Skill assessment for ENSO using ensemble prediction

A crucial component of any prediction system is the ability to estimate the predictive skill of a forecast so that the degree of confidence that can be placed in an individual forecast can be assessed. In this paper we have used ensemble prediction techniques to develop a means of estimating a priori the predictive skill of forecasts of El Niño Southern Oscillation (ENSO) using an intermediate coupled ocean-atmosphere model. Each member of an ensemble forecast is perturbed using either noise forcing fields or perturbations that are known to increase the low-frequency variability of the coupled model. These are respectively the so-called stochastic optimals and optimal perturbations of the coupled system; they are added to the model to mimic the presence of initial-condition errors and high-frequency stochastic noise and their effect on the predictability of the coupled system in the tropics. By performing ensemble predictions in hindcast mode, we have identified a usable relation between the skill of a model hindcast and the spread of the ensemble measured relative to some control hindcast. The practical nature of ensemble prediction is demonstrated by computing the relationship between model skill and the spread of an ensemble from hindcasts of ENSO for the period 1972-86, and then comparing the actual hindcast skills for the period 1987-93 with those suggested by the skill-spread relation from the preceding period. The relationship that we find between the model skill and spread of an ensemble appears to be robust in the sense that it is relatively insensitive to variations in the ensemble prediction procedure and to changes in some model parameters. However, crucial to the success of the ensemble predictions is the use of perturbations in the ensembles that are known to increase the low-frequency variability of the coupled model. These perturbations can efficiently probe the probability density function of possible coupled-model states. If randomly chosen perturbations are used in the ensembles, however, no practical relation between model skill and the spread of an ensemble emerges. The relationship identified here, between model skill and the spread of an ensemble prediction, offers a practical means of estimating the confidence that we can place in future forecasts of ENSO using the same coupled model.