finbert-finance-sentiment
A 3-class financial sentiment model that labels English financial text as positive, negative, or neutral. It is a fine-tuned version of ProsusAI/finbert, further trained on the Financial PhraseBank dataset.
Model description
The model takes a sentence of financial text — such as a line from a news headline, earnings report, or market commentary — and predicts its sentiment. It builds on FinBERT, a BERT model pretrained on a large financial corpus, and sharpens it on expert-annotated financial sentences.
- Base model: ProsusAI/finbert
- Task: 3-class sentiment classification (text-classification)
- Language: English (financial domain)
- Labels:
negative(0),neutral(1),positive(2)
Intended uses & limitations
Intended uses: Sentiment scoring of financial text — news headlines, earnings commentary, analyst notes, and similar short financial statements.
Limitations:
- English and financial domain only; performance on general or non-financial text will be weaker.
- Trained on relatively formal financial news sentences, so very informal text (e.g. social media slang, tickers, memes) may be less reliable.
- Sentiment here means market/financial tone, not personal opinion — "shares fell" is negative regardless of phrasing.
How to use
from transformers import pipeline
clf = pipeline("text-classification", model="rajaadil/finbert-finance-sentiment")
print(clf("The company reported record profits and raised its dividend."))
# [{'label': 'positive', 'score': 0.97}]
Training data
Trained on the sentences_allagree subset of the Financial PhraseBank dataset — the highest-quality subset, containing only sentences where all human annotators agreed on the label.
- Split: 80% train / 20% test
- Classes: negative, neutral, positive (neutral is the majority class)
Training procedure
| Hyperparameter | Value |
|---|---|
| Base model | ProsusAI/finbert |
| Epochs | 4 (best checkpoint kept) |
| Train batch size | 16 |
| Learning rate | 2e-5 |
| Max sequence length | 128 |
| Best model selected on | macro-F1 |
Evaluation results
Best checkpoint (epoch 3), evaluated on the held-out test set:
| Metric | Value |
|---|---|
| Accuracy | 0.820 |
| F1 (macro) | 0.811 |
Macro-F1 is reported alongside accuracy because the dataset is class-imbalanced (neutral is the majority class); macro-F1 reflects performance across all three classes rather than favouring the majority.
Framework versions
- Transformers
- PyTorch
- Datasets
Author
Built by rajaadil.