[optimiz3]

What signals is the AI bot using? Just a mix of the TA indicators?

I ask because price forecasting is a data modeling problem, and I'm skeptical of the bot's listed win rates. Unless one knows how their signals work (i.e. what real world market phenomenon you are modeling), there's no way one should trust that edge.

Separately, if you breakdown other indicators like CCI you realize they are crappier versions of basic statistical methods like Z-scoring. Or that anything using a Wilders average is using a crappier EWMA, and then that anything using EWMA can be improved to be stationary by filtering out noise outside the passband of information you want to target.

So I'm skeptical of the bot.

[stocknear]

I am trying to build a robust and reliable open-source model for the AI Bot. The challenges that you are mentioned are definitely one of the biggest problems.

However, I think we have to reformulate the question. Instead of maximizing the win rate (which relies on forecasting the future price) the real task would be minimizing risk (which does not rely on forecasting the future price instead adapting to the new environment).

[stocknear]

Yes TA indicators a part of the feature vector space. However, so far the model uses 200+ features where I pick the best 10 features that correlates the most with the target goal.

In the next version of the model I use more sophisticated features and architecture of the model to increase the win rate and minimize the loss.

[jzombie]

Are you training a single model that's supposed to represent the entire market?

[stocknear]

No,

I train each stock on a separated model to not interfere and make the dataset "dirty".

Each stock has its own dynamics thats why i split each model

[jzombie]

Yeah, that's how I thought it should be. Sounds like a lot of models you're building.

Do you tune each one separately as well or do you find a base tuning and go w/ that?