DerivX — Fixed Income & Currencies Analytics (v2)

https://derivx.streamlit.app/

DerivX is a portfolio-ready analytics platform showcasing quantitative engineering for Fixed Income & Currencies roles. The project combines Financial Mathematics theory with a friendly UX and production-style data plumbing.

Recruiter Snapshot

• Real market data: 24 trading days of SPY chains ingested automatically via yfinance + Selenium keep-alive workflow.
• Quantified edge: Constrained SVI calibration averages 0.052 RMSE (median 0.040) and beats spline baselines across 32 expiries.
• Robust validation: Walk-forward lab posts 71.8% directional hit-rate with live regime segmentation and model benchmarking (SVI vs SABR vs local-vol).
• Deployment-ready: Streamlit UI + GitHub Actions + wake-up automation keep the demo live for hiring managers.

The platform highlights:

• Equity mode — vectorised Black–Scholes with dividend/carry, Greeks dashboard, 3D price surface, implied-vol and SVI calibration.
• FX mode — Black–Scholes with foreign/domestic rates, SABR calibration and smile diagnostics.
• Rates mode — OIS/IRS bootstrapping, Black cap/floor pricing, PV/DV01/PV01 buckets, and scenario shocks (parallel/twist).

The app is built in Streamlit with a production-grade package layout, unit tests, sample data, and CI. Use this repository to demonstrate competency across term structures, volatility modelling, and fixed income risk.

---

Quick start

# Clone & enter
git clone https://github.com/ZealousEar/derivx-fic-analytics.git
cd derivx-fic-analytics

# Optional: create virtualenv
python -m venv .venv && source .venv/bin/activate

# Install requirements and package
pip install -r requirements.txt

# Launch Streamlit app
streamlit run src/derivx/ui/streamlit_app.py

# Run unit tests
pytest tests/ -v

Prefer to skip setup? Launch the hosted demo at https://derivx.streamlit.app/.

The project targets Python 3.10+ and has been validated on macOS and Ubuntu. The Streamlit app uses caching, so the initial load calibrates surfaces once and subsequent interactions remain responsive.

---

Repository tour

.
├── README.md                  # Overview + metrics
├── requirements.txt           # Runtime dependencies (numpy<2, pandas<2.2, yfinance>=0.2.66)
├── src/derivx/                # Python package
│   ├── ui/streamlit_app.py    # Streamlit interface (Equity/FX/Rates + Analytics)
│   ├── data/                  # Market data ingestion + filters
│   ├── calibration/           # Production SVI toolkit
│   ├── validation/            # Walk-forward backtesting utilities
│   ├── models/                # SABR/local-vol benchmarking
│   └── surface/ + rates/ + core/ etc.
├── tools/
│   ├── download_spy_chains.py # CLI helper to pull Yahoo option chains
│   └── compute_metrics.py     # Resume metrics generator
├── tests/                     # Pytest suite (calibration, data, backtests)
└── data/
    ├── examples/              # Sample CSVs for the UI
    └── metrics/               # Gitignored Parquet + CSV outputs

---

What’s inside

• Market data ingestion: yfinance downloader + CLI utility (tools/download_spy_chains.py) with liquidity filters, moneyness tagging, and gitignored Parquet cache.
• Production calibration: constrained SVI engine with spline benchmarking, diagnostics, and headless metrics pipeline (tools/compute_metrics.py).
• Walk-forward + model lab: regime-aware validation, SVI/SABR/local-vol comparison utilities, and Streamlit Analytics tabs that surface the full pipeline.
• Fixed-income/rates stack: OIS/IRS curve building, cap/floor analytics, PV/DV01/PV01 sensitivities, and scenario shocks (parallel/twist).
• Deployment plumbing: Streamlit UI, GitHub Actions CI, Selenium wake-up workflow, and a README aimed at recruiters/hiring managers.

---

Tests

• Put–call parity is checked numerically on every load (residual ~1e-8)
• Round-trip price ↔ implied vol via vol/implied.py returns <1e-6 error
• SABR calibration on synthetic data recovers parameters within ±10%
• DV01/PV01 analytics match bump-and-revalue within 2 bps
• Eight targeted unit tests cover the core price/calibration logic

Run the suite locally:

pytest tests/ -v

---

Streamlit UI highlights

---

Packaging and deployment

The project is configured for:

• Local runs via streamlit run
• Docker deployments via the included Dockerfile
• Render via render.yaml (one-click deploy)
• Heroku-style deployments via the included Procfile and setup.sh
• Continuous integration (pytest) through .github/workflows/ci.yaml
• Streamlit wake-up automation via .github/workflows/wake-streamlit.yml

Deploy to Render

1. Fork this repository
2. Go to render.com and create a new Web Service
3. Connect your GitHub repo — Render auto-detects render.yaml
4. Click Deploy — the app will be live at https://derivx.onrender.com

Deploy with Docker

# Build and run locally
docker build -t derivx .
docker run -p 8501:8501 derivx

# Or deploy to any container platform (Railway, Fly.io, Google Cloud Run, etc.)

Deploy to Hugging Face Spaces

1. Create a new Space at huggingface.co/spaces
2. Select Streamlit as the SDK
3. Upload the src/ folder and requirements.txt
4. Set the app file to src/derivx/ui/streamlit_app.py

To package the library (optional):

python -m build

---

Market data + analytics lab

The new Analytics mode in streamlit_app.py exposes the end-to-end pipeline built for real option chains:

1. Market Data tab – upload a Parquet chain or click “Pull from Yahoo Finance”. The app runs liquidity filters, shows summary metrics, and lets you download a cleaned CSV.
2. Production Calibration tab – pick expiries, enforce moneyness bands/no-arb checks, and visualise fitted SVI smiles vs a cubic-spline baseline (with RMSE/diagnostics tables).
3. Walk-Forward Validation tab – either upload a folder of daily chains or fetch a short window (≤20 business days) from Yahoo Finance, reapply filters, and run walk_forward_test with regime-level summaries.
4. Model Comparison tab – load a single expiry slice and compare in-sample RMSE for SVI/SABR/local-vol with an overlayed smile plot.

Yahoo Finance caveats: the public options endpoint only returns current chains and occasionally responds with “Invalid Crumb”/empty payloads. Each download attempt now retries 3× and surfaces a per-date status table. When a window keeps failing, run the offline helper and upload the files instead:

# Download a few SPY sessions into data/market/manual
python tools/download_spy_chains.py \
    --ticker SPY \
    --start 2025-11-10 \
    --end 2025-11-14 \
    --outdir data/market/manual

The Parquet files are gitignored; keep your historical downloads locally and use the upload flow whenever Yahoo throttles the live API.

---

Key Metrics & Performance

Quantified results from production calibration and backtesting on 24 trading days of SPY option chains (768 expiry slices):

• SVI Calibration: Achieved mean RMSE of 0.0520 (median 0.0401) across 32 unique expiries, with constrained no-arbitrage diagnostics enforced on all slices.
• Walk-Forward Validation: Out-of-sample backtest achieved 71.8% directional hit rate with mean RMSE of 0.0558 across 416 predictions in mid-volatility regimes (10-day calibration window).
• Model Comparison: SVI outperformed SABR by 25.3% in mean RMSE (0.2892 vs 0.3869) across 24 representative smile calibrations, demonstrating superior fit to equity option market data.

Metrics computed via tools/compute_metrics.py; full results available in data/metrics/. Regenerate with:

python tools/compute_metrics.py --chain-dir data/metrics/chains --outdir data/metrics