Deep and reinforcement learning with sentiment for stock market trading

In this talk we will review our recent work on combining news and social media sentiment signals with cutting edge Machine Learning techniques to forecast and trade stock markets. Our first application is developing a sentiment-based intraday trading strategy for a large stock index (EUROSTOXX 50). We utilize an ingenious combination of Kalman Filter for de-noising the news sentiment signal, Multi-Layer Perceptron for features selection and an LSTM network for actual forecasting, to obtain powerful forecasting framework. Out-of-sample forecasts show that downward movements in price can be predicted correctly in more than 75% of cases. Algorithmic trading strategy based of our forecasts shows significant economic gain, even after transaction costs.

Outperform the benchmark

Our second application is trading individual stocks that comprise S&P500 index. We use high-frequency stock-specific news sentiment signals, in combination with a reinforcement learning approach. We apply Neuro Evolution of Augmenting Topologies (NEAT), to learn the non-linear functional price response to different input features. The resulting trading strategy not only outperforms the benchmark, but also generates double the average profits made by the benchmark.

Our results demonstrate the great potential of combining external real-time information (such as sentiment) with Machine Learning techniques for algorithmic trading.

Want to know more?

If you want to know more about these subjects, you can contact Svetlana or Pim directly.