The paper introduces FinEAS (Financial Embedding Analysis of Sentiment), a model for financial sentiment analysis that leverages supervised fine-tuned sentence embeddings from Sentence-BERT. The authors argue that financial sentiment is a sentence-level task, and vanilla BERT's sentence embeddings are suboptimal. FinEAS uses a Sentence-BERT backbone (fine-tuned on NLI) to produce high-quality sentence embeddings, which are then fed into a linear layer with tanh activation for regression to continuous sentiment scores between -1 and 1.

Experiments are conducted on a large-scale RavenPack financial news dataset with three temporal splits (6, 12, and 24 months) and an out-of-sample two-week test set. FinEAS is compared against vanilla BERT, BiLSTM, and FinBERT (a domain-specific BERT). Results show that FinEAS significantly outperforms all baselines in terms of MSE. Notably, FinEAS with full fine-tuning achieves MSE of 0.0044 on 6-month data versus FinBERT's 0.0050, demonstrating that general-domain sentence embeddings can surpass domain-specific models for this task. The findings suggest that for financial sentiment analysis, the quality of sentence-level embeddings is more critical than domain-specific vocabulary or structure.

The paper concludes that FinEAS is a simple, effective, and robust approach for financial sentiment analysis, and it makes the code and model weights publicly available. Future work may explore other Transformer models fine-tuned for sentence or document-level tasks in finance.