Instruction Finetuning Foundation Models, Three-Stage Bubble Analysis, and Examining the Size Effect
Loading...
Authors
Atsiwo, Abraham
Issue Date
2024
Type
Dissertation
Language
en_US
Keywords
Bubble Prediction , Fine-tuning , Large Language Model , Long-Term Stability of the Capital Asset Pricing Model , Sentiment Analysis , The Size Effect
Alternative Title
Abstract
Macroeconomic indicators and financial news are important to financial professionals and market participants due to their importance in sentiment analysis, speculative bubble analysis, and portfolio selection. According to the Efficient Market Hypothesis, asset prices incorporate all available information. Within this context, information pertains to company statements, business reports, and headlines in financial news. Hence, classifying information as negative, neutral, or positive is paramount to investment decision making, portfolio selection, and detecting early warning signals for financial collapse. However, sentiment analysis in finance is quite challenging due to the "closed source" of labeled data. To overcome this, we leverage existing data sources with data generated by a large language model (through fine-tuning and prompting) and use the datasets to fine-tune another large language model for text classification in finance. We examine the impact of macroeconomic indicators and financial news in predicting speculative bubbles in a multilabel classification using ensemble methods, leveraging the fine-tuned model for text classification. We also explore how economic indicators such as dividend yield and interest rates guide portfolio selection and explain the size effect.