eeng


Total Posts: 23 
Joined: Dec 2014 


How have the factor models been validated historically? I mean, when the momentum or value factors were identified, did the researcher simply run a linear regression on past data, or perhaps some kind of train/test split was performed, to ensure that the newly identified factor was robust to new data?
The background on my query is, due to how the alpha in different factors has eroded over time (won’t go into the issue of whether timing factors is possible), surely a stablished process to detect and favor new factors will be required, but I’m unsure whether this already exists.

quantmatters.wordpress.com 


yanko


Total Posts: 63 
Joined: Nov 2009 


Original papers run a regression on the longest possible time series. If adding a factor pushes the intercept ("alpha") closer to 0 and if that factor's beta is deemed significant, then the factor is real. Yay!
Depending on your choice of factors, you get different results (value is real in a threefactor model, value is dead in a fivefactor model, and it's alive again in a sixfactor model).
So, you'd say: OK, let's crossvalidate. The result is that it's all just data mining:
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2897719
yanko 

