Contents - Max-Planck-Institut für Physik komplexer Systeme
Contents - Max-Planck-Institut für Physik komplexer Systeme
Contents - Max-Planck-Institut für Physik komplexer Systeme
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
where π and ¯π are probability distributions with πk =<br />
P(y = k|p), and ¯πk = P(y = k), respectively. The first<br />
term is the inherent uncertainty of y—this quantity is<br />
not affected by the forecasts. The second term quantifies<br />
reliability; note that this term is always positive unless<br />
p = π, which is the mathematical definition of reliability.<br />
The third term quantifies the information content<br />
of π. It contributes negatively to (i.e. improves) the<br />
score, unless π = ¯π, in which case π is constant. Hence,<br />
this term penalises lack of variability of π; the larger the<br />
variability of π, the better the score. As a whole, the decomposition<br />
If constructed yields credence from a to finite thenumber practiceof ofsamples, assessingthe<br />
forecast Talagrand quality diagram through is subject properto scoring randomrules. fluctuations. In<br />
-diagram, deviations from flatness due to finite<br />
Ensemble samples forecasts are taken into account. Under the assumption of<br />
Modern consistency,thenumberofcasesinwhichtheverification<br />
weather forecasts are generated using large<br />
dynamical atmospheric should models, followrunning a binomial on supercom- distribution<br />
puters. In order to initialise these simulations properly, is<br />
the current state of the atmosphere has to be known<br />
at least approximately, and subsequently projected into<br />
the state space of the model; this process is known as<br />
data assimilation. The fact that the initial condition is<br />
not known with certainty (and also that the model is incorrect)<br />
is accounted for by generating not one but several<br />
simulations with minutely perturbed initial conditions,<br />
resulting in an ensemble of forecasts.<br />
Although ensemble forecasts already provide vital information<br />
as to the inherent uncertainty, they need to<br />
be post-processed before they can be interpreted as<br />
probabilities. The interpretation of ensembles and how<br />
to generate useful forecast probabilities using ensembles<br />
is a very active area of research.<br />
Several different interpretations exist, a very common<br />
one being the following Monte–Carlo interpretation:<br />
An ensemble is a collection x1,...,xK of random variables,<br />
drawn independently from a common distribution<br />
function p, the forecast distribution. The forecast distribution<br />
p can be considered as the distribution of the<br />
ensemble members conditioned on the internal state<br />
of the forecasting scheme. The forecast distribution<br />
p however is but a mental construct and not operationally<br />
available.<br />
In this interpretation, the forecasting scheme is called<br />
reliable if the observation y along with the ensemble<br />
members x1 ...xK are independent draws from the<br />
forecast distribution p. Less formally stated, the observation<br />
behaves like just another ensemble member.<br />
A necessary consequence of reliability is that the rank<br />
of y among all ensemble members assumes the values<br />
1,...,K +1 with equal probability (namely 1/(K +1)).<br />
This means that the histogram of rank(y) should be flat,<br />
which can be statistically tested, see for example [3].<br />
In operational ensembles though, histograms are often<br />
found to be u–shaped, with the outermost ranks<br />
being too heavily populated (see Fig.1 for an example);<br />
in other words, outliers happen more often than<br />
they should in a reliable ensemble. This can have several<br />
reasons, such as insufficient spread or conditional<br />
bias. On the other hand, this means that we should be<br />
able to predict such outliers by looking at characteristic<br />
patterns in the ensemble. Indeed, as we could show,<br />
even for reliable ensemble forecasts, the spread of the<br />
actual specific ensemble is indicative of the probability<br />
that the future observation will be an outlier [10].<br />
v l<br />
0.999<br />
0.99<br />
0.9<br />
0.5<br />
0.1<br />
0.01<br />
0.001<br />
1 52<br />
Figure 1: Rank diagram for temperature forecasts in Hannover: The<br />
verification falls much too often into the first and last bins, indicating<br />
that outliers are too frequent for reliability. This ensemble features<br />
51 members. The y–axis shows Binomial probabilities, rather than<br />
actual counts.<br />
Unfortunately, rank based reliability tests are restricted<br />
to scalar predictions. In [9], a rank analysis for vector<br />
valued predictions was suggested by measuring the<br />
length of a minimum spanning tree. Thereby, a new<br />
scalar ensemble is created which however ceases to be<br />
independent, that is, the Monte–Carlo interpretation no<br />
longer applies. This puts the assumptions behind the<br />
entire rank histogram analysis into question. However,<br />
as it could be shown in [2] the rank based reliability<br />
analysis can still be applied to such ensembles, due to<br />
some inherent symmetries called exchangeability. In<br />
particular, these investigations demonstrated that the<br />
minimum spanning tree approach is mathematically<br />
sound.<br />
[1] Glenn W. Brier. Monthly Weather Review, 78(1):1–3, 1950.<br />
[2] J. Bröcker and H. Kantz. Nonlinear Processes in Geophysics,<br />
18(1):1–5, 2011.<br />
[3] Jochen Bröcker. Nonlinear Processes in Geophysics, 15(4):661–673,<br />
2008.<br />
[4] Jochen Bröcker. Quarterly Journal of the Royal Meteorological Society,<br />
135(643):1512 – 1519, 2009.<br />
[5] Jochen Bröcker, David Engster, and Ulrich Parlitz. Chaos, 19,<br />
2009.<br />
[6] Jochen Bröcker and Leonard A. Smith. Weather and Forecasting,<br />
22(2):382–388, 2007.<br />
[7] Thomas A. Brown. Technical Report RM–6299–ARPA, RAND<br />
Corporation, Santa Monica, CA, June 1970.<br />
[8] I. J. Good. Journal of the Royal Statistical Society, XIV(1):107–114,<br />
1952.<br />
[9] J.A. Hansen and L.A. Smith. Monthly Weather Review,<br />
132(6):1522–1528, 2004.<br />
[10] S. Siegert, J. Bröcker, and H. Kantz. Quarterly Journal of the Royal<br />
Meteorological Society, 2011 (submitted).<br />
2.15. Probabilistic Forecasting 71