Implementation of this paper: Paper
Paper describe how to optimize Sharpe Ratio using deep learning.
Defining Sharpe Ratio as loss function we need to implement custom training.
Also I tried to change the objective from LSTM to directly optimize Sharpe Ratio in LSTM for prediction and optimize Sharpe Ratio through quadratic programming. But Sharpe Ratio is not a convex function, so we have many possibilities:
- Dual problem
- Do some transformations in order to achieve convexity
- Change approach and implement a neural network to maximize numerator(returns) and a neural network to minimize denominator(volatility). This is a min-max game like what happens in Generative Adversarial Networks