This trading strategy is built on the concepts described in the paper by Molina Gabriel. The project implements the asset trader (agent) using recurrent reinforcement learning(RRL).
Key Elements : Utility Function, Policy and Gradient Ascent
One commonly used metric in financial engineering is Sharpe's ratio. For this project, sharpe ratio is our utility (or reward) function. The trader will attempt to maximize the sharpe's ratio which technically represents investment strategies with less volatile profit.
Sharpe ratio - This file calculates sharpe ratio for given investment returns.
Reward - This file calculates the sharpe ratios for all returns over window size M.
Here the trader function Ft, is basically a neuron activated with tanh with the output between -1 and 1. The value of Ft determines the current action.
Update Ft - This file updates the Ft for the whole time interval.
The optimization function needs to maximize the sharpe ratio. Hence, we require gradient ascent.