hindcasts) are produced with each version of the forecast system. Every real-time forecast data needs to be "analyzed" with the help of the relevant set of "hindcast" data, meaning there should be an unequivocal way to know a given forecast and the set of hindcasts come from the same forecast system. Therefore a forecast by itself is not useful without relating it to the relevant hindcasts.

As we said, hindcast can quantify only the systematic errors of a system. However; a system can also have random errors. The random errors of the forecast can be estimated by the ensembles and the effect of these errors on the prediction is quantified through the use of ensembles.

In other words, as we need to associate together all the ensemble members, in a similar way we need to associate the real-time forecast with the hindcast (by the version of the system). 

Finally, as well as playing an essential role in the correction of systematic errors, hindcasts are also used to assess the skill of seasonal forecast systems (by comparing each of the forecasts for the years in the reference period with the respective observed conditions). Information on forecast skills is important to avoid overconfident decision-making.

Production schedules

Burst ensembles

Both real-time forecast and hindcast initialize all their members on a given date (e.g. the 1st of the month)

Lagged-start ensembles

Real-time forecasts and hindcasts initialize a number of members over a set of dates (different start dates for the members). Both the number of members and the dates can be different between real-time forecasts and hindcasts

Examples: