@InProceedings{MaiaMein:2006:DiCoLi,
author = "Maia, Aline de Holanda Nunes and Meinke, Holger",
affiliation = "Embrapa Meio Ambiente, PO Box 69, Jaguari{\'u}na, SP, Brazil
(Maia) and Department of Primary Industries and Fisheries, PO Box
102, Toowoomba, Qld 4350, Australia (Meinke) and {}",
title = "Assessing uncertainty of seasonal probabilistic forecasts:
distribution-free confidence limits",
booktitle = "Proceedings...",
year = "2006",
editor = "Vera, Carolina and Nobre, Carlos",
pages = "569--573",
organization = "International Conference on Southern Hemisphere Meteorology and
Oceanography, 8. (ICSHMO).",
publisher = "American Meteorological Society (AMS)",
address = "45 Beacon Hill Road, Boston, MA, USA",
keywords = "probabilistic forecasts, uncertainty, confidence intervals.",
abstract = "Probabilistic climate forecasts often rely on information coming
from historical climate series of prognostic variables,
represented by cumulative distribution probabilities functions
(CDFs) or their complement, probability of exceeding functions
(POEs). They are a simple and convenient way to represent
probabilistic information arising from a time series that exhibit
no or only weak auto-correlation patterns. However, if the time
series shows moderate to strong auto-correlation patterns, a
CDF/POE summary will result in some loss of information. Yearly
sequences of rainfall data from a specific month or period
generally exhibit only weak auto-correlation, thus allowing the
CDF/POE representation to convey seasonal climate forecast
information. Useful information required by decision makers is
then derived from such distribution and expressed as the
probability of exceeding a certain threshold (e.g. probability of
exceeding historical median value of rainfall or any other derived
quantity such as agricultural yield or income). Such estimates are
frequently reported without any measure of uncertainty. The degree
of uncertainty depends on the length of the time series and its
internal variability. Lack of uncertainty assessments can lead to
misguided beliefs about the true performance of the forecast
systems (e.g. due to the possible existence of artificial skill,
especially if forecasts are based on short time series) possibly
resulting in inappropriate actions by the decision maker.
Parametric methods to assess uncertainty of percentiles and
probability of exceeding estimates are frequently based on
normality assumptions. However, distributions of some important
climate variables, such as rainfall, are notoriously skewed,
particularly in areas with strong seasonality that can result in
high frequencies of zero rainfall amounts. For such cases there
are often no mathematical transformations available that would
overcome this lack of normality. As an alternative for
Normal-based procedures, we propose the use of distribution free
methods for constructing percentile and POE confidence limits.
Those distribution-free tools are particularly useful for spatial
uncertainty assessments that would require a tedious,
location-by-location checking of assumptions regarding underlying
probability distributions. Normal-based and distribution-free
methods are both available in The Capability Procedure of the
Statistical Analysis System (SAS, version 7 and latter releases).
In this work, we discuss the rationale, advantages and limitations
of both, parametric and non-parametric approaches. We illustrate
these methods by assessing the uncertainty of percentiles and POEs
estimates for 3-monthly rainfall series from locations in
Australia and South America. The SAS codes for computing the
uncertainty measures will also be presented.",
conference-location = "Foz do Igua{\c{c}}u",
conference-year = "24-28 Apr. 2006",
language = "en",
organisation = "American Meteorological Society (AMS)",
ibi = "cptec.inpe.br/adm_conf/2005/10.31.10.20",
url = "http://urlib.net/ibi/cptec.inpe.br/adm_conf/2005/10.31.10.20",
targetfile = "569-573.pdf",
type = "Climate predictions",
urlaccessdate = "30 abr. 2024"
}