Financial Sentiment Classifier (DistilBERT fine-tuned)
Fine-tuned distilbert-base-uncased for 3-class sentiment classification
(negative / neutral / positive) on the Financial PhraseBank dataset —
short financial news sentences labeled by finance-background annotators.
Intended use
Classifies the sentiment of short financial/business news sentences, e.g. company earnings reports, market commentary, or corporate announcements. Not intended for general-purpose sentiment analysis outside a finance context.
Training details
- Base model: distilbert-base-uncased
- Dataset: Financial PhraseBank ({~4,850} sentences, train/val/test split)
- Epochs: 3
- Learning rate: 2e-5
- Hardware: free Colab T4 GPU
Evaluation results (held-out test set)
| Metric | Score |
|---|---|
| Accuracy | 0.838 |
| Weighted F1 | 0.84 |
| Class | Precision | Recall | F1 |
|---|---|---|---|
| Negative | 0.79 | 0.81 | 0.80 |
| Neutral | 0.87 | 0.88 | 0.88 |
| Positive | 0.79 | 0.76 | 0.78 |
Known limitations
The model performs well on sentences with explicit sentiment language, but struggles on sentences requiring implicit domain reasoning — e.g. "the company will deliver X units to a buyer" implies a sale (positive) without using any positive-sounding words. It also has difficulty with sentences containing mixed or competing signals (e.g. two market indices moving in opposite directions within one sentence). The most common error pattern is misclassifying positive or negative sentences as neutral, likely influenced by class imbalance in the training data (neutral is the majority class).
Usage
from transformers import pipeline
classifier = pipeline("text-classification", model="navu013/finbert-sentiment-distilbert")
result = classifier("The company reported record profits this quarter.")
print(result)
Training data note
This is a learning/portfolio project — trained on a public academic dataset, not intended for production financial decision-making.