After 10 days of coupled integrations, the model drift begins to be significant. It displays similar patterns to seasonal forecasting after 6 months of integrations, but with less amplitude. The strategy for dealing with model drift is straightforward. We initialize the ocean, atmosphere and land surface to be as close to reality as possible, and calculate the forward evolution of the system as best we can using numerical approximations of the laws of physics. No "artificial" terms are introduced to try to reduce the drift of the model and no steps are taken to remove or reduce any imbalances in the coupled model initial state: we simply couple the models together and start to integrate forward. The effect of the drift on the model calculations is estimated from previous integrations of the model in previous years (the re-forecast). This drift estimate is removed from the model solution during the post-processing.

One motivation for creating a model climatology is that after about 10 days of forecasts, the spread of the ensemble is very large (see, for instance, forecast plumes). Therefore, the probability distribution function (pdf) of the model climatology needs to be evaluated, in order to detect any significant difference between the ensemble distribution of the real-time forecast and climatology.

In the present system, the climatology (re-forecast). is based on an 11-member ensemble of 46-day ENS integrations, starting on the same day and month as each real time forecast for each of the past 20 years. For instance, if the first starting date of the real-time forecast is 27 March 2013. The corresponding climatology is a 11-member ensemble starting on 27 March 2012, 27 March 2011, ..., 27 March 1993. The 11-member ensemble is thus integrated with 20 different starting dates. This represents a total of 220 integrations.

The re-forecasts are created twice a week (Mondays and Thursdays) and are ready a week in advance.

Real-time extended range forecasts are actually calibrated using a 1- week window of re-forecasts, which represents a total of 660 (3 start dates x 20 years x 11 members) re-forecast integrations. For a Monday run we calibrate using the date of that Monday plus the dates of the Thursdays immediately before and after. Likewise Thursday runs use the date of that Thursday and the two Mondays either side.

Real-time medium range forecasts also need to define the climatology for some purposes (for example for the EFI, SOT and '15-day meteogram with climate' products). These actually use a 4-week window of re-forecasts (9 start dates x 20 years x 11 members = 1980 integrations), again centred around the data time of the forecast in question. This larger window brings the benefit of better defining the climatology tails - i.e. the more extreme values - which is important for consistent computations of EFI and SOT.