LLM-Based Financial Sentiment Analysis — NIFTY 50

Overview

Applied FinBERT and VADER sentiment models to approximately 29,000 financial news headlines spanning 2015–2025 for all NIFTY 50 constituents. Used panel regressions with two-way fixed effects to test whether news sentiment predicts stock returns and volatility.

Key Findings

• Sentiment magnitude significantly predicts short-run volatility (strongest result)
• Negative sentiment shows asymmetric impact — consistent with loss aversion theory
• Counterintuitively, VADER outperformed FinBERT for return prediction tasks

Methodology

• Data: NIFTY 50 daily price data via yfinance, ~29,000 headlines from NewsAPI, Google Trends data
• Sentiment: FinBERT (transformer-based) and VADER (lexicon-based) scoring
• Analysis: Panel regressions with two-way fixed effects, multicollinearity handling, look-ahead bias prevention, robustness checks
• Tools: Python, pandas, statsmodels, transformers (HuggingFace)