Automating the Creation of Trading Strategies Using Deep Reinforcement Learning: Algorithmic Trading

Abstract

As before, the RL state space's dimensionality is probably too low for the agent to learn a trading algorithm that is highly adaptable and appropriate for a variety of volatility regimes. The VG training data suggests that the best course of action in this case is to apply "more of the same". The main problem will now be persuading the RL agent to make use of the additional information that the 25 multi-asset futures contracts obviously provide. Third, more investigation is required to comprehend why evaluation frequency, with semi-annually being the terrible exception, is so important in preventing catastrophes like the one in 2008. In numerous respects, this appears strange. Fourth, the one strategy that has worked. Since investors are often risk adverse, the objective is to maximise a risk-adjusted performance function, such as the Sharpe ratio. This really results in a concave utility function. After learning the distribution, we may select actions that have the highest predicted Q-value and the lowest standard deviation from it to maximise the Sharpe ratio. The mean-variance portfolio theory and reward functions may be used to get an excellent anticipated algorithmic trade return with minimal volatility.