AAdevloper/market-regime-classifier

🚀 Market Regime Classifier - XGBoost Model

This is a production-ready XGBoost model for predicting market regimes (RISK_ON/RISK_OFF) in Indian financial markets (NIFTY 50). Part of a complete MLOps pipeline with experiment tracking, data versioning, and automated deployment.

🎯 Model Description

Task: Binary Classification
Algorithm: XGBoost (Extreme Gradient Boosting)
Target: Market Regime Prediction (RISK_ON vs RISK_OFF)
Framework: MLflow for experiment tracking and model registry

What are Market Regimes?

• 🟢 RISK_ON: Favorable market conditions - lower volatility, bullish momentum, suitable for aggressive trading
• 🔴 RISK_OFF: Cautious conditions - higher volatility, defensive positioning advised

📊 Model Performance

ML Metrics

Finance Metrics (Backtesting)

Production Thresholds

• ✅ F1 Score ≥ 0.65
• ✅ Sharpe Ratio ≥ 0.5

🔧 Features Used

The model uses four key technical indicators:

1. India VIX - Volatility index (market fear/greed indicator)
2. RSI (14-day) - Relative Strength Index (momentum)
3. 50-Day MA - Short-term moving average (trend)
4. 200-Day MA - Long-term moving average (trend)

💻 Usage

Using with MLflow

import mlflow
import pandas as pd

# Load the model
model_uri = "models:/market_regime_classifier/Production"
model = mlflow.xgboost.load_model(model_uri)

# Prepare features
features = pd.DataFrame([{
    'india_vix': 15.5,
    'rsi_14': 55.3,
    'ma_50': 18500.25,
    'ma_200': 18200.75
}])

# Predict
prediction = model.predict(features)[0]
proba = model.predict_proba(features)[0]

regime = "RISK_ON" if prediction == 1 else "RISK_OFF"
confidence = proba[prediction]

print(f"Regime: {regime} (confidence: {confidence:.2%})")

Using via REST API

The model is deployed on Hugging Face Spaces with a FastAPI endpoint:

curl -X POST "https://AAdevloper-mlops-finance-pipeline.hf.space/predict_regime" \
  -H "Content-Type: application/json" \
  -d '{
    "india_vix": 15.5,
    "rsi_14": 55.3,
    "ma_50": 18500.25,
    "ma_200": 18200.75
  }'

🏗️ MLOps Pipeline

This model is part of a complete MLOps system:

Components

• ✅ MLflow: Experiment tracking and model registry
• ✅ DVC: Data version control
• ✅ GitHub Actions: Automated CI/CD pipeline
• ✅ FastAPI: REST API for model serving
• ✅ Docker: Containerized deployment
• ✅ Hugging Face Spaces: Cloud deployment

Training Pipeline

1. Data preprocessing and feature engineering
2. Train/test split with stratification
3. XGBoost training with hyperparameters
4. Model evaluation (ML + Finance metrics)
5. Backtesting on historical data
6. Model registration in MLflow
7. Promotion to production if thresholds met

Automated Retraining

• Scheduled weekly retraining (GitHub Actions)
• Automatic model promotion based on performance
• Version control for models and data

📈 Training Configuration

model:
  type: xgboost
  params:
    max_depth: 6
    learning_rate: 0.1
    n_estimators: 100
    objective: binary:logistic
    eval_metric: logloss

📊 Dataset

• Source: Synthetic market data for NIFTY 50
• Features: India VIX, RSI-14, MA-50, MA-200
• Target: Market regime (binary classification)
• Split: 80/20 train/test

🎓 Skills Demonstrated

• MLOps pipeline architecture
• Experiment tracking (MLflow)
• Model registry management
• Data versioning (DVC)
• CI/CD automation (GitHub Actions)
• Model serving (FastAPI)
• Financial metrics & backtesting
• Docker containerization
• Cloud deployment (Hugging Face)

📚 Project Links

• GitHub Repository: mlops-finance-pipeline
• Live API Demo: Hugging Face Space
• Documentation: Full README with setup instructions

🔄 Model Versioning

This model is version-controlled using MLflow Model Registry:

• Model Name: market_regime_classifier
• Version: 1
• Stage: Production
• Run ID: 43544e1ed8eb4c2a9268948b6795bdf5

🤝 Contributing

This is a portfolio project demonstrating MLOps best practices. Feel free to:

• Fork and experiment
• Submit issues or suggestions
• Use as reference for your own MLOps projects

📄 License

MIT License - free to use for learning and projects

---

Built with ❤️ to showcase MLOps skills in Finance

For questions or collaboration, visit the GitHub repository