...
Seasonal forecasts are started from an observed state of (all components of) the climate system, which is then evolved in time over a period of a few months. Errors present at the start of the forecast (due to the imprecise measurement of the initial conditions and the approximations assumed in the formulation of the models) persist or, more often, grow through the model integration, reaching magnitudes comparable to that of the predictable signals. Some such errors are random; the effect of these on the outcome is quantified through the use of ensembles. Some errors, however, are systematic; if these systematic errors were determined, corrections could be applied to the forecasts to extract the useful information. This is achieved by comparing retrospective forecasts (reforecasts or hindcasts) with observations. The same forecast system is run 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. The resulting data set constitutes a 'climate' of the model, which can then be compared with the observed climate of the real world. The systematic differences between the model and the real world - usually referred to as biases - are thus quantified and used as the basis for corrections which can be applied to future, real-time forecasts. Given the relative magnitude of such biases, some basic corrections are essential to convert the data into forecast information - therefore a forecast by itself is not useful without relating it to the relevant hindcasts.
...