Algorithmic short-term trading using time series prediction and reinforcement learning

Lewys Davies


Supervised by Yuhua Li; Moderated by Alun D Preece

Stock market trading has historically been about fundamental analysis, but with the introduction of electronic trading (NASDAQ), short-term trading using technical analysis gained popularity, and is used by many to make a profit. Given the (mostly) purely technical aspect of this trading, it’s no surprise that machine learning is currently the direction financial institutions and traders are turning to.

The idea is to combine time-series prediction using LSTMs with a deep reinforcement learning algorithm for financial signalling, which can be used for short-term algorithmic trading. The main issue with time-series prediction is that stock prices tend to be non-stationary, and prone to regime changes, which essentially renders older data irrelevant. However, by taking the first order derivative of the price, this effectively turns the data stationary, with the added benefit of making it easy to spot where profit can be made, since this can only occur when the gradient is above zero. So the reinforcement learning algorithm will include the time-series prediction on the first order derivative of the stock price as input, amongst other inputs such as volume and technical indicators, which are used by technical analysis traders to help predict the direction of a stock’s price.

Final Report (18/09/2020) [Zip Archive]

Publication Form