Author ORCID Identifier
0009-0007-4307-1254
Document Type
Dissertation
Date of Award
5-31-2024
Degree Name
Doctor of Philosophy in Business Data Science - (Ph.D.)
Department
Data Science
First Advisor
Dantong Yu
Second Advisor
Cheickna Sylla
Third Advisor
Zuofeng Shang
Fourth Advisor
Xinyuan Tao
Fifth Advisor
Zhipeng Yan
Abstract
Time-series analysis is essential for a wide range of financial applications, including but not limited to bond valuation, firm earnings forecasts, firm fundamentals predictions, and firm characteristics imputations. Given its considerable value, the financial community has shown a strong interest in refining and advancing time-series analysis techniques. The study in this dissertation contributes to this field by employing advanced machine learning approaches, specifically graph neural networks, deep neural networks, and matrix/tensor methods. The primary objectives are twofold: first, to reveal complex correlations within financial time series to improve prediction accuracy, and second, to enhance the process of integrating and imputing missing financial data.
One breakthrough of this research is the conversion of institutional bondholding tabular data into a network structure that serves as the foundation for the Temporal Bipartite Graph Neural Network (TBGNN) model. This model supports the interaction between bonds and their surrounding networks, extracts pricing information for corporate bonds from the bond network, and demonstrates a remarkable ability to account for approximately 90% of the variations in in-sample returns. It also significantly outperforms traditional linear and non-linear forecasting models, providing at least a four-fold improvement in the accuracy of out-of-sample forecasts.
Furthermore, this work addresses the critical challenge of missing values in financial time series that affect the reliability and precision of financial models. It develops stochastic gradient descent algorithms for nonlinear sparse tensor factorization and completion methods. This algorithm enables financial data integration and imputation and demonstrates a significant leap in performance, achieving improvements of 40%-74% over linear tensor completion models and 2%-52% over contemporary nonlinear models. This research enriches the methodologies for financial time series analysis, with implications for both the accuracy of predictions and the imputation of financial time series.
Recommended Citation
Zhou, Dan, "Financial time series fusion, completion, and prediction with deep neural networks" (2024). Dissertations. 1762.
https://digitalcommons.njit.edu/dissertations/1762
