Decoding NASDAQ Trends with Sentiment and Transformers

In this research, we explore how integrating sentiment analysis with deep learning models enhances stock market forecasting, focusing on the NASDAQ-100 index. FinBERT-based sentiment scores from financial news were combined with technical indicators to train two models: Long Short-Term Memory (LSTM) and Informer, a transformer-based architecture for longrange time-series prediction. Both aimed to forecast daily closing prices. Results show that incorporating sentiment consistently improves predictive accuracy. Informer performed better in estimating price magnitudes, especially during volatile or downward markets, while LSTM had a slight edge in directional accuracy. This highlights a trade-off between accurately predicting trend direction versus price value. Limitations include the use of a fiveyear dataset, reliance on news-based sentiment, and exclusion of intraday data. Despite these, the findings demonstrate the value of sentiment-aware models and complementary model strengths. Future research may expand datasets, include social media sentiment, and refine domain-specific models to improve real-time performance.