🐂 FinSense distilbert v2 — financial news sentiment, tiny and fast

v2 — retrained on a cleaner recipe with a published split. Same 67M-parameter speed your pipelines already rely on, more accurate than v1 on a properly held-out benchmark.

from transformers import pipeline

clf = pipeline("text-classification", model="AnkitAI/distilbert-base-uncased-financial-news-sentiment-analysis")
clf("The company's quarterly earnings surpassed all estimates.")
# [{'label': 'positive', 'score': 0.99}]

positive / neutral / negative for headlines, news wires, analyst sentences. Built on ModernBERT-base — Flash-Attention-fast, 149M params, runs happily on CPU.

Benchmarks

Financial PhraseBank (the standard benchmark for this task), held-out test set, identical harness for every row:

Model	Accuracy	Macro-F1
🐂 This model (v2)	0.8447	0.8316
v1 (previous weights)	0.8323	0.8064
ProsusAI/finbert¹	0.8799	0.8761

+1.2 accuracy / +2.5 F1 over v1, at a third of FinBERT's size. We do not claim to beat FinBERT: it scores higher on this table, having been trained on effectively the whole corpus this split comes from.¹ Want maximum accuracy? The ModernBERT flagship scores 0.8675.

_{¹ Measured by us on the identical split — not quoted from another paper. FinBERT was trained on effectively all of Financial PhraseBank, so its score here reflects memorisation of the corpus rather than generalisation; a fair comparison needs data neither model has seen. A previous version of this card reported FinBERT at 0.8423/0.8439 citing an independent replication that could not be verified; both the number and the claim resting on it have been removed.}

Labels

id	label	example
0	negative	"Operating profit fell to EUR 35.4 mn from EUR 68.8 mn."
1	neutral	"The annual general meeting will be held on April 12."
2	positive	"Quarterly earnings surpassed all estimates."

Batch scoring (thousands of headlines):

headlines = ["Shares jumped 8% after the guidance raise.",
             "The company filed its annual report on Thursday.",
             "Regulators fined the bank EUR 20 mn."]
for h, r in zip(headlines, clf(headlines, batch_size=32)):
    print(f"{r['label']:<9} {r['score']:.2f}  {h}")

Built for

Trading & research pipelines — score news flow at scale (fast batch inference, CPU-friendly)
Fintech products — sentiment tags for news feeds, alerts, dashboards
Quant & academic work — reproducible split + eval script included, cite with confidence

Good to know

Tuned for financial news register — tweets and Reddit are a different dialect
English, sentence-level, three classes
Errors concentrate on positive-vs-neutral — the same boundary human annotators disagree on 25% of the time (structural ceiling of this task, affects every model including FinBERT)

Training details

Full fine-tune of distilbert-base-uncased on Financial PhraseBank (sentences_50agree, 4,846 expert-annotated sentences): 5 epochs, lr 2e-5, batch 16, max length 128, fp32, best checkpoint by validation macro-F1. Stratified 80/10/10 split with a fixed, published seed — the split script and raw evaluation outputs are in this repo, so every number above is reproducible end-to-end.

Support the Project

If this model is useful in your work, you can support independent research:

Citation

@misc{finsense2026,
  author = {Aglawe, Ankit},
  title = {FinSense: Financial News Sentiment Models},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/AnkitAI/FinSense-ModernBERT-Financial-News-Sentiment-Analysis}
}

Base & license

Apache-2.0 weights (ModernBERT-base, Answer.AI). Trained on Financial PhraseBank (Malo et al., 2014 — CC BY-NC-SA; commercial users, check dataset terms).

The FinSense family

Model	Size	Accuracy	Pick it for
FinSense ModernBERT	149M	0.8675	best accuracy, modern stack
This model (v2)	67M	0.8447	smallest & fastest

More sizes and a multilingual variant are on the roadmap. Sibling series: Parable — local agent LLMs from the same maker.

Version history

v2 (2026-07-17) — this release, in place: cleaner recipe, published stratified split (seed 42), honest held-out benchmark. Same labels, same API — drop-in for v1 users.
v1 (2024-11) — original release (0.9669 self-reported on the small allagree subset — not comparable to the held-out 50agree numbers above). Preserved in revision history.

More on the FinSense models: ankitaglawe.com/finsense