I have a hard time believing that you have found 50 independent LOB-based features.
List them.
Not everyone here is retail.
This thread begs credulity.
Well, I didn't say the features are independent, there are various ratios, slopes, and velocity and rate of change measurements on the same data - different price levels, aggressiveness of orders, volumes, etc... You can probably extract hundreds of measurements, but of course you have to think about Bellman's "curse of dimensionality" the more features you use the exponentially more data you need to train a model that will generalize.
So my optimization algo is actually reducing the complexity and number of features, so it uses only features it really needs which is probably 20 or so. During training it alternates between complexification and simplification to find the right balance. Feature engineering is bulk of the programming work and part of the secret sauce, so I'm not gonna list everything here. I feel like I told you too much already

P.S. I'm not trying to sell anything, so you don't have to believe anything I say, I just would like to get a contact with someone from a market making/HFT firm in NYC. But now I'm doubting that this kind of people come to this forum
Last edited: