FinBERT’s two primary Hugging Face models combine for over 4.3 million monthly downloads as of 2024, making it the most-used domain-specific language model in finance. The NLP in finance market reached $6.65 billion in the same year and is on track for $8.34 billion by the end of 2025. This article covers the latest FinBERT statistics, including download trends, accuracy benchmarks, enterprise adoption rates, and market data through early 2026.
FinBERT Statistics in 2026 — TL;DR
ProsusAI/finbert logged 2.3 million monthly downloads on Hugging Face, while yiyanghkust/finbert-tone reached 2.1 million.
FinBERT hit 97% accuracy on Financial PhraseBank sentences with full annotator agreement — 8 points above general BERT.
The model was pre-trained on 4.9 billion tokens from corporate reports, earnings transcripts, and analyst reports.
Hedge funds lead enterprise adoption at 68%, followed by investment banks at 56%.
Sentiment analysis holds 31% of the financial NLP market, the single largest application segment.
FinBERT is a BERT-based NLP model built by Prosus AI for financial sentiment classification. It processes earnings calls, 10-K filings, and analyst reports, outputting positive, negative, or neutral labels. The model’s edge over general-purpose financial language models comes from domain-specific pre-training on 4.9 billion financial tokens. Combined monthly downloads across both primary variants top 4.3 million, with 92 fine-tuned derivatives built by the community.
How Many Downloads Does FinBERT Get?
ProsusAI/finbert recorded 2.3 million monthly downloads as of 2024. The yiyanghkust/finbert-tone variant reached 2.1 million. Together, they exceed 4.3 million monthly downloads, based on Hugging Face platform data.
ProsusAI/finbert has accumulated over 6 million total downloads and 1,097 likes on Hugging Face. That version spawned 74 fine-tuned derivatives and 3 quantized versions. The yiyanghkust variant, trained on 4.9 billion tokens, generated 18 fine-tuned adaptations for specialized use cases.
| FinBERT Variant | Monthly Downloads | Fine-Tuned Derivatives |
|---|---|---|
| ProsusAI/finbert | 2.3 million | 74 |
| yiyanghkust/finbert-tone | 2.1 million | 18 |
| Combined | 4.3+ million | 92 |
Source: Hugging Face, Endor Labs
FinBERT Statistics: Accuracy and Benchmarks
FinBERT achieved 97% accuracy on Financial PhraseBank sentences where all annotators agreed on the label. On the complete dataset with all annotations included, accuracy was 86%. The original FinBERT paper, published in Contemporary Accounting Research, reported an 8 percentage point gap over standard BERT on financial sentiment tasks.
Only 5% of misclassifications occurred between negative and positive categories, meaning the model rarely confuses opposite polarities. FinBERT also maintained 81.3% accuracy using just 10% of the training data — a level that general BERT could not match even with 20% of the same data.
| Model | Accuracy | Gap vs FinBERT |
|---|---|---|
| FinBERT | 97% | — |
| ULMFiT | 93% | -4 pts |
| BERT (general) | 89% | -8 pts |
| LSTM | 84% | -13 pts |
| CNN | 82% | -15 pts |
| Random Forest | 78% | -19 pts |
| Loughran-McDonald Dictionary | 71% | -26 pts |
Source: Contemporary Accounting Research (Huang, Wang & Yang)
FinBERT Training Data and Model Architecture
FinBERT was pre-trained on 4.9 billion tokens. Corporate reports (10-K and 10-Q filings) made up 2.5 billion tokens or 51% of the corpus. Earnings call transcripts contributed 1.3 billion tokens at 27%. Analyst reports accounted for the remaining 1.1 billion tokens at 22%.
The architecture follows BERT-base specifications similar to other domain-adapted models: 110 million parameters, 12 attention heads, and a 768-dimension hidden layer. Fine-tuning used 10,000 manually annotated sentences from analyst reports across positive, negative, and neutral labels. FinBERT employs six pre-training tasks compared to BERT’s two.
| Specification | Value |
|---|---|
| Total Parameters | 110 million |
| Attention Heads | 12 |
| Hidden Dimension | 768 |
| Pre-Training Tokens | 4.9 billion |
| Fine-Tuning Sentences | 10,000 |
| Output Labels | Positive, Negative, Neutral |
| Pre-Training Tasks | 6 |
| Frameworks Supported | PyTorch, TensorFlow, JAX |
Source: Hugging Face model card, HKUST research
Financial NLP Market Size and Growth
The NLP in finance market was valued at $6.65 billion in 2024 and is projected to reach $8.34 billion by 2025, according to The Business Research Company. That is a 25.5% compound annual growth rate. Broader NLP market projections from MarketsandMarkets put the global NLP market at $70.11 billion in 2026, growing to $249.97 billion by 2031 at 29% CAGR.
