Algorithmic trading strategies are typically built using human intuition in combination with back-testing, and get continuously tuned over time. They depend on the market price curves of securities, derivatives, currencies, or commodities along with one or more indicators such as moving average or RSI. Maximising profit is usually a function of trial and error.
This white paper explores how machine learning, in particular deep learning, can be employed to improve algorithmic trading strategies. Using deep belief networks (DBN), we attempt to predict the case when the price will have a significant change in the near future and build the trading strategy based on this. Backtesting shows that significant profit improvements can be achieved using this technique. The paper also describes how the DBN can be trained with maximum performance using a combination of CPUs and GPUs.
Structure of a deep belief network