NeuroTraders

Inspiration

Stock price prediction based on News and financial historical database using recurrent neural networks

What it does

Use of state of the art of techniques to predict a stock's price. Moreover we introduce a new technique of twin-predictors to improve the choice of trades.

How I built it

The prediction process can be compared to a pipeline with 3 main stages: 1) Data collection: Not only we use stock price and volume data, but most importantly we generate a daily sentiment score based on news available online (more than 20000 news have been scrapped for predicting Apple's stock price). 2) The data is used to train two deep Recurrent Neural Network; one generates a price prediction while the other gives an error estimate for the output of the first. 3) The predictions are used to simulate a year of trading on a sophisticated trading strategy. The results outperform the stock's return.

Challenges I ran into

Finding sentiment data from unstructured sources such as web pages, news articles, etc. was a considerable challenge. Most of the times we had to filter the information or keep only the header to run the simulation in a considerable time span. Knowing the price and the error range is already an spectacular achievement but the goal of our team is to use in our benefit. Designing a Trading Strategy that exploits the prediction was quite tough; the result is an algorithm that generates additional return given the predictions are sufficiently accurate (about 7% accuracy has proven enough)

Accomplishments that I'm proud of

The incredible prediction of the stock price with the neural network system that achieves, in its latest iteration, a RMSE of 1.311, approximately 1% of the current price per share of Apple. Furthermore the most outstanding fact is that the Trading strategy based on the output of the Neural Network equals and sometimes outperforms the rise of the stock by making use of the predicted information, indicating that indeed alpha can be generated via the approach of training neural networks to predict stock prices and estimate prediction errors.

What's next for NeuroTraders

Improve the Neural Network by trying different structures; at the moment we are using a recurrent network. We are confident on the fact that the training of the network can be developed more and more in order to achieve higher accuracy by introducing more layers, which requires larger sets of input data. In further projects, we want to create a complete dictionary for the pointing system of the sentiment analysis. For instance taking into account the order of the words would give higher accuracy and a better understanding of the content of the news.