Summary
International Symposium on Nonlinear Theory and its Applications
2005
Session Number:3-4-4
Session:
Number:3-4-4-3
Prediction, Behaviour, and Ignorance
Devin Kilminster, Reason Machete,
pp.654-657
Publication Date:2005/10/18
Online ISSN:2188-5079
Summary:
In an "imperfect model scenario", the choice of values for the model’s parameters is not merely a matter of "estimation", rather we must "fit" the parameters according to some criterion for the performance of the model. For time-series, one such criterion is that of prediction. A "predictionist" model is fit so as to optimise for short term prediction. Predictionist models, however can fall short in other ways ? a model that makes very good short term prediction might fail to have reasonable long term behaviour when "free-run", for example. Interestingly, though, "behaviourist" models (ones which optimise a free-running behavioural criterion) often perform quite robustly when considered as predictors. We investigate these issues in the context of "ignorance" ? a score for measuring the performance of predictors. We develop a behavioural analog of ignorance and derive a few interesting connections between the two scores.