Deep learning-based stock market prediction using historical price and topic-distributed news sentiment

Predicting stock prices remains a challenging task because of the volatile and non-linear nature of financial markets. This study proposes a hybrid deep learning method that combines sentiment and topic modeling of financial news with historical stock data to predict Apple Inc. stock prices. FinBERT is used to obtain sentiment scores and BERTopic is used to derive topic distributions and capture topic-aware sentiment context. These features are aligned on a daily basis and processed in a dual-branch neural architecture. One branch uses either LSTM, GRU, or Temporal CNN (TCNN) to handle historical price sequences, while the other is used to process topic-sentiment vectors. The results show that GRU performs slightly better than the other two models, with the highest R2 (0.9866), the lowest RMSE (2.81), and the lowest MAE (2.27). This study focuses only on Apple stock so it gives a more focused view on how financial news sentiment can affect the price prediction of a single stock. Limitations related to dataset size, computational resources, and real-time data availability are also discussed for future research.