Transformers and tradition: using Generative AI and Deep Learning for financial markets prediction

Wade, Toby J. (2024) Transformers and tradition: using Generative AI and Deep Learning for financial markets prediction. PhD thesis, London School of Economics and Political Science.

Text - Submitted Version Download (3MB)

Abstract

Artificial intelligence has revolutionized numerous industries, and financial markets are no exception. With the ability to process vast amounts of data quickly and accurately, AI algorithms have been increasingly used in finance to predict stock prices, detect fraud, and optimize investment strategies. However, the full potential of AI in finance still needs to be explored, and researchers continue to explore new ways to apply machine learning techniques to financial challenges. This thesis investigates whether advanced Generative AI and Deep Learning techniques are more effective in extracting information for predicting financial markets than conventional natural language processing methods. The first part of this thesis analyzes quarterly SEC 10-Q filings for S&P 500 companies from January 2000 to December 2019 to show how artificial intelligence techniques can provide reasoning about changes in corporate disclosures indicative of future company performance. This thesis finds that by leveraging the reasoning capabilities of the Claude2 large language model on the Management Discussion & Analysis section of a 10-Q, negative excess returns of -5.5% over 180 days (- 11% annualized) can be avoided. The paper introduces two novel approaches: A) Concatenating Deep Learning architectures comparing quarterly filings, and B) Summarization methods using Claude2 to extract sentiment signals related to significant business risks, profitability, legal, and market pressures. Together, these techniques demonstrate new ways of expanding beyond rudimentary natural language processing approaches that many investment firms have historically used, such as lexicons and cosine similarity, to answer fundamental questions related to firm performance. The second part of the thesis takes a step further, developing an enhanced sentiment model and utilizing Bitcoin subreddit data from December 2010 to January 2022 to predict the price of Bitcoin 60 days in advance. The Reddit text data is known for its high noise level, with non-relevant price information such as advertisements or technical advice. This noise can significantly impact the accuracy of the predictions. To address this, the research proposes a novel approach that combines a Few-Shot RoBERTa topic classification model with sample augmentation on training data powered by ChatGPT. This approach effectively reduces the noise, creating a more robust sentiment signal. The enhanced sentiment signal is then integrated with other Bitcoin on-chain features in a nonlinear multivariate LightGBM model. The results clearly demonstrate the impact of noise reduction, with the F1 score for predicting the sign of Bitcoin 60 days in advance increasing from 0.26 to 0.63 on the test set.

Item Type:	Thesis (PhD)
Additional Information:	© 2024 Toby J. Wade
Library of Congress subject classification:	H Social Sciences > HG Finance T Technology > T Technology (General)
Sets:	Departments > Statistics
Supervisor:	Lam, Clifford and Kalogeropoulos, Konstantinos
URI:	http://etheses.lse.ac.uk/id/eprint/4657

Actions (login required)

Record administration - authorised staff only