finsentiment-deberta-v3-base
Fine-tuned microsoft/deberta-v3-base for three-class financial sentiment classification
(negative / neutral / positive) on a combined corpus of:
- Financial PhraseBank (via the
nickmuchi/financial-classificationmirror) - Twitter Financial News Sentiment (
zeroshot/twitter-financial-news-sentiment)
Test macro-F1: 0.876 (accuracy 0.892, ECE 0.034).
Evaluation
| Metric | Fine-tuned | FinBERT (zero-shot) |
|---|---|---|
| macro-F1 | 0.876 | 0.731 |
| accuracy | 0.892 | 0.765 |
| F1 (negative) | 0.858 | 0.661 |
| F1 (neutral) | 0.914 | 0.819 |
| F1 (positive) | 0.856 | 0.712 |
| Expected Calibration Error | 0.034 | 0.075 |
Evaluation was run on a stratified 10% test split of the combined corpus
(seed 42), with ProsusAI/finbert as a zero-shot domain-pretrained baseline.
How to use
from transformers import pipeline
clf = pipeline("text-classification", model="anabdd/finsentiment-deberta-v3-base", top_k=None)
clf("Apple beat earnings expectations on iPhone strength.")
# → [{'label': 'positive', 'score': 0.97}, ...]
Training details
Trained with Trainer (HuggingFace transformers >= 4.46) using:
- Class-weighted cross-entropy loss (sklearn
balancedweighting on the training split) - Label smoothing (factor 0.05)
- Cosine LR schedule with 10% warmup
- Mixed-precision (fp16) on a single NVIDIA T4 GPU
See the project repository for the full training script
(src/train.py) and the analysis notebook.
Intended use & limitations
Designed for short-form financial English text (headlines, tweets, brief analyst notes). It has not been trained for:
- Languages other than English
- Long-form financial documents (10-Ks, earnings call transcripts)
- Fine-grained sentiment beyond three classes
Labels follow the nickmuchi/financial-classification mirror's convention,
which is itself derived from the original PhraseBank annotation scheme.