Trying to evolve a neural network population maximizing return and minimizing MDD using genetic algorithms. Using direct encoding of NN parameters (weights and biases).

Simple architecture with a recursive (vanilla RNN, GRU or LSTM) module followed by a linear layer with the final tanh activation. Output $> 0.5$ results in a Buy signal, output $< -0.5$ results in a Sell signal. Buying is for the minimal possible order size, Sell — closes the whole position. Testing on 1s spot data for simplicity (all orders are Market orders).

After initial experiments with a pure Python implementation, it was decided that the performance is terrible. At first, I thought to rewrite everything in Rust using some NSGA-II Rust implementation, but could not find any having pymoo-like low-lever interface. So I decided to continue using pymoo, but rewrite the computationally-intensive part in C++, providing simple ctypes-compatible interface.