I currently have a basket of 200 investment portfolios which are partitioned into 3 styles: Long, Intermediate, and Short. Each portfolio has 10 interesting features X which define its style. I want to create a system whereby I can identify portfolios that are out of style and that that need some work to get them back into style.
My thought is to use a probabilistic framework to solve this problem. This will reduce my 10 interesting features down to a single number which will determine whether or not the portfolio needs attention. To this end I'm currently creating a binary classifier on each style using a neural network with a sigmoid activation function to estimate probabilities. If the probability that a portfolio belongs to a certain style is less than 75%, I will transact to bring the portfolio back into style.
Is this the best framework for an anomaly detection system given what I am trying to accomplish? A multinomial classifier could be another option, but it is quite similar the simple binary classifier I described above.
Best
For anyone curious, I ended up using a one-sided support vector machine. This I decided was best as I had very few negative examples to work with.
Best