CRESTA

Your Wealth, Powered by Intelligence

Production-grade AI Robo-Advisory for Indian Equity Markets

Live Demo → · Author →

---

Screenshots

Dark Mode Landing — live SENSEX/NIFTY ticker, 3D COBE globe, breathing emerald gradient, floating market data.	Light Mode Landing — same data, zero compromise on readability. Globe adapts palette automatically.
Intelligent Features — AI Risk Profiling, Real-time Market Data, Automated Rebalancing cards.	Login Page — Warren Buffett quote, Google OAuth, JWT authentication, sign up flow.
Email Verification — token-based verification for all new signups, 24-hour expiry.	Portfolio Dashboard — total invested, current value, P&L, asset allocation donut, live ticker tape, AI alerts.
Holdings + AI Advisor — sparkline trends, BUY/SELL/HOLD signals, AI Score breakdown (Sentiment/Risk Fit/Valuation), personalized news.	Market Watch — stock search, 30-day historical + 7-day AI forecast, dynamic red/green coloring, NIFTY/SENSEX/BANK NIFTY indices.
Personalized News Feed — top gainers/losers, Reuters/Bloomberg headlines filtered by portfolio holdings.	Settings — profile management, dark/light mode toggle, language selection (English, Hindi, Gujarati, Punjabi).
Full Hindi Mode — entire dashboard rendered in Hindi including nav, tooltips, AI reasoning, holdings table, and alerts. Gujarati and Punjabi also supported.

---

Why Cresta?

Zerodha and Groww give you a brokerage. Smallcase gives you curated baskets. Cresta gives you a reasoning engine.

	Zerodha / Groww	Smallcase	Wealthfront	Cresta
AI Stock Scoring	✗	✗	Partial	✅ Ensemble ML (LSTM + XGBoost + ARIMA + FinBERT)
Explainable Signals	✗	✗	✗	✅ Sentiment / Risk Fit / Valuation breakdown
Indian Language Support	✗	✗	✗	✅ Hindi, Gujarati & Punjabi
7-day Price Forecast	✗	✗	✗	✅ Walk-forward validated ensemble
Email Watchlist Alerts	✓	✗	✓	✅ Trigger-based, configurable
Risk Profiling (ML)	✗	✗	✓	✅ XGBoost + NFCS 2021 real survey data
Email Verification	✗	✗	✓	✅ Token-based, 24-hour expiry

Built for the 200M+ Indians who invest without institutional-grade tooling.

---

📊 ML Evaluation Metrics

1. 🧠 Intelligent Risk Profiling

Cresta dynamically classifies users as Conservative, Moderate, or Aggressive based on Age, Income, Investment Goals, and Risk Tolerance.

• Model: XGBoost Classifier
• Dataset: 25,000 profiles — 2,578 real NFCS 2021 Investor Survey respondents (FINRA Foundation) augmented with 22,422 synthetic profiles generated via SEBI income capacity guideline distributions and empirical behavioral noise.
• Accuracy: 68% (Validated on heterogeneous synthetic noise)
• Aggressive Recall: 97% — solved the high-bias "Moderate" issue

Explainable AI (XAI) — Feature Importance:

Feature	Importance
Risk Tolerance	61.35%
Income	17.81%
Investment Goal	14.75%
Age	3.88%
Experience	2.21%

---

2. 📈 Tier-2 Quant Stock Forecasting

• Architecture: Attention-LSTM Hybrid + XGBoost + ARIMA Ensemble
• Attention Mechanism: Applies learned temporal weights across LSTM hidden states, allowing the model to selectively focus on the most predictive time steps rather than treating all historical observations equally.
• Ensemble Weights: (0.70 LSTM / 0.10 XGBoost / 0.20 ARIMA) selected via validation MAPE minimization across walk-forward folds — LSTM dominates long-horizon trend capture while ARIMA stabilizes short-term variance.
• Features (16): Close, Volume, SMA (5, 20), RSI (14), MACD, Bollinger Bands, OBV, FinBERT Sentiment (daily NSE-listed company headlines via yfinance news API, aggregated as mean sentiment score per ticker across all articles published within 24 hours), USD/INR Exchange Rate, India VIX, Crude Oil Futures
• Validation: Strict Walk-Forward Validation (3-fold expanding window, minimum 45-day folds) — no look-ahead bias
• Seed: Fixed at 42 for full reproducibility
• Dataset: 20 years of historical Nifty50 daily data (via Kaggle & yfinance)

Forecasting Performance (Walk-Forward Validated):

