Neural networks in stock price jump prediction
Mäkinen, Milla (2017)
Mäkinen, Milla
2017
Tuotantotalous
Talouden ja rakentamisen tiedekunta - Faculty of Business and Built Environment
This publication is copyrighted. Only for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2017-12-07
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tty-201711222199
https://urn.fi/URN:NBN:fi:tty-201711222199
Tiivistelmä
This thesis proposes a new convolutional long short-term memory network with a feature-dimension attention model for predicting the occurence of stock price jumps by studying several popular neural network types for time series prediction and examining stock price jumps with data from NASDAQ limit order books for five different stocks. The proposed convolutional long short-term memory attention model network (CNN-LSTM-Attention) is further compared to a convolutional and a long-short term memory network from existing stock price prediction literature as well as a multi-layer perceptron. Normalized limit order book data with additional features is used to predict whether a jump will occur within the following minute.
Testing the models yields very promising results for the predictability of jumps, which is especially significant as there is very little existing research on predicting stock price jumps with machine learning methods. Additionally, the proposed CNN-LSTM-Attention method is found the best from the tested ones, with the average F1 of 0.72. Furthermore, predicting existing jumps is found significantly easier than their size or direction, supporting the existence of a jump counting process which is separate from both the regular stock price process and the jump sizes, and which is not fully unpredictable.
Testing the models yields very promising results for the predictability of jumps, which is especially significant as there is very little existing research on predicting stock price jumps with machine learning methods. Additionally, the proposed CNN-LSTM-Attention method is found the best from the tested ones, with the average F1 of 0.72. Furthermore, predicting existing jumps is found significantly easier than their size or direction, supporting the existence of a jump counting process which is separate from both the regular stock price process and the jump sizes, and which is not fully unpredictable.