Stock Market Prediction using a Deep Neural Network with A Large Set of Financial Indicators
Examensarbete för masterexamen
High-performance computer systems (MPHPC), MSc
Correctly predicting the trends or values of different stocks within the stock market has been of general interest for many years. This thesis investigates the possibilities for further increasing the Root Mean Squared Error accuracy of current stock market prediction methods using Long Short-Term Memory. Specifically, the derivative as a function of the stock price as well as the time of day during the trade is investigated closer. The hypothesis behind these two approaches is that the relative difference of the stock price is important as well as the time of day which is assumed to impact the stock market. The investigation of time is examined by predicting and training the model on specific times of the day. Furthermore, the main method which is attempted is training the model on a much larger amount of indicators than has been traditionally used. In addition to the stock data, 92 stock market indicators were introduced to the model. The many different methods also required a task distribution system to be implemented for faster execution. A scalable distribution system was then created, increasing execution speed by up to 7.25 times. Results show that using a lower window size is beneficial for a problem this large. Furthermore, that time is an important factor in stock price prediction. The highest directional accuracy obtained was 55.4 % for afternoon predictions on the Advanced Micro Devices stock using all of the indicators. The best Root Mean Squared Error value was 0.53 for this stock and obtained without any indicators and only using the closing price for training. These results are based on 8 months of data with 15 minute intervals.
Machine Learning , Stock Market Prediction , Long Short-Term Memory , Time Series