Prediction of stock price fluctuations using convolution neutral networks

Specialeforsvar ved Rasmus Aagaard Jensen

Titel: Prediction of stock price fluctuations using convolution neural networks

Abstract: This paper tries as many other papers beforehand to forecast the stock price movements through the field of machine learning applications. In contrary to most papers this paper take use of Deep Learning algorithms characterized by their automated feature selection in order to form predictions of the stock price movements one day ahead on every stock on the Danish OMXC25. More specifically the paper compares the use of 1-D, 2-D and 3-D convolutional neural networks. Special for this paper is the use of generalization. Each convolutional neural network is only tuned for a small sample of the initial stocks and then generalized to form predictions on all the stocks. By use of some rather simple data of the individual stocks, the paper illustrates that convolutional neural network constitute a powerful tool in forming the stock price movement predictions even in the generalized cases.
The prediction performance is evaluated by use of several statical measures. To further compare the predictive power of the different networks simple trading strategies utilizing the predictions are constructed. Despite the general low gain of these trading strategies, the paper demonstrates that by increasing the forecast time window, you can significantly increase the profit of the different strategies.

Vejleder: Rolf Poulsen
Censor:   David Sloth Pedersen, Danske Bank