Fintech related machine learning : experimental investigation on the effects of data variance towards difficulty of learning for deep learning models

• Main Author: Chua, Shearman Wei Jie
• Other Authors: Althea Liang
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/148075
• Institution: Nanyang Technological University

The application of Artificial Intelligence models in the domain of Financial technology, more commonly known as Fintech, has been widely studied by researchers throughout the years. With Deep Neural Networks gaining traction in the field of Artificial Intelligence, the use of Deep Neural Networks, or Deep Learning models, for Fintech applications has become more prevalent in recent years. One of the most widely studied application of Deep Learning models for Fintech, is the utilization of Deep Learning models for tasks surrounding the trading and investing of financial instruments such as Stocks or Forex. Some of these tasks include the prediction of next-day stock prices, stock portfolio management, valuation of assets, etc. Although past research have tackled stock market prediction tasks using Deep Learning models, much of these research conducted, allow the Deep Learning models to learn by training on a large amount of training data, without full understanding of the underlying data that is used to train these models from a financial perspective. Past research also often do not take into consideration if the data used to train the Deep Learning models fully captured the various variations and economic trends present in stock data, which is necessary for the Deep Learning models to learn the stock market prediction task and enable them to generalize better and make more accurate predictions. In addition, minimal work has been done to study how variations present in the training data used to train the Deep Learning models, affect the ability of the Deep Learning models to learn the stock market prediction task, as well as if the introduction of specific variations to the training data(based on knowledge of stock trends and factors affecting stock prices) used to train the Deep Learning models, will allow them to be able to better learn and make predictions on stocks with certain characteristics(bullish stocks, small-cap stocks etc.). Therefore, there is a need to study how different variations introduced to the training data of the Deep Learning models affects the learning ability of the Deep Learning models for learning the task of stock market prediction, or in other words, how different variations introduced to the training data of the Deep Learning models affects the difficulty of learning of these models to learn the task of stock market prediction. This paper proposes a novel study on the effects of introducing different variations to the training data used to train Deep Learning models(CNN and LSTM models), towards the ability of these models to learn the task of stock trading action prediction. The variations were introduced to the training data in a manner that allowed us to better understand how various variations present in the training data used for stock trading action prediction affect the ability of Deep Learning models to learn the task of stock trading action prediction. The variations introduced to the training data also took into consideration stock market specific factors likely to affect stock prices and movement. From the results obtained from the study, we were able to better understand the effects of the introduction of certain types of data variation towards the ability of Deep Learning models to learn the task of stock trading action prediction. In addition, we also learnt which variations, when introduced to the training data used to train the Deep Leaning models, are beneficial towards helping Deep Learning models better overcome the difficulty of learning from the training data during training, to enable them to solve the task of stock trading action prediction. The results of the study will be of assistance towards future studies conducted on the use of Deep Learning models for stock market prediction tasks, by providing insights on how to improve the training of Deep Learning models for stock market prediction tasks with the use of training data introduced with variations beneficial towards helping the Deep Learning models better overcome the difficulty of learning from the training data during training.