Comparative Study of Deep Learning Architectures for Financial Sentiment Analysis in Digital Market Environments

This study investigates the effectiveness of deep learning architectures for financial sentiment classification using textual data from financial reports and market statements. Four models, namely Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), Gated Recurrent Unit (GRU), and Convolutional Neural Network (CNN), were trained and evaluated on a dataset of 5,842 labeled financial sentences categorized as positive, negative, or neutral. The evaluation employed accuracy and weighted F1-score metrics to assess predictive performance and class balance. The results revealed that the CNN model achieved the highest performance with an accuracy of 70.32 percent and an F1-score of 70.22 percent, followed by the BiLSTM model with 67.49 percent accuracy and 66.86 percent F1-score. The CNN’s superior performance demonstrates its ability to capture local sentiment-bearing features efficiently, while BiLSTM excels in understanding bidirectional contextual dependencies. In contrast, the LSTM and GRU models performed less effectively due to their reliance on sequential memory mechanisms, which are less applicable to short and factual financial sentences. These findings highlight that CNN provides the best balance between accuracy, efficiency, and adaptability for financial sentiment classification and can serve as a foundation for intelligent digital market analysis systems and AI-driven financial decision support tools.