![Hier klicken, um den Treffer aus der Auswahl zu entfernen](images/unchecked.gif) |
Titel |
Skill Scores for Probabilistic Climate and Weather Prediction |
VerfasserIn |
T. Maynard, E. B. Suckling, L. A. Smith |
Konferenz |
EGU General Assembly 2012
|
Medientyp |
Artikel
|
Sprache |
Englisch
|
Digitales Dokument |
PDF |
Erschienen |
In: GRA - Volume 14 (2012) |
Datensatznummer |
250068585
|
|
|
|
Zusammenfassung |
A wide variety of skill scores are in use for the evaluation of probability forecasts. While the
importance of using proper scores is well recognised, researchers often face requests to
present results under a variety of different scores. Is there any sense in which considering
many “different" skill scores makes a case more (or less) persuasive? Which set of scores
makes the most persuasive case?
A number of scores are considered, and their strengths and weaknesses contrasted,
including the most commonly used metrics such as the proper linear score, continous ranked
probability score and logartithmic score, amongst others. The role of moving from ensembles
of point forecasts to continuous probability distributions is discussed. Illustrations with
meteorological forecasts are considered, as are mathematical experiments where the size of
the forecast-outcome archive can be made arbitrarily large. The aim is to restrict the number
of skill scores considered, reduce the use of misleading scores, identify independent
evidence and ensure that the set of scores published are specified before the results are
in.
I J Good’s logarithmic score is known to be the only local, proper score for continuous
variables. Proposed criticisms of locality are considered and rejected and it is argued that
proper, non-local scores tend to include a term that is independent of the probability mass
associated with the outcome. Thus they reward the forecast for some other property of the
distribution function itself, having for example a particular shape; it is conjectured that this
may be a general result. This is depricated, except in cases where the forecaster is being
rewarded for something other than the relation between the forecast and reality, or whose
distribution functions are unavoidably constrained by some particular structure.
Remaining challenges to the evaluation of probability forecasts are noted. In partucular
the effect of small samples and the implications of uncertinaty in the outcome. |
|
|
|
|
|