Aristotle University Of Thessaloniki (AUTH)
Deep learning techniques for financial data
Abstract
dc:descriptionFinancial transactions and products have been evolving and becoming more democratized over the past century. Peer-to-peer transactions and contracts are available to an increasing number of people, and interest in wealth management is growing. The introduction and evolution of electronic trading and automated financial services have spurred the development of automated trading strategies for investing and managing assets. At the same time, Machine Learning (ML) is seeing a meteoric rise in interest due to the unprecedented achievements of Deep Learning (DL). The computational capacity of processors has been growing exponentially, and in the last decade, it has reached a critical point for feasible large-scale DL. That, along with an abundance of data generated on the world wide web, has led to breakthroughs in Artificial Intelligence research. These range from unparalleled accuracy in classification tasks, to surpassing human-level performance in intricate games such as chess. These improvements are attributed to DL model growth in depth and number of parameters. As we probe more research fields under the lens of DL, one that is picking up speed is finance and financial trading. Applying DL as a black box to financial trading is not simple and does not yield the same success as other DL applications. Throughout this dissertation, we explore the applicability of DL to financial trading. Novel models and methods leverage the inherent power of DL to benefit existing and new approaches to financial trading. We develop market simulation environments that use historical price data and train agents with Reinforcement Learning to optimize their trading performance. The reward signal in these approaches is the achieved trading profit. Novel Deep Reinforcement Learning (DRL) methods are proposed for Reward Shaping and Knowledge Distillation. In the proposed Reward Shaping method, we implement an auxiliary reward, stemming from the empirical measure of price trailing proximity controlled by the agents. For Knowledge distillation, we examine the effects of a diversified distillation scheme by training teachers using distinct subsets of a dataset. We show that both the Reward Shaping and the Knowledge Distillation methods significantly improve the performance of the resulting models. Next, we study the efficacy of submitting Limit Orders as a more effective order type for achieving profitable trading outcomes. We develop a DRL environment to support Limit Orders and train agents that outperform the baselines by using Limit Orders. We explore and compare Supervised Learning DL models for predicting price changes using high-frequency Limit Order Book data. Finally, since many market actors use qualitative or algorithmic strategies, a signal cloning model is developed that manages to accurately reproduce the decisions of the original algorithm. In this dissertation, the efficacy of DL approaches for Financial Data is thoroughly explored. The experimental evaluation of our proposed methods implies that DL can be used to successfully analyze, predict and model financial trading with clear avenues for improvements in future work.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2021
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Tsantekidis, Avraam
- Τσαντεκίδης, Αβραάμ
Subjects
dc:subject × 11Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/51390