Comparative Sentiment Analysis of Financial Text: Traditional Models vs. LLM-Based Evaluation with Gemini

This notebook demonstrates a comparative approach to sentiment analysis of financial text, focusing on stock-related news. It applies three established methods—VADER, FinBERT, and RoBERTa—to evaluate how traditional rule-based and transformer-based models interpret the same input differently. VADER provides quick, lexicon-driven polarity scores; FinBERT, trained on financial corpora, captures domain-specific sentiment; and RoBERTa, fine-tuned on Twitter data, generalizes sentiment detection in social contexts. Beyond these models, the notebook introduces a large language model (LLM)–based evaluation using Gemini. Instead of producing a single polarity, Gemini is prompted to assess text across multiple rubrics, including sentiment strength, event severity, confirmation level, novelty, and reasoning quality. This rubric-based framework highlights the potential of LLMs not only to classify sentiment but also to explain and calibrate it in more interpretable ways. Together, these methods showcase the evolution of sentiment analysis in finance—from lexicon rules to specialized transformers, and now toward transparent, rubric-driven assessments with LLMs.