|
Abstract:
|
In this talk, we propose a new predictor by using a median version of stacking in a regression context. This leads us to compare the predictive performance of Bayes model averaging (BMA), stacking and our new median stacking in some scenarios. Preliminary studies suggest that when the true model is on a model list, BMA is better than stacking or median stacking. By contrast, when the true model is not on the list, we suggest that both median stacking and regular stacking can outperform BMA. Moreover, the comparison between stacking and median stacking seems to also depend on the tails of the error term. When the tails are heavy, median stacking seems to give better predictive performance than regular stacking. However, when the tails are light, we get a reverse result.
|
