Speaker: Phil Dawid (University of Cambridge)
Abstract: We consider modelling a stream of data arriving sequentially, but making no assumptions of independence, identical distribution, etc. Many standard concepts such as parametric consistency are not wholly adequate in this context, and need to be replaced by suitable predictive versions. We describe the concept of prequential (predictive sequential) consistency, and show how it holds more generally than parametric consistency. Extension is made to deal with misspecified models, and connections are drawn with statistical learning theory, and prediction with expert advice.
Screen reader support enabled.