YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
AION Market Sentiment
Developer toolkit for financial sentiment analysis
AI-powered sentiment intelligence for Indian financial markets.
98.55% accuracy | <100ms latency | 592 NSE tickers | 957K training samples
What This Is
A developer toolkit for building financial sentiment models:
- Reusable Components - Sentiment analysis, sector mapping, VIX adjustment
- Experimentation - Quick prototyping for quant research
- Clean APIs - Simple Python interfaces, no infra complexity
- India-Focused - NSE/BSE tickers, Indian financial news
What This Is NOT
This is NOT a production trading system. For production infrastructure (live ingestion, Redis streams, ClickHouse pipelines, execution engines), see our internal systems.
This toolkit is for:
- ML engineers building sentiment models
- Quant researchers prototyping strategies
- Developers experimenting with financial NLP
- Students learning financial sentiment analysis
This toolkit is NOT for:
- Direct production trading (use internal systems)
- Live tick data ingestion (use internal systems)
- Order execution (use internal systems)
- System governance/audit (use internal systems)
Quickstart (30 seconds)
# Install
pip install aion-sentiment aion-sectormap aion-volweight
# Run
python -c "
from aion_sentiment import AIONSentimentAnalyzer
analyzer = AIONSentimentAnalyzer()
print(analyzer.predict(['Market reaches all-time high']))
"
Output:
[{'label': 'positive', 'confidence': 0.9389}]
Example Output
Sentiment Analysis
Headline | Sentiment | Confidence | VIX-Adjusted
--------------------------------------|-----------|------------|-------------
Reliance reports record profits | positive | 93.8% | 93.8% (VIX=12)
Market crashes on recession fears | negative | 90.5% | 45.2% (VIX=28)
TCS wins major digital deal | positive | 88.8% | 88.8% (VIX=12)
Sector Sentiment Heatmap
Sector | Bullish | Neutral | Bearish | Net Sentiment
--------------------|---------|---------|---------|---------------
Banking | 65% | 25% | 10% | +55%
IT | 72% | 20% | 8% | +64%
Auto | 45% | 35% | 20% | +25%
FMCG | 58% | 30% | 12% | +46%
Metal | 30% | 40% | 30% | 0%
Historical Impact Analysis
Query: "Market crashes on recession fears"
Similar Historical News (last 30 days):
1. "Stock market tumbles on recession fears" → -2.5% (next day)
2. "Investors panic as banks collapse" → -3.8% (next day)
3. "Banking crisis spreads across Europe" → -2.9% (next day)
Average 1-day Impact: -3.07%
Architecture
┌─────────────────┐
│ Data Sources │
│ • News APIs │
│ • RSS Feeds │
│ • Social Media │
└────────┬────────┘
↓
┌─────────────────┐
│ Sentiment Engine│
│ • Transformer │
│ • NRC Emotions │
│ • 98.55% Acc │
└────────┬────────┘
↓
┌─────────────────┐
│ Sector Mapper │
│ • 592 Tickers │
│ • 44 Sectors │
│ • 340 Groups │
└────────┬────────┘
↓
┌─────────────────┐
│ VIX Adjustment │
│ • LOW <12 │
│ • NORMAL 12-15 │
│ • HIGH 16-25 │
│ • PANIC ≥25 │
└────────┬────────┘
↓
┌─────────────────┐
│ Signal Output │
│ • JSON API │
│ • DataFrame │
│ • WebSocket │
└─────────────────┘
Packages
| Package | Purpose | Install |
|---|---|---|
| aion-sentiment | Sentiment & emotion analysis | pip install aion-sentiment |
| aion-sectormap | NSE ticker → Sector mapping | pip install aion-sectormap |
| aion-volweight | VIX-based confidence adjustment | pip install aion-volweight |
| aion-newsimpact | Historical news impact analysis | pip install aion-newsimpact |
| aion-sentiment-in | Training pipeline | pip install aion-sentiment-in |
Models
AION-Sentiment-IN-v1
| Metric | Value |
|---|---|
| Accuracy | 98.55% |
| F1 Score | 98.65% |
