The examples of filter and meta-labeling are cheating about mlfinlab HOT 3 OPEN

I don't know if it cheating but yeah that's the main issue with their framework which seems very well made overall but in order to use classification you need to know a-priori before the training at which time to enter in position and there is no proposal to do that with an ml model. I think that is the main missing part of their system to get a fully build ml architecture, A first model for events, a second for side and a third for bet sizing (metalabeling)

from mlfinlab.

The most important thing is the label .
If you know which point profit most and label it, build a relative good model is pretty easy.
They provide a lot of labeling methods all have lookahead bias, I don't think they don't aim to cheat .

Maybe the valueable part is only metalabeling, it does help bet sizing, but that's all, it can't bring outperform to moden model(LGB/XGB/DNN) .

from mlfinlab.

But for classifiers it is necessary to set right labels before the training that comes from the constraints of the algo itself. Maybe one way to solve that is to voluntary decrease the amont of the accuracy of labels like they did for backtesting to find optimal rules in one of their video, but the model will learn on fake labels and that will false his learning and predictions.

We can also view this problem from the other side if you set labels on events that are not proof to be particularly relevant it will miss opportunities that the model could have find otherwise.

Another way is to do the opposite, label all the market at each point with just the trading rules and let it filter events itself instead of outsourced predefined filter that could false it in both ways.

from mlfinlab.