Forums  > Basics  > Unit root test of returns  
Page 1 of 1
Display using:  


Total Posts: 10
Joined: Mar 2019
Posted: 2019-03-31 22:45
Without delving into the details of one of the strategies I employ, it is basically highly reliant on walk forward optimization techniques. In doing this, one of the most important problems to solve has been what metric to optimize for. Sharpe, Sortino, and some other well known methods logically have issues - there was a good study by quantopian recently about their ineffectiveness at predicting out of sample performance - and so much of what I have done is create metrics that serve as a better indication of out of sample performance.

One factor that I have found that has been working very well is find a set of returns without a unit root (constant and trend). I do not use this as the only factor, and I actually optimize for a different metric before incorporating an augmented dickey-fuller test, but this seems to do an excellent job of providing a linear relationship between in sample and out of sample performance.

With walk forward optimization, the problem of overfitting has been serious, but all my results indicate that using a unit root test really helps (I have also had decent results utilizing autocorrelation tests). Does anyone know if there is some metric I am not familiar with that is exactly what I am describing here? Anyways I have found this idea helpful and would be interested to know what some of you think.
Previous Thread :: Next Thread 
Page 1 of 1