https://github.com/adrenaline03/nifty-lens
Volatility intelligence and portfolio analytics for NIFTY 50 equities with PostgreSQL + XGBoost + Streamlit.
https://github.com/adrenaline03/nifty-lens
data-analytics finance machine-learning nifty50 portfolio-analysis postgresql python streamlit time-series xgboost
Last synced: 26 days ago
JSON representation
Volatility intelligence and portfolio analytics for NIFTY 50 equities with PostgreSQL + XGBoost + Streamlit.
- Host: GitHub
- URL: https://github.com/adrenaline03/nifty-lens
- Owner: adrenaline03
- Created: 2026-04-17T04:52:39.000Z (2 months ago)
- Default Branch: main
- Last Pushed: 2026-04-27T13:08:23.000Z (2 months ago)
- Last Synced: 2026-04-27T14:20:49.004Z (2 months ago)
- Topics: data-analytics, finance, machine-learning, nifty50, portfolio-analysis, postgresql, python, streamlit, time-series, xgboost
- Language: Jupyter Notebook
- Homepage: https://nifty-lens.streamlit.app
- Size: 1.17 MB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# 📈 Nifty-Lens
Volatility intelligence and portfolio analytics for NIFTY 50 equities.
🔗 **[Try it live →](https://nifty-lens.streamlit.app)**
An end-to-end analytics platform: PostgreSQL backend with materialized views and stored procedures, an XGBoost classifier predicting next-5-day volatility regimes, and a Streamlit dashboard wiring it all together. Five years of daily NIFTY 50 data, 67.8% model accuracy on a 3-class classification problem (33% random baseline), and 86% accuracy on the high-confidence subset.
---
## Dashboard
The app has five pages. **Market Overview** shows NIFTY 50 index history, a sector YTD heatmap, and top movers across configurable time periods. **Stock Deep-Dive** drills into a single ticker with a price chart overlaying 50/200-day moving averages, rolling volatility, and a drawdown underwater plot. **Portfolio Analyzer** lets you build custom portfolios with weight sliders and returns annualized return, volatility, Sharpe, max drawdown, and sector exposure (computed via PostgreSQL stored procedures). **Volatility Predictions** visualizes the XGBoost classifier — current predictions for all 50 stocks, confusion matrix, calibration analysis, and per-ticker prediction history. **Methodology** documents the design decisions and key findings.
---
## Architecture
```mermaid
flowchart LR
A[yfinance API] -->|daily OHLCV| B[(PostgreSQL
on Neon)]
B --> C[SQL Analytics Layer
views + procedures]
B --> D[Feature Engineering
pandas + ta library]
D --> E[XGBoost
Classifier]
E -->|predictions| B
C --> F[Streamlit Dashboard
Plotly charts]
B --> F
style A fill:#1f6feb,color:#fff
style B fill:#0969da,color:#fff
style C fill:#3fb950,color:#fff
style D fill:#bf8700,color:#fff
style E fill:#bf8700,color:#fff
style F fill:#8957e5,color:#fff
```
Built with PostgreSQL (Neon serverless), Python, pandas, the `ta` technical analysis library, scikit-learn, XGBoost, Streamlit, and Plotly. Data comes from yfinance and is split/dividend-adjusted for return calculations. The full pipeline runs end-to-end via `scripts/refresh_pipeline.py`.
---
## The most important finding
The single most impactful decision in this project was the choice of prediction horizon. The initial model targeted **next-day volatility** and achieved 39% accuracy — barely above the 33% random baseline. Single-day moves are dominated by idiosyncratic noise. Reframing the target to **next-5-day average volatility** moved accuracy to 67.8%, a 29-percentage-point jump from changing one line of code.
This aligns with the well-documented _volatility clustering_ phenomenon in finance: realized volatility is much more autocorrelated over 3-10 day windows than over single days. Hyperparameter tuning across seven configurations moved accuracy by less than 0.5 percentage points. **Target horizon mattered far more than model complexity.**
---
## Calibration matters more than headline accuracy
The classifier is well-calibrated: when it reports >70% confidence, it's right 86% of the time across 4,874 test predictions (40% of the test set). Predictions in the 60-70% confidence band hit 63%, 50-60% hit 53%, and below 50% the model is essentially uncertain (49.8% accuracy on those — close to random for the cases the model itself flags as low-confidence).
In production this means you can apply a confidence threshold and get genuinely high-quality predictions on a meaningful fraction of inputs, with honest uncertainty on the rest. That tiered behavior is arguably more valuable than the headline number.
---
## Local setup
Requires Python 3.11+ and a PostgreSQL database (this project uses Neon's free tier).
```bash
git clone https://github.com/adrenaline03/nifty-lens.git
cd nifty-lens
python -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env # then edit .env with your Postgres credentials
```
Then run the full pipeline:
```bash
python scripts/refresh_pipeline.py
streamlit run streamlit_app/app.py
```
Or run individual stages — see `scripts/refresh_pipeline.py` for the order. Skip flags `--skip-ingest` and `--skip-ml` are available for partial reruns.
---
## Disclaimer
This is a portfolio project. It is not investment advice. All metrics use historical data; past performance does not predict future results. Transaction costs, slippage, and taxes are not modeled.