Ticker	Sector	Avg Ensemble MAPE	Status
RELIANCE.NS	Energy/Conglomerate	1.33%	✅ Verified
TCS.NS	IT Services	3.85%	✅ Verified
INFY.NS	IT Services	2.69%	✅ Verified
HDFCBANK.NS	Banking	2.67%	✅ Verified
ICICIBANK.NS	Banking	2.51%	✅ Verified
SUNPHARMA.NS	Pharma	0.82%	✅ Verified
MARUTI.NS	Auto	7.66%	✅ Verified
ONGC.NS	Energy	4.30%	✅ Verified

Average Ensemble MAPE across test set: 3.23% (Dramatic improvement from ~11-18%)

---

ML Pipeline

Raw Market Data (OHLCV + News Headlines)
         │
         ▼
┌─────────────────────────────────────────────┐
│              Feature Engineering             │
│  RSI · MACD · Bollinger · SMA · OBV ·       │
│  India VIX · USD/INR · Crude · FinBERT      │
└───────────────────┬─────────────────────────┘
                    │  16 features
         ┌──────────┴──────────┐
         ▼                     ▼
  ┌─────────────┐       ┌─────────────┐
  │ Attention-  │       │  XGBoost    │
  │   LSTM      │       │  Regressor  │
  │ (temporal)  │       │ (tabular)   │
  └──────┬──────┘       └──────┬──────┘
         │                     │
         ▼                     ▼
  ┌─────────────┐       ┌─────────────┐
  │    ARIMA    │       │   FinBERT   │
  │ (baseline)  │       │  (NLP sent) │
  └──────┬──────┘       └──────┬──────┘
         └─────────┬───────────┘
                   ▼
          ┌────────────────┐
          │    Ensemble    │
          │  0.70/0.10/0.20│
          └───────┬────────┘
                  ▼
     AI Score + BUY / SELL / HOLD
          7-day Price Forecast

---

🏗️ System Architecture

graph TD
    Client[React + Vite Frontend] <-->|HTTPS / REST API| Nginx[Nginx Reverse Proxy]
    Nginx <-->|Gunicorn| Django[Django 5 Backend]
    Django <-->|Task Queue| Redis[Redis Broker]
    Redis <--> Celery[Celery + Celery Beat]
    Celery <-->|Train/Predict| PyTorch[Attention-LSTM Engine]
    Celery <-->|Analyze News| FinBERT[HuggingFace NLP]
    Django <-->|Classify Risk| XGBoost[XGBoost Profiler]
    Django <--> Postgres[(PostgreSQL)]
    PyTorch <--> YF[yfinance API]

Loading

🗄️ Database Schema

erDiagram
    USER ||--o{ PORTFOLIO : owns
    USER {
        int id PK
        string email
        string password_hash
    }
    INVESTOR_PROFILE ||--|| USER : belongs_to
    INVESTOR_PROFILE {
        int id PK
        int age
        decimal income
        int risk_tolerance
        string user_class "Aggressive/Moderate/Conservative"
        boolean email_verified
    }
    PORTFOLIO ||--o{ HOLDING : contains
    PORTFOLIO {
        int id PK
        decimal total_value
        decimal cash_balance
    }
    HOLDING ||--o{ ALERT : triggers
    HOLDING {
        int id PK
        string ticker
        decimal quantity
        decimal average_buy_price
    }
    STOCK_PREDICTION {
        string ticker PK
        json history_array
        json future_forecast_array
        timestamp last_updated
    }
    ALERT {
        int id PK
        string ticker
        string signal "Buy/Sell/Hold"
        boolean is_active
    }
    PAPER_TRADE {
        int id PK
        string ticker
        string action "BUY/SELL"
        decimal quantity
        decimal price_at_trade
        timestamp created_at
    }
    WATCHLIST_ALERT {
        int id PK
        string ticker
        decimal target_price
        string condition "ABOVE/BELOW"
        boolean triggered
    }

Loading

---

✨ Core Features

🌍 Interactive 3D Global Exchange Globe

• 7 Global Exchanges: BSE SENSEX, NSE NIFTY 50, FTSE 100, NYSE/DOW, IBOVESPA, NIKKEI 225, ASX 200
• Geographic Sync: Angular-distance phi detection ensures correct exchange card appears when that continent faces the camera
• Live Data: BSE SENSEX and NSE NIFTY 50 prices fetched live from backend API
• Theme-Aware: Globe adapts palette for Light and Dark mode via ThemeContext

🎨 Premium Emerald Design System

• Dark Mode: Deep charcoal #121212 · Light Mode: Cool off-white #f0f4f8
• Animated Background: Canvas-based breathing gradient with 3 independent radial emerald glows + 45 floating market data numbers (SENSEX values, ₹ prices, ▲▼ indicators)
• Complete cyan→emerald migration across all components, Tailwind config, and CSS variables

