David Spiegelhalter (Cambridge)
Abstract: Classical model selection focuses on the ability to fit the data in hand, generally with some form of penalty to prevent over-fitting. Bayesian model uncertainty attempts to identify a plausible model that gave rise to the current observations, which can then be used for predictions. In each case the future is assumed to arise from essentially the same process that gave rise to the past: in particular Bayesian uncertainty statements are conditional on the 'truth' of the assumed model or set of models. But since 'all models are wrong', these statements inevitably understate reasonable uncertainty. Recent experience with financial models, and disputes about climate change, show that model uncertainty may be more than a technical statistical issue. We shall look at how model uncertainty is dealt with in a range of areas, and in particular examine whether any account is taken of the almost inevitable 'discrepancy' between any formal representation and the actual events that occur.