Not too surprising if by "ols" one means chucking all parameters in the regression and calling it a day, despite possible multicollinearity. But this isn't an issue with ols, but rather with parameters introduction algorithms.
I would be much surprised if it turns out that simple methods outdo regression where parameters are introduced piecemeal based on greatest absolute correlation with error from earlier regression (not to mention stepwise and other such tricks)
They did use step-wise regression based on cp, but I wasn't too interested in this outdated approach, so I only used ridge (which won every comparison).
Not too surprising if by "ols" one means chucking all parameters in the regression and calling it a day, despite possible multicollinearity. But this isn't an issue with ols, but rather with parameters introduction algorithms.
I would be much surprised if it turns out that simple methods outdo regression where parameters are introduced piecemeal based on greatest absolute correlation with error from earlier regression (not to mention stepwise and other such tricks)
They did use step-wise regression based on cp, but I wasn't too interested in this outdated approach, so I only used ridge (which won every comparison).