LSTM and Evolutio-Bayesian-Strategy combinantion

[AlconDivino]

My goal is it to make a script that automatically buys and sells cryptocurrency on either binance or bitmex (or both)
apparently with this repo it is possible to do that like huseinzol said in the issues.

I now fail to understand what he means with "use lstm to predict the trend and then let the agent do the decicions"
I thought maybe you can tell me how to do that, or maybe even show me a basic script how its done.

[bitnom]

My goal is it to make a script that automatically buys and sells cryptocurrency on either binance or bitmex (or both)
apparently with this repo it is possible to do that like huseinzol said in the issues.

I now fail to understand what he means with "use lstm to predict the trend and then let the agent do the decicions"
I thought maybe you can tell me how to do that, or maybe even show me a basic script how its done.

Sure. So LSTM is pretty good (Not always the best) at making a forecast. It's good enough though to want to use it. LSTM does not trade as a bot though. He only makes predictions. For the trading part, you need a model/agent that will either optimize for a fitness function or a reward function (Depends on algorithm). One agent that can do this is a Q learner. Genetic algorithms can do it (Fitness) too but Q is probably better. I think I have a Bitcoin Q learner on my computer somewhere. I will try to find and add it to the repo. These agents can use the LSTM predictions to make decisions (Buy/Sell) in order to maximize the reward (User's account balance $$$).

He has an example of an agent already in his repo. That's what the bayesian evolution agent is doing. It is a good agent but we can do better. It's just a matter of how much time we spend on the repo.

[bitnom]

Actually, this repo has a Q-learner. I just remembered that. Both are a way to search the problem space and make decisions to maximize fitness or reward. Same as the Atari Q agent.

[AlconDivino]

Thanks. So basically use the Data the LSTM outputs and feed that to the (for exampe Evolution-Bayesian) Agent and see what he outputs as decicion? I cannot use the native buy() function but have to use the predict() function from the Model class right?

[AlconDivino]

Also would be great if you can include the Bitcoin Q learner in the repo :D