Type
Text
Type
Dissertation
Advisor
Zhu, Wei | Wu, Song | Kuan, Pei Fen | Xiao, Keli.
Date
2015-12-01
Keywords
Artificial Neural Networks, Genome-wide Association Study, High Frequency Trading | Statistics
Department
Department of Applied Mathematics and Statistics.
Language
en_US
Source
This work is sponsored by the Stony Brook University Graduate School in compliance with the requirements for completion of degree.
Identifier
http://hdl.handle.net/11401/77543
Publisher
The Graduate School, Stony Brook University: Stony Brook, NY.
Format
application/pdf
Abstract
Artificial neural network model is a powerful method that has been widely applied in many different areas. It is essentially a nonlinear statistical model, empirically proved with good prediction accuracy, and has been applied in both regression and classification problems. One challenge in applying artificial neural network models is constructing proper structure adaptive to specific problems. This thesis work is to introduce a novel, double-layered feed-forward neural network (DNN) model with special link patterns. Its applications to genome-wide association studies and stock price prediction in high frequency time scale have been explored. Detecting gene-gene interactions in traditional Genome-wide associate studies (GWAS) is mostly at the SNP level, called SNP-SNP interactions, which ignores the existence of large amount of correlations embedded among nearby SNPs. Popular existing methods with this mechanism, such as multifactor-dimensionality reduction (MDR) and random forests, would usually suffer from redundant interaction tests, due to the correlations between SNPs, and subsequently from less powers. With our new DNN model, we can take advantage of the correlations between SNPs and perform interaction test at the level of SNP blocks. Extensive simulation studies have been conducted to compare our new method with Random Forests. And our simulation results suggest that the DNN model can have higher power than Random Forests in detecting the existence of causal SNPs – no matter the effect is interactive or marginal. We also have applied the DNN model to financial markets, forecasting changes of stock prices in high frequency. One advantage of our DNN model is that it utilizes correlation information between different stocks, a pattern more commonly observed in high-frequency data but ignored in most existing methods. Our method has been tested on the 100 stocks with largest capital in S&P 500 using 5-minute data, and its performance has been benchmarked with a single layer neural network model and the classical ARMA-GARCH model. The DNN model clearly outperforms to the other models in terms of prediction accuracy and Sharpe ratio. Given the parallelizable scheme of our method with DNN models, it may be capable for designing profitable trading strategies in high frequency time scale. | Artificial neural network model is a powerful method that has been widely applied in many different areas. It is essentially a nonlinear statistical model, empirically proved with good prediction accuracy, and has been applied in both regression and classification problems. One challenge in applying artificial neural network models is constructing proper structure adaptive to specific problems. This thesis work is to introduce a novel, double-layered feed-forward neural network (DNN) model with special link patterns. Its applications to genome-wide association studies and stock price prediction in high frequency time scale have been explored. Detecting gene-gene interactions in traditional Genome-wide associate studies (GWAS) is mostly at the SNP level, called SNP-SNP interactions, which ignores the existence of large amount of correlations embedded among nearby SNPs. Popular existing methods with this mechanism, such as multifactor-dimensionality reduction (MDR) and random forests, would usually suffer from redundant interaction tests, due to the correlations between SNPs, and subsequently from less powers. With our new DNN model, we can take advantage of the correlations between SNPs and perform interaction test at the level of SNP blocks. Extensive simulation studies have been conducted to compare our new method with Random Forests. And our simulation results suggest that the DNN model can have higher power than Random Forests in detecting the existence of causal SNPs – no matter the effect is interactive or marginal. We also have applied the DNN model to financial markets, forecasting changes of stock prices in high frequency. One advantage of our DNN model is that it utilizes correlation information between different stocks, a pattern more commonly observed in high-frequency data but ignored in most existing methods. Our method has been tested on the 100 stocks with largest capital in S&P 500 using 5-minute data, and its performance has been benchmarked with a single layer neural network model and the classical ARMA-GARCH model. The DNN model clearly outperforms to the other models in terms of prediction accuracy and Sharpe ratio. Given the parallelizable scheme of our method with DNN models, it may be capable for designing profitable trading strategies in high frequency time scale. | 90 pages
Recommended Citation
Chen, Hao, "DEVELOPMENT OF A NOVEL DOUBLE NEURAL NETWORK AND ITS APPLICATIONS" (2015). Stony Brook Theses and Dissertations Collection, 2006-2020 (closed to submissions). 3343.
https://commons.library.stonybrook.edu/stony-brook-theses-and-dissertations-collection/3343