FY 1986

Simulation data sets for testing MOS (Model Output Statistics) prediction methods

Preisendorfer, R.W.

NOAA Tech. Memo. ERL PMEL-65, NTIS: PB86-173911/XAB, 53 pp (1985)

It is shown how to construct a set of time series with prescribed autocorrelations and cross correlations which will serve as simulators of real fields drawn from Nature for use in testing methods of prediction. These constructions are applied in particular to the problem of evaluating the skill of prediction methods in the context of a model output statistics (MOS) framework. In an MOS framework there are three main ingredients: (i) a model (e.g. an atmospheric general circulation model) that produces a set of predictor time series called the model predictors, (ii) an observer (a human or an instrument) that produces the observed predictand, and (iii) a prediction method (usually a set of statistical algorithms) that forecasts the predictand given the predictors. The distinguishing feature of an MOS framework is that the observed predictands are not in the list of outputs of the model. It therefore falls upon the prediction method to link the model predictors and the observed predictand during a training period for the method. Then, when fresh realizations of the predictors are produced by the model, the prediction method will produce its forecasts of the predictand. The skill of the prediction method is determined by comparing its forecasts with subsequent estimates of the predictand by the observer. The model and observer in (i) and (ii) above are generally imperfect and these imperfections find their way into the skill scores of the prediction method--itself an imperfect instrument in practice. To help sort out the various contributions of these three types of imperfection to the final skill scores of the prediction method, we use the time series construction techniques mentioned above to produce controlled simulators of the real predictor and real predictand fields, and we also develop controlled simulators of the fields produced by the model and observer. Hence in the simulation of an MOS prediction setting there are basically these four fields to generate and interrelate. In such a controlled experimental setting a prediction method's inherent and apparent skills and its robustness to changing prediction conditions may be measured and studied.