Sentiment analysis represents 31% of the financial NLP market — the largest single segment. Risk management and fraud detection account for 24%. Customer service automation holds 19%. North America leads with 36% market share, and software solutions make up 65% of total industry spending. Financial institutions alongside general-purpose AI tools generate unstructured data at rates between 55% and 65% annually, based on John Snow Labs estimates.
FinBERT Enterprise Adoption by Sector
Hedge funds lead FinBERT adoption at 68%, using sentiment scores as alpha signals for quantitative trading strategies. Investment banks follow at 56%, primarily for processing analyst reports and market research. Asset managers report 48% adoption, concentrated on ESG screening automation.
Retail banking sits at 38%, with customer feedback analysis as the main application. Research from Credgenics showed that sentiment-driven analytical approaches in lending helped recover 70% to 95% of bad debts, with collection rates rising 15% to 20% through automated analysis.
| Sector | Adoption Rate | Primary Use Case |
|---|---|---|
| Hedge Funds | 68% | Quantitative trading signals |
| Investment Banks | 56% | Analyst report processing |
| Asset Management | 48% | ESG screening |
| Retail Banking | 38% | Customer feedback analysis |
Source: CompaniesHistory.com industry analysis
FinBERT Statistics: ESG and Specialized Variants
FinBERT outperformed traditional methods by 18% on ESG text classification, based on HKUST research published in January 2026. The model identifies discussions related to environmental, social, and governance issues in corporate filings and analyst reports more accurately than dictionary-based approaches.
Specialized variants extend its reach. FinBERT-ESG targets sustainability reporting. FinBERT-FLS identifies forward-looking statements in regulatory disclosures. FinBERT-FOMC, developed by Ziwei Chen, was fine-tuned on FOMC meeting minutes from 2006 to 2023 for monetary policy sentiment. FinBERT2 was introduced at ACM SIGKDD 2025 as a successor following a pattern seen in other domain-specific BERT models where second-generation variants improve on the original.
Financial NLP Market by Application Segment
How Does FinBERT Compare to BloombergGPT?
FinBERT and BloombergGPT target different ends of the financial NLP stack. FinBERT has 110 million parameters and is open-source on Hugging Face. BloombergGPT has 50.6 billion parameters and remains closed within Bloomberg’s infrastructure. FinBERT’s strength is focused sentiment classification. BloombergGPT handles a wider range of tasks — named entity recognition, news classification, and query translation — but at far higher computational cost.
GPT-4 outperformed BloombergGPT on most financial benchmarks, scoring 68.79% on FinQA. FinBERT’s 97% accuracy applies to its specific task of financial sentiment classification, where smaller models with domain training can still outperform much larger general-purpose ones. For teams that need comparisons with broader AI platforms, the tradeoff is specialization versus scope.
FinBERT Research Impact and Community
The primary publication appeared in Contemporary Accounting Research in 2023 by Allen Huang, Hui Wang, and Yi Yang. The ProsusAI GitHub repository holds 2,100 stars and 511 forks. Hugging Face community discussions generated 49 threads across both primary variants.
Research building on FinBERT continued through 2025 into 2026. A November 2025 study in MDPI Electronics fine-tuned FinBERT for sector-specific financial news with a macro F1 score of 0.707 after fine-tuning. The EMNLP 2025 Workshop on Financial Technology featured work on model compression via quantization. A January 2026 HKUST paper confirmed FinBERT captures 18% more textual informativeness from earnings calls than baseline models. These studies appear in IEEE Big Data, ACM AI in Finance, and other venues that track domain-specific NLP models.
| Metric | Value |
|---|---|
| GitHub Stars | 2,100+ |
| GitHub Forks | 511 |
| Hugging Face Likes | 1,097 |
| Community Threads | 49 |
| Total Derivatives | 92 |
| Primary Publication | Contemporary Accounting Research (2023) |
Source: GitHub, Hugging Face
FAQs
What accuracy does FinBERT achieve on financial sentiment analysis?
FinBERT reaches 97% accuracy on Financial PhraseBank sentences with full annotator agreement. On the complete dataset, it scores 86%. Both figures outperform general BERT by 8 percentage points.
How many downloads does FinBERT get each month?
The two primary variants on Hugging Face combine for over 4.3 million monthly downloads. ProsusAI/finbert accounts for 2.3 million and yiyanghkust/finbert-tone for 2.1 million.
What data was FinBERT trained on?
FinBERT was pre-trained on 4.9 billion tokens from three sources: 2.5 billion from corporate filings, 1.3 billion from earnings call transcripts, and 1.1 billion from analyst reports.
Which financial sectors use FinBERT the most?
Hedge funds lead at 68% adoption. Investment banks follow at 56%, asset management at 48%, and retail banking at 38%, each with different primary applications.
How large is the financial NLP market?
The NLP in finance market was valued at $6.65 billion in 2024 and is projected to reach $8.34 billion in 2025, growing at a 25.5% compound annual growth rate.
Sources:
https://huggingface.co/ProsusAI/finbert
https://www.marketsandmarkets.com/Market-Reports/natural-language-processing-nlp-825.html
