Derivatives Pricer

Stochastic process pricing engine with a Streamlit dashboard.

Highlights

• 6 stochastic process families (GBM, Heston, Bates, SABR, LocalVol, RBergomi).
• 12+ pricers covering analytic, lattice, finite difference, Monte Carlo, COS Fourier and MLMC.
• 4 calibrators (SABR, Heston via COS, SVI raw, Dupire local vol bootstrapper).
• Greek engines: full analytic vector, lattice, MC pathwise, MC LRM, CRN bumping.
• AOT JIT warmup, batched CRR, vectorised Black-Scholes surface, pluggable cache adapter.
• 9 tab Streamlit dashboard (pricing, surface, greeks, convergence, strategy, IV, process lab, calibration, risk).

Layout

models/        domain types, pricers, processes, calibration
utils/         plotting, IV inversion, greeks, sensitivities, synthetic quotes
services/     application service layer between dashboard and core
ui/           streamlit components and sidebar
app.py        streamlit entry point
controller.py legacy shim, DeprecationWarning on import
tests/        regression and parity smoke (pytest)

Install

pip install -e .[dev]

pyproject.toml pins python>=3.10 and gates yfinance, pytest, ruff, black, mypy behind extras.

Run

streamlit run app.py

Set DERIVATIVES_PRICING_NO_WARMUP=1 to skip the eager numba warmup at import; call models.pricers.prewarm() manually instead.

Test

pytest -q

The conftest exports DERIVATIVES_PRICING_NO_WARMUP=1 so the suite stays cold-start fast.

Architecture

• Domain: models.options ships slot-frozen value objects. models.processes holds stochastic process classes that self register through @autoregister. models.pricers exposes the pricer surface, a PricerRegistry with MRO walk, and a prewarm() opt-in.
• Service: services is the only layer that knows about Streamlit, but it does so through a runtime-adaptive cached() adapter so the same helpers run in plain Python, notebooks or tests with a transparent lru_cache fallback.
• Greeks: utils.greeks ships analytic, MC pathwise + LRM, lattice and CRN bumping calculators. GreekCalculatorRegistry dispatches via (pricer_class, method) keys and walks the pricer MRO.
• Calibration: models.calibration carries SVI, SABR, Heston (COS-driven) and Dupire bootstrapper. All return a CalibrationResult dataclass.

Performance

• JIT kernels (numba.njit) on every hot path, with cache=True and parallel=True where applicable.
• AOT warmup pass eagerly compiles every kernel at import; cached .nbi files keep subsequent boots fast.
• Vectorised Black-Scholes surface bypasses the per-cell Python loop (40x speedup at 25x25).
• cached() decorator fingerprints ndarrays through tobytes and dicts recursively so non-streamlit callers also hit the cache.
• slots=True on every frozen dataclass for smaller per-instance footprint.

Decoupling

• Streamlit imports live only in services._cache, ui and app.py.
• New processes register via @autoregister; new Greeks register via GreekCalculatorRegistry.register.
• Sidebar option type branching dispatches through _BUILDERS mapping.