FinBERT-Indian-Sentiment
Overview
FinBERT-Indian-Sentiment is a fine-tuned financial sentiment analysis model designed specifically for Indian financial news and market-related text.
The model classifies input text into three sentiment categories:
- Negative
- Neutral
- Positive
It is based on ProsusAI/finbert and fine-tuned on an India-focused financial news dataset to better capture domain-specific language such as RBI policy statements, market movements, and macroeconomic commentary.
Motivation
Generic financial sentiment models often struggle with:
- Indian market terminology
- RBI policy language
- Macro-economic neutrality
- Mixed-signal financial news
This model aims to improve sentiment understanding in the Indian financial context, where cautious and neutral language is common.
Training Data
- Dataset:
kdave/Indian_Financial_News
- Total samples: ~22,000
- Classes: Negative, Neutral, Positive
- Split: 85% training / 15% test (stratified)
The dataset consists of Indian financial news articles covering:
- Stock markets
- Banking and finance
- RBI announcements
- Corporate earnings
- Macroeconomic indicators
Model Details
- Base model: ProsusAI/finbert
- Architecture: BERT-based sequence classification
- Number of labels: 3
- Label mapping:
0 → Negative1 → Neutral2 → Positive
- Max sequence length: 512
- Framework: PyTorch / Hugging Face Transformers
Evaluation Results
Evaluation was performed on a held-out test set.
| Metric | Score |
|---|
| Accuracy | ~0.89 |
| Weighted F1-score | ~0.89 |
Confusion Matrix Summary
- Strong diagonal dominance across all classes
- Minimal confusion between positive and negative
- Neutral sentiment remains the most challenging class (expected for financial text)
- False positives and false negatives remain below 10% across classes
These results indicate balanced and reliable performance suitable for real-world applications.
Intended Use
This model is suitable for:
- Financial news sentiment analysis
- Market sentiment monitoring
- Academic and research projects
- NLP experimentation in finance
- Backend APIs for sentiment classification
Limitations
- Long, mixed-signal macroeconomic articles may lead to overconfident predictions
- Neutral sentiment may lean toward positive or negative in ambiguous cases
- Confidence calibration may be required for high-stakes production use
⚠️ This model is not intended for investment advice or automated trading decisions.
