suha-memon/finbert-stock-sentiment

FinBERT Stock Sentiment Analyzer

Fine-tuned FinBERT model for financial sentiment analysis of stock-related news headlines.

Model Description

This model is a fine-tuned version of ProsusAI/finbert on financial sentiment data. It classifies financial text into three categories:

• Negative (0)
• Neutral (1)
• Positive (2)

Training Details

Training Approach

• Base Model: ProsusAI/finbert
• Fine-tuning Method: Focal Loss with class weighting
• Training Epochs: 10 (with early stopping)
• Learning Rate: 3e-5
• Batch Size: 16 (effective 32 with gradient accumulation)
• Warmup Ratio: 0.1

Performance Metrics

Metric	Score
Accuracy	81.88%
F1 Macro	0.8009
F1 Weighted	0.8259

Per-Class Performance

Class	F1 Score
Negative	0.6719
Neutral	0.8203
Positive	0.9104

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model and tokenizer
model_name = "suha-memon/finbert-stock-sentiment"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Example usage
headline = "Apple announces record-breaking quarterly earnings"
inputs = tokenizer(headline, return_tensors="pt", truncation=True, max_length=128)

with torch.no_grad():
    outputs = model(**inputs)
    predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)

# Get probabilities
negative_prob = predictions[0][0].item()
neutral_prob = predictions[0][1].item()
positive_prob = predictions[0][2].item()

print(f"Negative: {negative_prob:.2%}")
print(f"Neutral: {neutral_prob:.2%}")
print(f"Positive: {positive_prob:.2%}")

Intended Use

This model is designed for:

• Analyzing sentiment of financial news headlines
• Stock market sentiment analysis
• Financial document classification
• Real-time news sentiment tracking

Limitations

• The model performs best on financial news headlines (similar to training data)
• Negative sentiment detection (F1: 67%) is weaker due to class imbalance in training data
• May not generalize well to non-financial domains
• Limited to English language text

Training Data

The model was trained on financial sentiment data from Kaggle, consisting of:

• Training set: ~4,800 labeled examples
• Validation set: ~580 examples
• Test set: ~580 examples

Class distribution was imbalanced with fewer negative examples, addressed using Focal Loss.

Citation

If you use this model, please cite:

@misc{finbert-stock-sentiment,
  author = {Your Name},
  title = {FinBERT Stock Sentiment Analyzer},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/suha-memon/finbert-stock-sentiment}}
}

Team

• Suha Memon
• Nick Cirillo
• Kalen Truong
• Bruce Zhang

Acknowledgments

• Base model: ProsusAI/finbert
• Framework: Hugging Face Transformers