AION-Sentiment-IN-v3

Financial sentiment analysis model built by AION Analytics.

Model Details

Developer: AION Analytics

Model Type: Transformer-based sequence classification

Input: Text (financial news headlines)

Output: Sentiment classification (negative, neutral, positive)

License: Apache License 2.0

Version: 3.0.0 (March 2026)

Model Description

AION-Sentiment-IN-v3 is a sentiment analysis model built by AION Analytics for Indian financial news. This model uses AION's proprietary taxonomy-based labeling system with 136 predefined market events.

Key Features

India-Tuned: Trained on 400K Indian financial news headlines from 2024-2026
AION Taxonomy Labels: Training labels derived from 136 known sentiment events
Three-Class Classification: negative, neutral, positive
99.63% Accuracy: Validated on 100K holdout test set
2x Improvement: Fixed critical misclassifications from v2 (e.g., "Markets Crashing" now correctly predicts negative)

What Changed from v2

Issue	v2	v3 (AION Taxonomy)
"Markets Crashing"	Positive ❌	Negative ✅
"TCS Record Profits"	Neutral ❌	Positive ✅
Validation Accuracy	98.55%	99.63%
Test Headlines (6 cases)	33% (2/6)	67% (4/6)

Training Data

Source: AION Analytics news corpus (Indian financial news)
Size: 400,000 headlines (100,000 validation)
Labeling Method: AION Taxonomy event sentiment
- 136 events with known sentiment (e.g., "RBI repo hike" = negative, "Record earnings" = positive)
- Headlines matched to events → assigned correct sentiment
Time Period: 2024-01-01 to 2026-03-31
Label Distribution:
- negative: 78,066 (15.6%)
- neutral: 246,990 (49.4%)
- positive: 175,410 (35.0%)

AION Taxonomy Methodology

AION's taxonomy uses 136 predefined market events with known sentiment and sector impacts. Each event has:

base_impact: Sentiment strength (mild, normal, severe)
sector_impacts: How the event affects different sectors
market_weight: Overall market importance

This methodology ensures consistent, accurate labels that understand financial context.

Training Procedure

Framework: PyTorch with HuggingFace Transformers
Architecture: Transformer encoder
Epochs: 3
Batch Size: 32
Learning Rate: 2e-05
Hardware: Apple M4 (MPS acceleration)
Training Time: ~10 hours

Evaluation

Validation Set Metrics (100K samples)

Epoch	Accuracy	F1 Score	Loss
1	99.65%	99.56%	0.0100
2	99.63%	99.54%	0.0103
3 (Final)	99.63%	99.54%	0.0098

Per-Class Performance (Epoch 3)

Class	Precision	Recall	F1
negative	0.99	0.99	0.99
neutral	0.99	0.99	0.99
positive	0.99	0.99	0.99

Test Headlines (6 Problematic Cases)

Headline	Expected	v2 Prediction	v3 Prediction
"Markets Crashing"	negative	positive ❌	negative ✅
"Stocks to buy...10-30% return"	positive	neutral ❌	neutral ❌
"Gold slides over 3%"	negative	negative ✅	negative ✅
"RBI hikes repo rate"	negative	neutral ❌	neutral ❌
"TCS record earnings"	positive	neutral ❌	positive ✅
"Market crashes on recession"	negative	negative ✅	negative ✅

Accuracy: v2 = 33% (2/6) → v3 = 67% (4/6) ✅

Limitations

Ambiguous Headlines: May misclassify headlines that don't clearly signal sentiment:
- "Stocks to buy in 2026" → predicts neutral (should be positive)
- "RBI hikes repo rate by 25 bps" → predicts neutral (should be negative)
These are being addressed in future taxonomy updates.
Overconfidence: Model often returns 100% confidence scores. Use confidence scores as relative indicators rather than absolute probabilities.
Domain Specificity: Works best on Indian financial news. Performance may degrade on:
- Non-financial text
- Non-Indian market context
- Social media or informal text
Taxonomy Coverage: ~40% of headlines match taxonomy events directly.

Use Cases

Intended Uses

Financial News Analysis: Automated sentiment classification of Indian financial news headlines
Market Research: Sentiment trend analysis across sectors or time periods
Portfolio Monitoring: Track sentiment around specific stocks or sectors
Academic Research: Study sentiment-market relationships in emerging markets

Out-of-Scope Uses

Direct trading decisions without human oversight
Non-financial sentiment analysis
Non-English text
Real-time trading systems (use internal production systems)

Usage

Using aion-sentiment Package (Recommended)

from aion_sentiment import SentimentAnalyzer

analyzer = SentimentAnalyzer()  # Uses v3 by default
result = analyzer.predict("RBI hikes repo rate by 25 bps")
print(result)
# {'label': 'neutral', 'confidence': 0.89}

Using Transformers Directly

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="aion-analytics/aion-sentiment-in-v3"
)
result = classifier("RBI hikes repo rate by 25 bps")
print(result)
# [{'label': 'LABEL_1', 'score': 0.89}]
# Note: LABEL_0=negative, LABEL_1=neutral, LABEL_2=positive

Batch Processing

from aion_sentiment import SentimentAnalyzer
import pandas as pd

analyzer = SentimentAnalyzer()

# Single prediction
result = analyzer.predict("Stock market reaches all-time high")

# Batch prediction
texts = [
    "RBI hikes repo rate",
    "Stock market crashes",
    "TCS reports record earnings"
]
results = analyzer.predict(texts)

# DataFrame integration
df = pd.DataFrame({'headline': texts})
df = analyzer.analyze(df, text_column='headline')

Label Mapping

Label ID	Label Name	Description
0	negative	Bearish sentiment, negative outlook
1	neutral	Balanced reporting, factual statements
2	positive	Bullish sentiment, positive outlook

Model Versions

Version	Release Date	Training Data	Accuracy	Notes
v3	2026-03-26	AION Taxonomy (400K)	99.63%	Current release
v2	2026-03-26	External lexicon (823K)	98.55%	Deprecated - labeling issues
v1	2026-03-14	SHAM labels	N/A	Initial release

Related Models

aion-analytics/aion-sentiment-in-v2: Previous release (deprecated due to labeling issues)
cardiffnlp/twitter-roberta-base-sentiment: General sentiment model

Citation

@software{aion_sentiment_v3_2026,
  author = {AION Open Source Contributors},
  title = {AION-Sentiment-IN-v3: Indian Financial News Sentiment Analysis (Taxonomy-Corrected)},
  version = {3.0.0},
  year = {2026},
  url = {https://github.com/AION-Analytics/market-sentiments}
}

Acknowledgements

AION Taxonomy: 136 events with known sentiment used for label creation
NRC Emotion Lexicon: Mohammad & Turney (2013) - Used for emotion analysis
HuggingFace Transformers: Wolf et al. (2020) - Model framework
AION Analytics: News corpus and infrastructure

Contact

GitHub: https://github.com/AION-Analytics/market-sentiments
Issues: https://github.com/AION-Analytics/market-sentiments/issues
Email: contributors@aion.opensource

Changelog

v3.0.0 (2026-03-26)

Built on AION Taxonomy labeling methodology
Trained on 400K headlines with AION taxonomy labels
Validation accuracy: 99.63%
Test headlines: 67% accuracy (4/6)
Fixed: "Markets Crashing" now predicts negative (was positive in v2)
Fixed: "TCS Record Earnings" now predicts positive (was neutral in v2)

Known Issues

Ambiguous headlines like "Stocks to buy" or "RBI hikes repo" may predict neutral
Working on improving these edge cases in future releases

Downloads last month: 37

Safetensors

Model size

67M params

Tensor type

F32