Hi,
do some of you, when you perform an out-sample test, have a preference toward the walk forward vs. the cross validation? If yes why (other that the relative longer out-sample period with the cross validation)?
To be sure that we talk about the same thing, I take an example. We have 15 years data (1985-1999). For the walk forward, I will typically take 5 years for the in sample period (1985-1989) to calibrate the model and test out-sample on the following year (1990). Then I will roll the in-sample (1986-1990) and test on 1991, so that at the end we have 10 years out-sample.
For the cross validation, I will hold out each time one year for the out-sample test (ie 1995) and calibrate the model on the 14 remaining years (1985-1994 + 1996-1999 have to be linked). At the end we have 15 years out-sample (each year performance and signals being calculated with different parameters of the model).
I have been using the walk forward for a while without thinking about the cross validation. Now that I am coming accross this alternative, I cannot see any argument against it.
Alex
walk forward or cross validation?
Out-sample results
Hi,
any feedback on this previous post?
Alexandre
any feedback on this previous post?
Alexandre