Individual forecast ensemble

The design and configuration of each ensemble forecast system depend on specific requirements, priority, and resources in each contributing centre. The full description of each system configuration is available here. For example, each system comes with a set of reforecasts covering different periods and with different ensemble sizes. This variety of system configuration presents a big challenge.

To create consistent individual forecast products and multi-model products we need to use model climate estimates based on a set of reforecasts covering a fixed period of 10 to 20 years common to all the models. Unfortunately, reforecasts for a common and sufficiently long set of years is not always available. As a compromise we use the longest common period currently available covering 1999-2010. However, for some systems, we are forced to use even a shorter climate period. This approach might be revisited in due course. (see Table).

Individual and multi-model combination are constructed using the forecast initiated around each Thursday. For systems with ensemble size smaller than 8 members we consider forecast initiated 1 or 2 days earlier depending on the system configuration. Although not ideal, for Bejing system we combine the forecast of Thursday and previous Monday. It is worth noting that for Bejing system only 3 years of common reforecast period are available (see Table). The method used to create the forecast for Bejing is going to be revised in due course and a different climate period will be used. The entries in parenthesis indicate that, for some systems, the method to sample the model climate has been recently revised.

It is important to note that for the systems with reforecast produced on the fly and covering just the most recent years, the use of a fixed common period is not always possible. This is the case of Bejing system which (at the time of writing) has only 3 years of reforecast for the common period 1999-2010 (see here). Because of the above, the Beijing forecast is defined as the departure from the model climate estimated by the whole reforecast set covering the most recent 15 years.

The table below shows, for each model configuration, the forecast and re-forecast data used to construct the forecast charts.

Multi-model forecast ensemble

Multi-model forecasting systems are a simple but remarkably powerful way of enhancing forecast reliability. Different forecasting systems have different errors. If we combine different forecasts, the effects of some of the model-specific errors are averaged out, giving a better estimate of the most likely outcome. At the same time, the differences between models give a measure of the system-related uncertainty in the forecast. Multi-model systems are not perfect – some errors can be common to all models, for example. Nonetheless, they tend to give forecasts which are both more accurate and more reliable.

The multi-model forecast is just a simple combination of all contributing forecast available at the time. The mean of the multi-system ensemble is an unweighted average of the ensemble means of all available contributing systems (each calculated as described under ‘individual forecast ensemble’). Multi-model probabilities are calculated as the unweighted mean of the probabilities from the contributing systems.