This repository contains 8 trained V5 Transformer models with 126 optimized features for predicting next-day price movements of major Indian banking stocks.

🎯 Key Features: Professional-Grade Macro Integration

Feature Category	Count	Weight	Purpose
Technical Indicators	35	1x	Price action, momentum, volatility
FinBERT Sentiment	4	2x	News sentiment analysis
Advanced Signals	12	2x	Analyst ratings, risk, earnings
Nifty Bank Index	3	8x	Market correlation
USD/INR Forex	7	5x	FII sentiment, INR weakness
Total Features	126	-	Optimized to prevent overfitting

💡 Why USD/INR Integration Matters

Critical Macro Indicator for Indian Markets:

INR Weakening → FII selling pressure → Bearish market sentiment
INR Strengthening → FII buying interest → Bullish market sentiment
Current Signal (Jan 2026): ₹91.54, +0.72% weakness = STRONG BEARISH

This model captures what institutional traders watch: currency movements as a leading indicator of foreign capital flows.

📊 Latest Training Results (January 2026)

Stock	Directional Accuracy	Avg Confidence	High Conf Accuracy
HDFC Bank	100%	99.94%	100%
ICICI Bank	100%	93.80%	100%
Kotak Mahindra Bank	100%	99.86%	100%
Axis Bank	100%	75.65%	100%
State Bank of India	100%	97.29%	100%
Punjab National Bank	100%	46.02%	100%
Bank of Baroda	100%	84.70%	100%
Canara Bank	100%	75.38%	100%

Average: 100% directional accuracy, 84.08% confidence

🏗️ Model Architecture

Type: V5 Transformer with Multi-Task Learning
Features: 126 (35 technical + 4 FinBERT + 12 advanced + 3 Nifty Bank + 7 USD/INR + duplicates)
Lookback: 30 days (optimized for responsiveness)
Parameters: ~517,534 per model
Tasks: Direction classification (70%) + Magnitude regression (30%)

Feature Breakdown (126 Total):

Base Features (61):

Technical Indicators (35): Price, volume, moving averages, RSI, MACD, Bollinger Bands, ATR, ADX
FinBERT Sentiment (4): sentiment_polarity, sentiment_score, news_volume, earnings_event
Advanced Signals (12): technical_signal, analyst_rating, macro_signal, risk_score, leadership_signal, earnings_signal
Nifty Bank Index (3): 1d, 5d, 20d returns (market correlation)
USD/INR Forex (7): rate, 1d/5d/20d changes, momentum, volatility, INR weakness score

Weighted Duplicates (65):

Nifty Bank 8x: 21 duplicates (strong market correlation signal)
USD/INR 5x: 28 duplicates (critical FII sentiment indicator)
Sentiment 2x: 16 duplicates (FinBERT + Advanced combined)

🎨 USD/INR Forex Features (Critical Innovation)

7 Features Capturing FII Sentiment:

usd_inr_rate: Current exchange rate
usd_inr_change_1d/5d/20d: Multi-horizon rate changes
usd_inr_momentum: 5-day rolling momentum
usd_inr_volatility: 20-day rolling volatility
inr_weakness_score: Weighted composite (1d × 40% + 5d × 30% + momentum × 30%)

Why This Matters:

INR weakening (USD/INR ↑) → FII selling → Market downturn
INR strengthening (USD/INR ↓) → FII buying → Market rally
Real-time macro signal that professional traders watch

💡 What Makes This Different?

Previous Models

51 features (technical + sentiment + advanced signals)
Missed critical macro indicators
No FII sentiment integration

Current Models (126 Features)

Macro-aware: USD/INR forex as leading indicator
Market-correlated: Nifty Bank 8x weightage
Sentiment-enhanced: 2x weight on FinBERT + Advanced
Optimized: Reduced from 248 → 126 features to prevent overfitting
Result: 100% accuracy with 84.08% confidence (+15.13% vs previous)

📥 Usage

python

from huggingface_hub import hf_hub_download
import tensorflow as tf

# Download a specific model
model_path = hf_hub_download(
    repo_id="RohithKoripelli/indian-bank-stock-models-advanced",
    filename="HDFCBANK/best_model.keras"
)

# Load the model
model = tf.keras.models.load_model(model_path)

# Download scaler
scaler_path = hf_hub_download(
    repo_id="RohithKoripelli/indian-bank-stock-models-advanced",
    filename="HDFCBANK/scaler.pkl"
)

📊 Training Data

Date Range: January 2019 - January 2026
Records: ~1,544 per stock (after cleaning)
Features: 126 (35 technical + 4 FinBERT + 12 advanced + 3 Nifty + 7 USD/INR + duplicates)
News Articles: 963 articles analyzed
Forex Data: 1,837 days of USD/INR rates
Market Index: Nifty Bank daily returns
Training Split: 70% train, 15% validation, 15% test

Data Sources:

Stock Prices: Yahoo Finance (NSE)
News: Google News API (30-day lookback)
Sentiment: FinBERT (yiyanghkust/finbert-tone)
Advanced Signals: Custom NLP extraction from headlines
Forex: USD/INR rates from Yahoo Finance
Market Index: Nifty Bank (^NSEBANK)

🤖 Automation

These models are used in an automated GitHub Actions workflow that:

Collects latest stock data daily
Calculates technical indicators
Collects Nifty Bank index data
Collects USD/INR forex rates (NEW)
Fetches news articles via GNews API
Analyzes sentiment with FinBERT
Extracts advanced trading signals
Prepares 126 features with optimized weights
Downloads models from Hugging Face
Generates predictions with confidence scores
Deploys to Vercel

Scheduled: Daily at 10 PM IST (4:30 PM UTC)

🔬 Training Details

Epochs: 21-61 (early stopping with patience 20)
Batch Size: 32
Learning Rate: 0.0001 with ReduceLROnPlateau
Optimizer: Adam with gradient clipping
Loss: Binary cross-entropy (direction) + Huber (magnitude)
Training Time: ~7 minutes total for all 8 stocks
Trained: January 23, 2026

📜 License

MIT License - Free to use for research and educational purposes.

🔗 Links

GitHub Repository: Stock-Price-Prediction
Live Predictions: Vercel Deployment
Previous Models (FinBERT): indian-bank-stock-models-finbert
Original Models (VADER): indian-bank-stock-models

📝 Citation

If you use these models in your research, please cite:

bibtex

@misc{indian-bank-advanced-models-2026,
  author = {Koripelli, Rohith},
  title = {Indian Bank Stock Price Prediction Models with Advanced Signals},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/RohithKoripelli/indian-bank-stock-models-advanced}}
}

Last Updated: January 2026 Model Version: V5 Transformer with Advanced Signals Status: Production Ready ✓