Empowered PG in Forex Trading

Abstract

With the assumption that capital markets follow the semi-strong form of efficient market hypothesis (EMH), numerous efforts have been taken to defeat the non-stationary financial market, ranging from time series analysis, artificial intelligence for prices prediction, to automated decision making by reinforcement learning. This experiment integrated the power of time series forecasting of neural network with the competence of actions selecting of the reinforcement learning. CNN was trained first to predict future prices, and then it fed the output to the policy gradient (PG) model together with historical data to empower the trading decisions. The experiment was conducted on 30 minutes interval of EUR/USD pair in Forex between 2014 and 2018. Our experimental results showed that our model can achieve higher return in both train and validate samples than buy and hold strategy.