FinRoBERTa-Mendeley
A fine-tuned DistilRoBERTa model for financial sentiment analysis,
trained on a balanced dataset of financial news headlines.
It predicts Positive, Neutral, or Negative sentiment for text related to the stock market, company earnings, or financial trends.
Model Details
| Property | Description |
|---|
| Base model | distilroberta-base |
| Task | Financial sentiment classification (3 labels) |
| Dataset | Balanced Financial News Headlines (≈18k samples) |
| Languages | English |
| Domain | Finance / Stock Market / Economy |
| Fine-tuning epochs | 3 |
| Batch size | 16 |
| Optimizer | AdamW (lr=2e-5) |
| Framework | PyTorch + Transformers |
| Accuracy | ~78% |
| F1 (macro) | ~0.72 |
Labels
| ID | Label | Description |
|---|
| 0 | Negative | Bearish / pessimistic financial outlook |
| 1 | Neutral | Mixed or uncertain tone |
| 2 | Positive | Bullish / optimistic market sentiment |
Usage
1from transformers import AutoTokenizer, AutoModelForSequenceClassification
2import torch
3import torch.nn.functional as F
4
5# Load model and tokenizer
6model_name = "AurelPx/FinRoBERTa-Mendeley"
7device = "cuda" if torch.cuda.is_available() else "cpu"
8
9tokenizer = AutoTokenizer.from_pretrained(model_name)
10model = AutoModelForSequenceClassification.from_pretrained(model_name).to(device)
11
12text = "Tesla shares rally after strong earnings report"
13
14# Tokenize and move to correct device
15inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True).to(device)
16
17# Predict
18with torch.no_grad():
19 logits = model(**inputs).logits
20
21# Compute probabilities and predicted label
22probs = F.softmax(logits, dim=-1)
23pred = probs.argmax().item()
24
25labels = ["Negative", "Neutral", "Positive"]
26print(f"Sentence: {text}")
27print(f"Predicted Sentiment: {labels[pred]}")
28print(f"Probabilities: {probs.cpu().numpy()}")
Evaluation
The model was evaluated on a held-out test set of 20% of the dataset.
| Metric | Score |
|---|
| Accuracy | 0.78 |
| F1 (macro) | 0.72 |
| Precision | 0.75 |
| Recall | 0.76 |
Performance Comparison
| Model | Accuracy | F1 Macro | Observations |
|---|
| CardiffNLP/twitter-roberta-base-sentiment | 0.44 | 0.36 | Performs poorly on financial news; biased toward Neutral |
| Sigma/financial-sentiment-analysis | 0.55 | 0.51 | More balanced, but weak recall on Negative headlines |
| RashidNLP/Finance-Sentiment-Classification | 0.41 | 0.35 | Strong Positive recall but poor Neutral precision |
| 🟢 FinRoBERTa-Mendeley | 0.78 | 0.72 | Best overall balance across all three classes |
The model shows robust generalization to finance-related texts such as:
-
Company earnings releases
-
Market commentary
-
Analyst outlooks
-
Economic updates
Citation
@misc{distilroberta_financial_sentiment_mendeley,
title={FinRoBERTa-Mendeley},
author={Your Name},
year={2025},
publisher={Hugging Face},
howpublished={\url{
https://huggingface.co/your-username/FinRoBERTa-Mendeley}},
note={Fine-tuned DistilRoBERTa for financial sentiment analysis on the Mendeley/Financial News dataset}
}
Limitations & Future Work
The dataset is limited to English financial headlines.
It may not generalize perfectly to social media or retail investor sentiment (e.g., Reddit).
Next step: multi-source fine-tuning on tweets, Reddit posts, and news summaries.
License
MIT License — free for commercial and research use.
