Hi everyone,
This time, I thought of discussing possibilities of trading with neural nets. So this thread is mainly for discussion with people that do know a bit about deep and/or reinforcement learning, although all opinions are welcome.
I recently stumbled upon these two very mysterious (obviously) YouTube channels that seem to use some sort of neuro-evolution of augmented topologies (or NEAT for short) for trading purposes. I myself tried to create a small experiment in Python but so far, I did not quite get the results I was looking for.
Basically, the way I laid out the experiment is the following:
Link #2: another interesting video:
Cheers,
m10021
This time, I thought of discussing possibilities of trading with neural nets. So this thread is mainly for discussion with people that do know a bit about deep and/or reinforcement learning, although all opinions are welcome.
I recently stumbled upon these two very mysterious (obviously) YouTube channels that seem to use some sort of neuro-evolution of augmented topologies (or NEAT for short) for trading purposes. I myself tried to create a small experiment in Python but so far, I did not quite get the results I was looking for.
Basically, the way I laid out the experiment is the following:
- I used EURUSD hourly data from past N years (let's say 5), that I split into training (80 first percent of the data) and testing (remaining 20%).
- I considered that at each moment, the agent has the info on the given closing price delta, trend/indicators data, and also its own state information (in or out of position, current P/L, current holding time, etc.). I also tried variants of "handmade RNN structure" where the current input also has info on the past states of the market and the agent.
- I took into account expected spreads and swaps for holding.
- I made a small logic where the input at each time step is as detailed above, and the output is a softmax on four possibilities: buy, sell, hold, close. The output is then interpreted to take action and apply corresponding changes to the agent's performance.
- I then run thousands of generations with different types of settings on NEAT on this experiment, using final P/L as a fitness function (although I tweaked it a bit to force the agent to be active and not just go long because the EURUSD went up, by setting: fitness = P/L * log(#long_trades) * log(#short_trades) for example). Other fitness functions such as Sharpe ratios or custom mean/deviation measures could be interesting here too...
Anyone here tried similar stuff? I'm talking deep/reinforcement learning, genetic evolutionary algorithms, or even stochastic control theory? I'd be super happy to discuss and possibly find out/share interesting things.
Link #1: see the video and the channel for more examples
Inserted Video
Link #2: another interesting video:
Inserted Video
Cheers,
m10021