Stock Market Trend Prediction for Next Day Direction using Sentiment Analysis with FinBERT

This study investigates the predictive performance of various machine learning models and neural network architectures using a comprehensive dataset comprised of price data from the E-mini Nasdaq100 Futures and sentiment data derived from news articles. The price dataset, spanning from April 2002 to September 2024, is meticulously curated to exclude non-trading days, while sentiment scores are generated through advanced techniques, including VADER and FinBERT. Additionally, the research incorporates intermarket influences by integrating data from E-mini S&P500, E-mini Dow Jones, VIX, and Dollar Index. Exploratory data analysis reveals significant seasonal tendencies within the market data, which enhance model training when utilized effectively. Results show that the Conv1D+LSTM model combined with intermarket data and FinBERT sentiment analysis achieves an accuracy score of 0.552 and an F1 score of 0.712. Additionally, the Gradient Boosting model attains an accuracy of 0.588 and 0.555 when applied SMOTE to balance the dataset, illustrating its effectiveness in predicting both bullish and bearish market directions. This research underscores the importance of incorporating sentiment features alongside traditional price data in financial forecasting.