Model Averaging Is Asymptotically Better Than Model Selection For Prediction

Tri M. Le

Link To Paper: https://jmlr.org/papers/v23/20-874.html

Journal Of Paper: JMLR

Confirmed Open Access: Yes

Topics From Call For Papers: Bayesian Optimization for AutoML.

Broader Impact Statement On Ethical And Societal Implications:

Key achievements from our paper are 1) providing inequalities showing the error of model averages is less than model selection, 2) the development of comprehensive expressions for stability/complexity for commonly occurring predictor classes, 3) assessment of the performance of such comprehensive expressions, 4) theoretical justifications, 5) application of the new expressions to assess whether predictors are reliable and neither over stable nor under stable.

Societal impact: developing ways to assess performance and stability of predictors and developing predictors for empirical prediction can be used with agronomic data to help ensure that the best varieties of species of grains, for instance, are effectively found and efficiently grown. The use of traditional agronomic data is a good starting point for developing principles and techniques that may be applicable to the vastly more complex agronomic data that has been collected in recent years.