HR: 14:10h
AN: A13C-03 [Abstracts]
TI: Towards Defining Probability Forecasts of Likely Climate Change
AU: * Smith, L A
EM: lenny@maths.ox.ac.uk
AF: CATS, London School of Economics, Department of Statistics
Houghton Street, London, WC2A 2AE
United Kingdom
AU: Allen, M R
EM: m.allen1@physics.ox.ac.uk
AF: AOPP University of Oxford, Parks Road, Oxford, OX1 1DW
United Kingdom
AU: Stainforth, D A
EM: d.stainforth1@physics.ox.ac.uk
AF: AOPP University of Oxford, Parks Road, Oxford, OX1 1DW
United Kingdom
AB:
There is strong desire for probabilistic forecasts of climate change, both
for policy making and risk management, as well as scientific interest. The extent to which this desire can be satisfied
scientifically is unclear. The aim of
this paper is to explore (i) current methods for extracting probability
forecasts and (ii) alternative deliverables which have a firm scientific
basis given the current limitations to the state of the art.
Even ``physics-based" models contain empirically determined paramters and parameterizations. While it is straightforward to
make `ensembles' over initial conditions, parameter values and even several model structures, the interpretation of the
resulting ensemble of simulations requires some care.
Methods for extracting probability forecasts from ensembles of model simulations will be discussed in terms of their
relevance and internal consistency, a particular example being provided by Murphy et al (Nature, 2004). This approach will be
compared and contrasted with one proposed by climateprediction.net (Stainforth et al, Nature, in review), which strives to
produce policy relevant information when no coherent probability forecast can be extracted from the limited ensembles
available in practice. The role of the Bayesian paradigm will be considered in both cases. Extracting probability
distributions in the context of climate change requires consideration of a discrete sample drawn from a high
dimensional space. The analysis of small ensembles requires further
assumptions of linearity and smoothness which must be verified explicitly;
even when ``large" ensembles are to hand, the analysis of the collection
of simulations requires combining mutually exclusive runs, sampling a
restricted region of the parameter space, under a set of models with
related shortcoming. Historical observations serve to lift some of these
difficulties, but attempts to fold observations into the analysis (say, in
terms of weighting sets of simulations differently) are complicated by the
fact that we are trying to model climate (a distribution) while we have
only one realization of the earth system, and that we are extrapolating
into regions where we have no empirical data (in CO2 concentration, for
example). We conclude that while new ensembles of model simulations
provide significant insights into the future of the earth system, current
attempts to extract physically relevant, objective probability
distributions are not coherent. Nevertheless, policy and risk management
decisions can be better informed through a coherent interpretation of
these simulations, which also provide useful guidance toward the design of
improved climate studies.
DE: 3339 Ocean/atmosphere interactions (0312, 4504)
DE: 3344 Paleoclimatology
DE: 3200 MATHEMATICAL GEOPHYSICS (New field)
DE: 3210 Modeling
DE: 3220 Nonlinear dynamics
SC: Atmospheric Sciences [A]
MN: 2004 AGU Fall Meeting