Unsupervised Learning in Finance: From Bibliometric Analysis to Investment Applications

Abstract

This thesis explores the applications of unsupervised learning methods in finance through three interconnected studies.
Chapter One provides a comprehensive bibliometric analysis of unsupervised clustering applications in finance, examining 392 research documents from 2000 to 2024.
The study reveals a significant surge in research post-2018 and identifies three primary areas of application: Market Modelling and Portfolio Analysis, Risk Analysis, and Banking and Credit Assessment.
This systematic review establishes the theoretical foundation and identifies research gaps in the field.
Chapter Two develops a novel pairs trading strategy that integrates firm characteristics with price information through unsupervised learning techniques.
Applied to the US stock market from 1980 to 2020, the strategy demonstrates remarkable performance, achieving a statistically significant annualised return of 24.8 and a Sharpe ratio of 2.69.
The robustness of the strategy is confirmed through various tests, including transaction costs and size filters.
Chapter Three investigates the effectiveness of firm characteristics and unsupervised clustering in constructing partial index tracking portfolios.
Using data from 2003 to 2022, the study demonstrates that firm characteristics alone can achieve tracking performance comparable to traditional return-based methods.
The proposed Proportional () stock selection method proves particularly effective in managing uneven cluster sizes across various portfolio sizes.
In conclusion, this thesis advances our understanding of unsupervised learning applications in finance, demonstrating their effectiveness in portfolio management through both theoretical analysis and empirical implementation.
The findings suggest that incorporating firm characteristics through clustering techniques can significantly enhance investment strategies, offering new perspectives on portfolio construction and management.