Hodges' Superefficiency and Oracle Property in Model Selection

Abstract: 

This work introduces a new type of Hodges' estimators that can serve model selection with desired oracle and other properties. It demonstrates that finding an oracle model selection procedure with certain desired properties is in fact not a difficult exercise.  Furthermore, by analyzing the performance of this type of oracle estimators in finite sample sizes, it is theoretically justified that such oracle estimators can perform poorly at some values of the parameters to be estimated. As a result, an oracle model selection procedure does not guarantee to produce superior estimates over the maximum likelihood or least square methods, contrary to the claims or belief in the mainstream of the statistical literature regarding the merits of the oracle property in model selection. 

The seminar is followed by wine and cheese in the staff room RC-3082. All attendees are welcome!