Fine-tuned FinBERT (ProsusAI/finbert) that classifies a stock-market / financial news headline or sentence as bullish / neutral / bearish (mapped to BUY / HOLD / SELL signals in the demo app).

Model details

Base model: ProsusAI/finbert (BERT-base, pretrained on financial text)
Task: 3-class text classification — bearish (0), neutral (1), bullish (2)
Language: English
Fine-tuned by: shubhampm
Demo: shubhampm/stock-news-sentiment (HF Space) — FastAPI POST /predict + Gradio UI

Training data

lmassaron/FinancialPhraseBank — 3,872 financial news sentences annotated by finance experts (Malo et al., 2014, "Good Debt or Bad Debt"). 90/10 train/test split (seed 42).

Training procedure

Fine-tuned with Hugging Face Trainer on a Kaggle GPU:

Hyperparameter	Value
Epochs	3
Batch size	16
Learning rate	2e-5 (AdamW)
Max sequence length	128

Evaluation (held-out 10% test split, n=388)

Metric	Value
Accuracy	0.825
Macro F1	0.811

Class	Precision	Recall	F1
bearish	0.81	0.78	0.80
neutral	0.92	0.80	0.85
bullish	0.70	0.90	0.78

Example predictions

python

from transformers import pipeline
pipe = pipeline("text-classification", model="shubhampm/stock-news-sentiment-finetuned")
pipe("Company X beats earnings estimates, shares surge 8%")
# bullish (BUY) — confidence ≈ 0.93
pipe("Regulator opens probe into accounting practices at the company")
# bearish (SELL)

Limitations

Trained on short financial sentences; long articles are truncated at 512 tokens.
English only; forced to choose one of 3 classes even for non-financial input (the demo app labels predictions below 55% confidence as UNCLEAR).
Not investment advice — a small educational fine-tune, not a trading system.