Definitely worth a read: https://alphaarchitect.com/predict-factor-returns/.
Primary takeaways from my point of view:
- The reason machine learning models may outperform vanilla ranking systems in some instances is that they take advantage of factor momentum.
- Factor momentum is pretty simple: factors that have worked in one month tend to keep working the next month.
- Successful use of factor momentum requires replacing 1/3 to 2/3 of all factors each month. That may be why machine learning models tend to have much higher turnover than ranking systems using the same set of factors.
- One can develop a factor momentum system without recourse to machine learning tools and be equally successful, though on Portfolio123 this is quite difficult to accomplish.
I tried using factor momentum prior to Portfolio123's introduction of machine learning models and found it extremely expensive to implement due to the high turnover. It was a very unprofitable experiment.