Thanks for the reply.
In hindsight I should have read your initial post about TC more carefully.
I think I was under the wrong assumption that your optimizer would be selecting the stocks for the final portfolio prior to looking at the meta model predictions and then apply the calculations based on the reduced set of stocks. In this scenario it made intuitively sense to me that being correlated to the target under as many viewpoints as possible (high FNC), would lead to a high TC, provided the prediction is unique enough.
However, it seems to me that a high exposure to a particular feature can lead to the optimizer systematically dropping out those predictions you were correlated to begin with, leaving basically some uncorrelated or even systematically anti correlated predictions. Which seems to be the better reason for the importance of FNC.
The real reason I was under the assumption TC is noisy is because one of my models submits random numbers and has a wide range of TC results. While the averaged TC is actually negative, I was not expecting it to provide any “contribution” in any round.
Is there a timeframe defining “a while”? I have other models that seem to have rather inconsistent TC results, like having 9 rounds of positive TC and then a long streak of negative TC, eating up the cumulative TC from the previous 9 rounds.
I really like the essence of TC, and obviously I am still experimenting, but somehow it feels a little bit like driving with no dials and a muddy windscreen, having only the rearview mirror available.