| Training Data | 957K Indian financial news headlines |
| Inference Time | <100ms per headline |
| Model Size | 438 MB |
| Download | HuggingFace |
Use Cases
1. Real-Time Trading Signals
from aion_sentiment import AIONSentimentAnalyzer
analyzer = AIONSentimentAnalyzer()
# Analyze breaking news
news = "RBI announces surprise rate cut"
result = analyzer.predict(news)
if result[0]['label'] == 'positive' and result[0]['confidence'] > 0.9:
print("BUY Signal: Banking sector")
2. Sector Rotation Strategy
from aion_sentiment import AIONSentimentAnalyzer
from aion_sectormap import SectorMapper
mapper = SectorMapper()
analyzer = AIONSentimentAnalyzer()
# Get sentiment by sector
sector_sentiment = {}
for sector in mapper.get_all_sectors():
tickers = mapper.get_tickers_in_sector(sector)[:10]
# Analyze news for each ticker
# Aggregate sector sentiment
sector_sentiment[sector] = avg_sentiment
# Rotate to highest sentiment sectors
top_sectors = sorted(sector_sentiment.items(), key=lambda x: x[1], reverse=True)
3. Risk Management
from aion_volweight import get_regime, weight_confidence
# Check VIX regime
regime = get_regime(vix=28) # Returns "PANIC"
# Adjust position sizing based on sentiment confidence
if regime == "PANIC":
position_size = 0.5 # Reduce by 50%
elif regime == "HIGH":
position_size = 0.8 # Reduce by 20%
else:
position_size = 1.0 # Full size
Edge Cases & Confidence Scoring
Edge Case Handling
| Scenario | Handling | Output |
|---|---|---|
| Empty text | Return neutral with 0% confidence | {'label': 'neutral', 'confidence': 0.0} |
| Very short text (<5 chars) | Return neutral with low confidence | {'label': 'neutral', 'confidence': 0.3} |
| Ambiguous news | Low confidence score (<0.6) | {'label': 'neutral', 'confidence': 0.55} |
| Conflicting signals | Confidence reduced proportionally | {'label': 'positive', 'confidence': 0.65} |
| Unknown ticker | Sector mapping returns 'Unknown' | Sector: 'Unknown' |
| High VIX (>25) | Confidence discounted 50% | adjusted_conf = conf * 0.5 |
Confidence Scoring System
def calculate_system_confidence(sentiment_confidence, vix_value, sector_weight, source_reliability):
"""
Calculate final system confidence score for trading decisions.
Args:
sentiment_confidence: Raw model confidence (0-1)
vix_value: Current India VIX value
sector_weight: Sector-specific weight (0.8-1.2)
source_reliability: News source reliability score (0.5-1.0)
Returns:
Final system confidence (0-1)
"""
# VIX regime adjustment
if vix_value >= 25:
vix_adjustment = 0.5 # PANIC: 50% discount
elif vix_value >= 16:
vix_adjustment = 0.8 # HIGH: 20% discount
elif vix_value >= 12:
vix_adjustment = 1.0 # NORMAL: no adjustment
else:
vix_adjustment = 1.0 # LOW: no adjustment
# Calculate final confidence
system_confidence = (
sentiment_confidence *
vix_adjustment *
sector_weight *
source_reliability
)
return min(1.0, max(0.0, system_confidence))
# Example usage
confidence = calculate_system_confidence(
sentiment_confidence=0.92,
vix_value=18, # HIGH regime
sector_weight=1.0, # Neutral sector weight
source_reliability=0.95 # High reliability source
)
# Result: 0.92 * 0.8 * 1.0 * 0.95 = 0.699 (69.9% system confidence)
Sector-Specific Confidence Weights
| Sector | Weight | Rationale |
|---|---|---|
| Banking | 1.1 | High news coverage, reliable signals |
| IT | 1.0 | Standard weight |
| FMCG | 1.0 | Standard weight |
| Auto | 0.9 | Moderate volatility |
| Metal | 0.8 | High volatility, noisy signals |
| Realty | 0.8 | Low liquidity, noisy signals |
| Unknown | 0.7 | Unmapped tickers |
Multi-Source Aggregation
def aggregate_sentiment(signals):
"""
Aggregate sentiment from multiple news sources.
