final_program_abstracts[1]

11 IMSC Session Program
Probabilistic forecasts of future climate based on an objective
Bayesian approach using Jeffreys' Prior
Monday - Poster Session 9
Steve Jewson 1 , Dan Rowlands 2 and Myles Allen 2
1 Risk Management Solutions
2 Oxford University, UK
Bayesian statistical methods can be divided into ‘subjective’ and ‘objective’
approaches, the difference being whether the prior is based on judgement, or on some
kind of rule. Several authors have used subjective Bayesian statistics to convert
numerical climate model runs into probabilistic forecasts of future climate. We
believe that there are three particular weaknesses of such an approach. Firstly, that
since the prior is based on judgement rather than on repeatable scientific methodology
it is perfectly reasonable to disagree with the resulting predictions. This is fine if the
predictions are for the personal use of the modellers, but less appropriate if the
predictions are to be used to try and achieve wider consensus about likely future
climates. Secondly, predictions based on subjective priors can never be subjected to
credible evaluation by hindcasting. Thirdly, the prediction produced from such an
analysis cannot be usefully compared with predictions from other models, since any
differences may simply be due to differences in the judgements included in the prior.
To avoid these problems we believe that it would also be useful to produce
probabilistic climate forecasts from numerical climate models using objective
Bayesian statistics. In particular, we believe that it would be useful to use the (nonlocation
version of) Jeffreys' Prior, since (a) it is the most commonly used and most
widely discussed objective prior, (b) it has various attractive mathematical properties,
and (c) it is already used in probabilistic predictions of future climate based on pure
statistical models. We argue that the calculation of Jeffreys’ Prior for complex
numerical climate models can be made feasible by making the reasonable assumption
that model errors are normally distributed. Based on this assumption we show how
expressions for the Jeffreys’ Prior can be derived, and present some results from
application to very simple climate models. We discuss some of the issues that are
likely to arise when applying Jeffreys’ Prior to fully complex numerical climate
models, discuss whether similar ideas could or should be applied to seasonal and
decadal forecasts, and mention some aspects of experimental design, including the
possible benefits of using initial condition ensembles of size one.
Finally we propose that the time has come for seasonal forecasts, decadal forecasts,
and climate predictions to be tested using rigourous out-of-sample hindcasting. To
achieve that the model parameters should be fitted using either maximum likelihood
or (preferably) objective Bayesian inference. We believe that computer power has
reached the point where the complexity of the models should no longer be an excuse
for not testing them in this way. Such testing would yield invaluable information
about model skill, would increase the confidence in projections of the future, and
would open the door to the development of much better methods for calibration and
model combination.
Abstracts 48