finnews-topic-single-classify

Financial News Topic Classifier

This model is a fine-tuned BERT-based classifier for financial news topic classification based on fuchenru/Trading-Hero-LLM, supporting 20 distinct financial topics. It is designed for use in financial NLP applications, news analytics, and automated trading systems.

Model Description

• Architecture: BERT (for sequence classification)
• Framework: PyTorch, Transformers
• Topics: 20 financial news categories (see below)
• License: MIT

Intended Uses & Limitations

• Intended Use:
  • Classify financial news headlines or short texts into one of 20 financial topics.
  • Use in financial analytics, news monitoring, and trading agent pipelines.
• Limitations:
  • Trained on zeroshot/twitter-financial-news-topic; may not generalize to all financial news sources.
  • Not suitable for non-financial or long-form text.

Topics

Example Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline

tokenizer = AutoTokenizer.from_pretrained("leonas5555/finnews-topic-single-classify")
model = AutoModelForSequenceClassification.from_pretrained("leonas5555/finnews-topic-single-classify")

nlp = pipeline("text-classification", model=model, tokenizer=tokenizer)

# Example text
text = "LIVE: ECB surprises with 50bps hike, ending its negative rate era. President Christine Lagarde is taking questions"

result = nlp(text)
print(result)
# Output: [{'label': 'Fed | Central Banks', 'score': 0.98}]

Example Inputs & Outputs

Training Data

• Dataset: zeroshot/twitter-financial-news-topic
• Size: 21 107 samples
• Class Distribution: Unbalanced; class weights used during training.

Training Procedure

• Framework: HuggingFace Transformers (Trainer API)

• Arguments:

  • num_train_epochs: 10
  • per_device_train_batch_size: 32
  • per_device_eval_batch_size: 32
  • gradient_accumulation_steps: 1
  • learning_rate: 2e-5
  • fp16: True (Native AMP mixed precision)
  • warmup_ratio: 0.1
  • label_smoothing_factor: 0.05
  • max_grad_norm: 1.0
  • max_length: 256
  • evaluation_strategy: "steps"
  • save_strategy: "steps"
  • save_total_limit: 3
  • load_best_model_at_end: True
  • metric_for_best_model: "f1"
  • run_name: "topic_classifier"
  • seed: 42

• Early Stopping: Patience of 2 evaluation steps (via EarlyStoppingCallback)

• Optimizer: Adam (betas=(0.9, 0.999), epsilon=1e-08)

• Scheduler: Linear

• Metrics: F1 (for best model selection), plus accuracy, precision, recall

Evaluation Results

ONNX Export

An ONNX version of this model {TBD} for use with high-performance inference engines such as Infinity.

optimum-cli export onnx -m leonas5555/finnews-topic-single-classify

License

MIT

Inspired by:

• nickmuchi/finbert-tone-finetuned-finance-topic-classification

---

References:

• fuchenru/Trading-Hero-LLM