🤖 Explainable AI (XAI) & Fiduciary Scoring

• Sentiment (40 pts): FinBERT NLP on daily NSE-listed company headlines
• Risk Fit (40 pts): Stock Beta matched against ML-classified user risk profile
• Valuation (20 pts): Price positioning relative to 52-week high/low
• Natural language reasoning generated per stock for full fiduciary transparency

📈 Performance-Driven Data Visualization

• Dynamic green/red chart coloring across all charts based on price direction
• Growth Forecast: historical line colored by trend, AI forecast in blue with confidence shading
• Portfolio chart, holdings sparklines, market indices all follow same logic

🔐 Security & Authentication

• JWT with 15-minute access tokens and HttpOnly refresh rotation
• Email Verification: Token-based verification for all new signups (24-hour expiry, SMTP delivery)
• Google OAuth integration
• Stock ticker inputs validated against NSE/BSE suffix whitelist
• Computational DoS mitigation via PostgreSQL prediction cache
• Production: SECURE_SSL_REDIRECT, X-Frame-Options: DENY, strict HSTS

🌐 Deep Localization (i18n)

• react-i18next supporting English, Hindi, Gujarati, and Punjabi
• All UI components, ML reasoning strings, and alerts translatable

📊 Portfolio Management

• Real-time P&L tracking via yfinance
• Smart Buy/Sell/Hold alerts based on moving average crossovers
• Email watchlist price trigger alerts via SMTP
• Paper trading and watchlist alert models
• Live ticker tape: SENSEX, NIFTY, BANK NIFTY, NIFTY IT, NASDAQ, S&P 500, USD/INR, GOLD

---

🛠️ Technology Stack

Category	Technology	Purpose
Frontend	React 18, Vite, TailwindCSS, COBE	UI, routing, globe visualization
Backend	Django 5, DRF	API layer, ML serving, alert triggers
Database	PostgreSQL	Users, holdings, watchlist, alerts
Cache	Redis	ML results, live ticker data
Auth	JWT + Google OAuth	Authentication, email verification
ML	PyTorch, XGBoost, statsmodels, FinBERT	Ensemble advisory engine
MLOps	MLflow	CV loss curves, model versions, hyperparameters
DevOps	Docker, Nginx, Gunicorn	Containerization, production serving
i18n	react-i18next	English, Hindi, Gujarati, Punjabi

---

🚀 Quick Start

With Docker (Recommended)

git clone https://github.com/Cypher-redeye/Cresta.git
cd Cresta
docker-compose up --build

• Frontend: http://localhost:5173
• Backend API: http://localhost:8000/api/

Without Docker

# Backend
cd backend
python -m venv venv
source venv/bin/activate   # Windows: venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env       # Fill in values
python manage.py migrate
python manage.py runserver

# Frontend
cd frontend
npm install
npm run dev

---

Environment Variables

Backend (backend/.env)

Variable	Description
SECRET_KEY	Django secret key
DEBUG	False in production
DATABASE_URL	PostgreSQL connection string
REDIS_URL	Redis connection URL
ALLOWED_HOSTS	Comma-separated allowed hosts
CORS_ALLOWED_ORIGINS	Frontend URL(s)
EMAIL_HOST_USER	SMTP sender address
EMAIL_HOST_PASSWORD	SMTP app password
FRONTEND_URL	Frontend URL for email verification links

---

🗺️ Roadmap

• Oracle Cloud Deployment — full ML stack on 4 OCPU / 24GB RAM free tier
• Portfolio Backtesting — simulate historical strategy vs Nifty50 benchmark with Sharpe ratio, drawdown, CAGR
• Options Chain Analyzer — IV and Greeks visualization for F&O traders
• Mutual Fund Coverage — extend AI scoring to top 50 Indian MFs
• Mobile App — React Native port with push-based watchlist alerts
• SEBI Compliance Pipeline — auditable reasoning logs for RIA regulations

---

📚 Academic References

• Hochreiter, S., & Schmidhuber, J. (1997). Long Short-Term Memory. Neural Computation, 9(8), 1735–1780.
• Vaswani, A., et al. (2017). Attention Is All You Need. NeurIPS.
• Araci, D. (2019). FinBERT: Financial Sentiment Analysis with Pre-trained Language Models. arXiv:1908.10063.
• Chen, T., & Guestrin, C. (2016). XGBoost: A Scalable Tree Boosting System. KDD '16.
• Markowitz, H. (1952). Portfolio Selection. Journal of Finance, 7(1), 77–91.

Contributors

Ankit Mishra	Shivam Panchal	Shubham Kumar Jha	Om Sharma

---

Author

Om Sharma AI Engineering Student · Parul University · Batch 2023–2027

Built end-to-end as a final year capstone project. Every architectural decision is production-intent.

---

Cresta is a production-ready, highly localized Robo-Advisory platform demonstrating the viable intersection of behavioral finance, deep learning, and modern web architecture.