Echoes of Exuberance? Detecting Bubble Patterns in AI through the Lens of Dot-Com
Master thesis
Permanent lenke
https://hdl.handle.net/11250/3184939Utgivelsesdato
2024Metadata
Vis full innførselSamlinger
- Master Thesis [4657]
Sammendrag
This thesis explores the predictive capabilities of machine learning models in identifying speculative financial bubbles, with the focus on emerging technologies such as the internet and artificial intelligence. The thesis seeks to draw parallels between the dotcom bubble following the exuberance of the late 1990s internet boom and the current AI-driven market excitement. The goal is to investigate whether machine learning models trained on overpriced stocks during the dotcom bubble can detect bubble patterns in AI companies of the present.
The study employs diverse machine learning algorithms including LASSO logistic regression, random forest, gradient boosting, and support vector machines. The models are trained on firm-level data retrieved from the databases CRSP, Compustat and IBES. Key financial indicators and analyst sentiment data are incorporated to ensure the capture of thecomplex dynamics of stock valuation during speculative phases. Furthermore, the bubble indicators are classified through the Generalized Supremum Augmented Dickey-Fuller (GSADF) test, an econometric tool for identifying periods of explosive price growth.
The findings highlight the complexities of bubble-prediction, which proves to be a challenging endeavor. The findings highlight the ongoing expansion for AI firms, but the lack of true labels creates a challenge in terms of making any conclusion. Our findings show signs of speculative tendencies, and whether these reflect rational optimism in a growth-oriented market or are based on over-optimism will only be proved in time. Our findings should therefore be considered a sign of caution, rather than a conclusion.
By applying machine learning in bubble prediction on a company level, this thesis contributes a new angle to the growing field of financial- and bubble prediction. Also, by applying the GSADF test on individual stocks rather than entire indexes, as previously done in the literature, we explore the tests applicability in terms of detecting periods of explosive growth on a smaller scale. These findings might provide practical insights for further research into this complicated field of bubble detection.