One paper that I really like and enjoy reading is the paper by Cheng, Pelger and Zhu (2021), “Deep Learning in Asset Pricing“.
In addition to the smart idea behind the paper and the excellent writing, one thing that I found interesting (maybe even a minor thing in the paper), is that they train their model only once and use it for many years to map firm characteristics to SDF weights, etc. It shows that the relationship between firm characteristics and SDF is almost stable over time. I.e., whatever relationship firm characteristics had in the 1970s with SDF, they have the same relationship in the 2020s.
It is something that I also found in my own research. Even in “Neighbouring Assets” I have had the same conclusion. (Although I found that this stability is stronger when you consider “tiny” stocks.)
However, there is quite a big literature about “data snooping” in the anomalies literature. For instance, McLean and Pontiff (2016) find the return predictability disappears after the publication. Linnainmaa and Roberts (2018) have also the same conclusion and they go even one step beyond that. They find that return predictability not only drops in out-of-sample post-publication but also it is weaker, if any, in pre-sample (the sample before the original studies).
What I am trying to understand now is whether firm characteristics’ relationship with expected returns drops out-of-sample (either post or pre-sample) or they are fairly stable over decades.
I will try to address this question in my next paper …
Leave a Reply