Args:
signals: List of dicts with 'sentiment', 'confidence', 'source'
Returns:
Aggregated sentiment and confidence
"""
if not signals:
return {'label': 'neutral', 'confidence': 0.0}
# Weight by source reliability
source_weights = {
'reuters': 1.0,
'bloomberg': 1.0,
'economictimes': 0.9,
'moneycontrol': 0.85,
'twitter': 0.5,
}
weighted_sentiment = 0.0
total_weight = 0.0
for signal in signals:
weight = source_weights.get(signal['source'].lower(), 0.7)
sentiment_score = 1 if signal['label'] == 'positive' else (-1 if signal['label'] == 'negative' else 0)
weighted_sentiment += sentiment_score * signal['confidence'] * weight
total_weight += weight
avg_sentiment = weighted_sentiment / total_weight if total_weight > 0 else 0
# Convert back to label
if avg_sentiment > 0.3:
label = 'positive'
elif avg_sentiment < -0.3:
label = 'negative'
else:
label = 'neutral'
return {
'label': label,
'confidence': abs(avg_sentiment),
'sources_aggregated': len(signals)
}
# Example usage
signals = [
{'label': 'positive', 'confidence': 0.92, 'source': 'reuters'},
{'label': 'positive', 'confidence': 0.88, 'source': 'economictimes'},
{'label': 'neutral', 'confidence': 0.75, 'source': 'twitter'},
]
result = aggregate_sentiment(signals)
# Result: {'label': 'positive', 'confidence': 0.82, 'sources_aggregated': 3}
Confidence Thresholds for Trading Signals
| System Confidence | Signal Strength | Action |
|---|---|---|
| ≥ 0.85 | Very Strong | Full position size |
| 0.70 - 0.84 | Strong | 80% position size |
| 0.55 - 0.69 | Moderate | 50% position size |
| 0.40 - 0.54 | Weak | 25% position size |
| < 0.40 | Very Weak | No action (skip) |
Benchmarks
Tested on: Apple M4 Mac, 16GB RAM | Dataset: 957K Indian financial news headlines
Model Performance
| Metric | Score |
|---|---|
| Accuracy | 98.55% |
| F1 Score (macro) | 98.65% |
| Precision (macro) | 98.70% |
| Recall (macro) | 98.60% |
Latency & Throughput
| Task | Latency | Throughput |
|---|---|---|
| Single headline | <50ms | - |
| Batch (100) | <200ms | 500/sec |
| Sector mapping | <10ms | 10,000/sec |
| VIX adjustment | <5ms | 50,000/sec |
Model Evaluation
| Metric | Score |
|---|---|
| Accuracy | 98.55% |
| F1 Score (macro) | 98.65% |
| Precision (macro) | 98.70% |
| Recall (macro) | 98.60% |
| Training Samples | 957K headlines |
| Validation Samples | 2K headlines |
Dataset: 957K Indian financial news headlines (Oct 2025 - Feb 2026) Classification source: UNIFIED_ROUTER_V4 (99.99% confidence)
Installation
Full Suite
pip install aion-sentiment aion-sectormap aion-volweight aion-newsimpact
Individual Packages
pip install aion-sentiment # Core sentiment analysis
pip install aion-sectormap # Sector mapping (592 tickers)
pip install aion-volweight # VIX adjustment
pip install aion-newsimpact # Historical impact
Development
git clone https://github.com/AION-Analytics/market-sentiments.git
cd market-sentiments
pip install -e ".[dev]"
pytest
Data Coverage
| Asset | Count | Description |
|---|---|---|
| NSE Companies | 592 | Mapped to sectors |
| Sectors | 44 | NSE classification |
| Business Groups | 340 | Tata, Birla, Ambani, etc. |
| Training News | 957K | Indian financial news |
| Emotion Lexicon | 14,182 | NRC emotions |
Performance Benchmarks
| Task | Latency | Throughput |
|---|---|---|
| Single headline | <50ms | - |
| Batch (100) | <200ms | 500/sec |
| Sector mapping | <10ms | 10,000/sec |
| VIX adjustment | <5ms | 50,000/sec |
Tested on: Apple M4 Mac, 16GB RAM
License
Apache License 2.0
Attribution:
This project uses AION Analytics open-source packages.
Visit https://github.com/AION-Analytics for more information.
Contact
- Email: aionlabs@tutamail.com
- GitHub: https://github.com/AION-Analytics
- HuggingFace: https://huggingface.co/AION-Analytics/aion-sentiment-in-v1
Built for the Indian financial community
- Downloads last month
- 110
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support