Aadhil-rog/finbert-indian-sentiment-v2

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 → Negative
  • 1 → Neutral
  • 2 → 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.

---