Sentiment Analysis on Financial News (HuggingFace DistilBERT)

Fine-tune DistilBERT on financial headlines for 3-class sentiment (Positive / Neutral / Negative). Tracked with MLflow. Deployed via Azure ML endpoint. Project Structure

DS2_FinSentiment__config.py       ← Model + training params
DS2_FinSentiment__data_gen.py     ← 300-headline labelled dataset
DS2_FinSentiment__model.py        ← DistilBERT setup, tokenisation, metrics
DS2_FinSentiment__trainer.py      ← MLflow-tracked training loop
DS2_FinSentiment__inference.py    ← Real-time inference pipeline
DS2_FinSentiment__dashboard.py    ← Results visualisation
DS2_FinSentiment__main.py         ← Entry point
DS2_FinSentiment__requirements.txt

Run

pip install -r DS2_FinSentiment__requirements.txt
python DS2_FinSentiment__main.py

Results

+12% accuracy over zero-shot FinBERT baseline

Macro F1 improved from ~0.68 → ~0.81

Model saved to ./fin_sentiment_model (reload without retraining)