FIN-RoBERTa-Custom: Financial Sentiment Analysis

A fine-tuned RoBERTa model for financial sentiment analysis that classifies text into negative, neutral, or positive sentiment.

Model Description

This model is a fine-tuned version of roberta-base specifically trained for financial text sentiment classification. It achieves 99.6% accuracy and 99.5% F1 (macro) on the Financial PhraseBank (100% agreement) benchmark.

Labels

Label	ID	Description
negative	0	Negative financial sentiment
neutral	1	Neutral financial sentiment
positive	2	Positive financial sentiment

Usage

from transformers import RobertaForSequenceClassification, RobertaTokenizer
import torch
# Load model and tokenizer
model_name = "alasteirho/FIN-RoBERTa-Custom"
tokenizer = RobertaTokenizer.from_pretrained(model_name)
model = RobertaForSequenceClassification.from_pretrained(model_name)
model.eval()
# Predict sentiment
def predict_sentiment(text):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)
    with torch.no_grad():
        outputs = model(**inputs)
        probs = torch.softmax(outputs.logits, dim=1)
        pred = torch.argmax(probs, dim=1).item()
    labels = ["negative", "neutral", "positive"]
    return labels[pred], probs[0].tolist()
# Example
text = "The company reported strong quarterly earnings, beating analyst expectations."
sentiment, confidence = predict_sentiment(text)
print(f"Sentiment: {sentiment}")
print(f"Confidence: {confidence}")

Using Pipeline (Simple)

from transformers import pipeline
classifier = pipeline("sentiment-analysis", model="alasteirho/FIN-RoBERTa-Custom")
result = classifier("Revenue declined 15% year-over-year due to weak demand.")
print(result)

Training Data

The model was trained on a combined dataset from five financial sentiment sources:

Dataset	Samples	Description
Financial PhraseBank (50% agreement)	4,846	Financial news sentences annotated by 16 domain experts
Twitter Financial News Sentiment	9,543	Financial tweets with sentiment labels
FiQA 2018	961	Financial question-answer pairs with continuous sentiment scores
SemEval 2017 Task 5 Subtask 2	1,142	Financial news headlines with continuous sentiment scores
SemEval 2017 Task 5 Subtask 1	1,700	Financial microblogs (tweets and StockTwits) with continuous sentiment scores

Datasets with continuous sentiment scores (FiQA, SemEval Subtask 1, and SemEval Subtask 2) were discretised into three classes using fixed thresholds (below -0.2 = negative, above +0.2 = positive, otherwise neutral).

Total training samples: 16,127 (after deduplication and cleaning) Validation samples: 3,774

Evaluation Results

Financial PhraseBank (100% Agreement) - Held-out Test Set (2,264 samples)

Metric	Score
Accuracy	0.9956
F1 (Macro)	0.9951

Per-Class Performance (FPB All Agree)

Class	Precision	Recall	F1-Score	Support
Negative	1.000	0.990	0.995	303
Neutral	0.995	0.998	0.996	1,391
Positive	0.995	0.993	0.994	570

Combined Validation Set (3,774 samples)

Metric	Score
Accuracy	0.8755
F1 (Macro)	0.8629
F1 (Weighted)	0.8746

Per-Class Performance (Combined Validation)

Class	Precision	Recall	F1-Score	Support
Negative	0.86	0.83	0.84	588
Neutral	0.88	0.81	0.84	1,171
Positive	0.88	0.93	0.90	2,015

Comparison with FinBERT (Combined Validation Set)

Metric	FIN-RoBERTa-Custom	FinBERT	Delta
Accuracy	0.8755	0.4290	+0.4465
F1 (Macro)	0.8631	0.4672	+0.3959
F1 (Weighted)	0.8746	0.4097	+0.4648

Training Procedure

Hyperparameters

Parameter	Value
Base Model	roberta-base
Max Epochs	30
Early Stopping Patience	8 (on F1 macro)
Best Checkpoint	Epoch 12
Training stopped at	Epoch 20
Batch Size (per device)	16
Gradient Accumulation Steps	2
Effective Batch Size	32
Learning Rate	2e-5
LR Scheduler	Cosine with warmup
Warmup Ratio	6%
Weight Decay	0.01
Max Sequence Length	128
Precision	FP16 (mixed precision)
Seed	42

Training was performed on an NVIDIA GeForce RTX 5080 GPU (16 GB) using the Hugging Face Transformers library with PyTorch. The training run completed in approximately 14 minutes, consuming 3.35 GB of peak GPU memory.

Intended Use

This model is designed for:

• Financial news sentiment analysis
• Stock/market sentiment classification
• Financial social media sentiment analysis
• Earnings call/report sentiment extraction

Limitations

• Trained primarily on English financial text
• May not generalise well to non-financial domains
• Performance may vary on very long documents (max 128 tokens)
• Cryptocurrency and emerging market terminology may be underrepresented

Citation

If you use this model, please cite:

@misc{alasteir_ho_2026,
    author       = { Alasteir Ho },
    title        = { FIN-RoBERTa-Custom (Revision 1fb31e2) },
    year         = 2026,
    url          = { https://huggingface.co/alasteirho/FIN-RoBERTa-Custom },
    doi          = { 10.57967/hf/8367 },
    publisher    = { Hugging Face }
}

License

MIT License