HKUST's FinBERT-Tone Reads Financial Filings With 89.7% Negative Accuracy
A compact BERT fine-tuned on 10,000 hand-labeled analyst report sentences keeps racking up downloads for one specialized job: labeling financial text tone.
Takeaways
- finbert-tone is a BERT model fine-tuned on 10K hand-labeled analyst report sentences for financial sentiment.
- Base FinBERT was pretrained on 4.9B tokens from 10-Ks, earnings calls, and analyst reports.
- Reports 89.7% accuracy on negative sentiment vs under 60% for non-BERT baselines.
- Beats vanilla BERT (85.0%) thanks to finance-domain pretraining, per the CAR paper.
- Drops into Hugging Face
pipeline; three-class output (positive, negative, neutral) at sentence granularity. - Approaching 1M downloads and used in 100+ Spaces on Hugging Face.
Generic sentiment classifiers trained on tweets and product reviews often misread financial language. Terms such as liability, volatility, and restructuring depend on accounting context, company guidance, and market expectations. FinBERT-Tone, a checkpoint published by Yi Yang at HKUST, classifies that language as positive, negative, or neutral at sentence level.
At the time of writing, the Hugging Face model page showed nearly one million downloads and more than 100 Spaces using the checkpoint. Its compact size and fixed three-class output suit systems that need to score large collections of filings, earnings calls, analyst reports, and financial news.
Finance is in the training data
FinBERT-Tone builds on FinBERT, a BERT model pretrained on 4.9 billion tokens of financial communication. The corpus concentrates on documents written and read by companies, analysts, and investors.
| Corpus | Training volume |
|---|---|
| Corporate 10-K and 10-Q filings | 2.5 billion tokens |
| Earnings-call transcripts | 1.3 billion tokens |
| Analyst reports | 1.1 billion tokens |
The tone checkpoint adds supervised training on 10,000 manually annotated sentences from analyst reports. Each sentence carries one of three labels: neutral, positive, or negative. This alignment between the training material and the target task gives the model direct exposure to hedged forecasts, accounting terms, and analyst phrasing.
BERT converts each token into a representation informed by the surrounding words. The classification head then emits three logits, which a softmax function converts into class probabilities. The checkpoint maps class ID 0 to neutral, 1 to positive, and 2 to negative.
Negative language drives the benchmark gap
The accompanying FinBERT paper reports 89.7% accuracy for detecting negative sentiment, while the evaluated non-BERT methods remained below 60%. FinBERT also surpassed the vanilla BERT comparison, which reached 85.0% accuracy, along with the paper’s CNN and LSTM baselines.