Abstract:
|
Prediction with models specified through information external of the data will potentially suffer losses due to misspecification. In turn, loss functions dependent on correct model specification (eg. expected square loss) will also become unrepresentative. Misspecified loss functions lead to inaccurate predictions as well as the inability to accurately provision for a wrong prediction. The effect is particularly significant at early stages of knowledge discovery during which no "experts" exist to provide reliable external justification for model specification.
In this talk, we will explore the approach of avoiding model specification not rigorously justified in favor of assumptions which are driven by rational principles. Under this framework, we begin with the single assumption that the underlying data generating process and its state space are stationary and can, in principle, be found. In partial answer to the problem posed, we will motivate the multinomial distribution as an alternative data driven model, and explore how to evaluate losses which are consistent with model criticism and doubt induced by "black swan" observations.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.