Wednesday, 16 January 2002: 4:30 PM
End-to-end ensemble forecasting: Ensemble interpretation in forecasting and risk management
The individual members of a (Monte Carlo) ensemble forecasts are often
interpreted as single deterministic forecasts; a more coherent
interpretation involves dressing these delta function forecasts before the
prediction system evaluated (against a dressed `best guess' forecast), or
deployed for commercial use. Good probability distributions of future
outcomes provide a crucial tool in risk management. The recent trend
towards probabilistic weather forecasts instead of traditional `best
guess' forecasts should enable businesses with weather risk exposure to
assess and manage this risk with greater rigour.
Quantifying the skill of ensemble weather forecasts as weather forecasts
requires dressing the ensembles before comparison with the observations.
The ideal forecast product for an end-user, however, is not a
probabilistic weather forecast at all but a probabilistic forecast of a
weather dependent economic quantity (e.g. soft drink sales, electricity
demand, wind energy production).
In both applications, the first challenge is to convert the ensemble of
deterministic forecasts into a bona fide probabilistic forecast. This
requires knowledge of the error statistics associated with each ensemble
member. Extracting such statistics is not as straightforward as with a
single deterministic forecasts. In the second application, the
probabilistic forecast must be transformed into a function of the economic
quantity. The dependence of economic quantities of interest on the weather
is generally nonlinear. This means that the expected value of the economic
quantity is not the value associated with the expected weather.
We present several examples to illustrate this `end-to-end' forecasting
approach. Moving away from a simple `cost-lost' scenario,
several examples, including (i) U.K.
electricity demand, and (ii) Wind energy production, are employed to
demonstrate how the ECMWF ensemble forecast product can be translated into
probability functions familiar to risk management professionals.