Research into 13F Holdings
beginning feature engineering, calculates turn over based on known data as a proxy for actual fund turn over.
Dynamic Time Warping, attempted temporal clustering of funds with kmeans.
Results were poor. Some obvious clustering mostly near zero.
As I've sat on this problem for a bit, while DTW seemed novel, it was much too naive. The intention again, is to identify similar funds but with slight change in attributes. For instance, personally, I want funds that respond similar to BRK, long holding periods and high concentration BUT I want them with smaller cap funds than BRK could buy.
I think this is a high-dimension sparse recommendation problem. Not having experience with that, will spend some time researching the Netflix-Problem and see what SOTA is and how best to attack this.