[iliicit]

i've been trading crypto in large volumes at high frequencies for quite some time now. my models were plain as yogurt feed-forward neural nets. i would engineer some dumb features, sample the at random data, assign the labels (that translate into trading decisions), and train the model. then push to prod, sit back, and relax while the balance grows like a mushroom cloud. just kidding, before that i would grow gray hair while backtesting, debugging issues, etc.

one of the hard problems was labeling the data. knowing that the price is going up 10 bps one minute from now, should i buy? maybe. but what if it's going to crash 100 bps right after this? probably should sell instead.

reinforcement learning promises to eliminate the need to assign labels in the training data. the agent will try a bunch of different variants at random and eventually will choose the most optimal one knowing the state of the working, i.e. the state of the markets. at training time i only need to feed it the features data. another benefit is that backtesting and model training is sort of fused into a single process. rl model is optimizing pnl, and not the label classification score (as in the nn model). with proper train-test-validation split, the most performant rl model can go straight into production (helping me to keep some of my hair brown)

while all the bits and pieces seem straightforward i never managed to tune rl model to work better in the backtest compared to the good old old nn models. maybe i have never been closer to the gold vein, but for now, i abandoned my efforts to build a performant rl agent if favor of nn models.

amen.