stochastic-rs

Quantitative finance in Rust — a high-performance library for stochastic process simulation, option pricing, model calibration, volatility surfaces, fixed income, risk, statistics, copulas, and neural-network volatility surrogates. Generic over f32 / f64, with SIMD acceleration on CPU and CUDA / Metal / Accelerate / cubecl backends where they pay off, and first-class Python bindings via PyO3.

Documentation

📖 stochastic.rust-dd.com — full docs site (Fumadocs + Next.js, deployed on Vercel).

Highlights:

• 120+ stochastic processes — diffusion, jump, fractional / rough, short-rate, HJM, LMM, fBM, Hawkes, Lévy. Generic-precision ProcessExt<T> impl, SIMD on CPU, optional CUDA / Metal for FGN / fBM.
• Pricing & calibration — closed-form (BSM, Bachelier, Black76, Bjerksund-Stensland, …), Fourier (Heston / Bates / Merton-jump / Kou / VG / CGMY / HKDE / double-Heston), Monte Carlo (basket, rainbow, cliquet, autocallable, spread), finite difference, Bermudan LSM, Heston SLV. Heston / SABR / SVJ / Lévy / rough Bergomi / double-Heston / Hull-White swaption-grid calibrators.
• Statistics & risk — Hurst (Fukasawa), MLE for 1-D diffusions with 6 transition densities, ADF / KPSS / Phillips-Perron, realised variance with BNHLS bandwidth, HMM, changepoint, particle filter, UKF. VaR / CVaR / drawdown, Sharpe / Sortino / IR / Calmar.
• Fixed income & credit — yield-curve bootstrapping, Nelson-Siegel / Svensson, multi-curve, IRS / inflation swaps, Vasicek / CIR / Hull-White / G2++ short-rate engines, Merton structural model, reduced-form survival curves, CDS pricing, JLT migration matrices.
• Microstructure — Almgren-Chriss, Kyle (1985), Bouchaud propagator, full price-time priority order book.
• Distributions & copulas — 18 SIMD distributions with closed-form pdf / cdf / cf / moments and Python bindings (29 distribution structs total). 15 bivariate (Clayton / Frank / Gumbel / BB1 / BB7 / Independence / AMH / FGM / Galambos / Gaussian / Hüsler-Reiss / Joe / Marshall-Olkin / Plackett / Student-t) and 8 multivariate (Gaussian / Student-t / nested Archimedean / C-vine / D-vine / R-vine / two Gaussian-collapsed tree / vine approximations) copulas.
• Python bindings — 234 entries (218 PyO3 classes + 16 functions) spanning every sub-crate except AI surrogates. Numpy-in / numpy-out.

Installation

Rust

[dependencies]
stochastic-rs = "3.0.0-rc.1"

use stochastic_rs::prelude::*;
use stochastic_rs::stochastic::diffusion::gbm::Gbm;
use stochastic_rs::quant::pricing::heston::HestonPricer;

For per-sub-crate (lean) builds, CUDA / Metal / cubecl / Accelerate feature flags, native CPU optimisation, and SIMD details, see the installation guide on the docs site.

Python

pip install stochastic-rs

Source build (requires the Rust toolchain):

pip install maturin
maturin develop --release --manifest-path stochastic-rs-py/Cargo.toml

Linear algebra is pure Rust (faer), so every wheel — Linux, macOS and Windows — ships the identical full surface with no system BLAS to install. See the Python bindings page for the parity table.

Quickstart

use stochastic_rs::prelude::*;
use stochastic_rs::simd_rng::Unseeded;
use stochastic_rs::stochastic::diffusion::ou::Ou;
use stochastic_rs::quant::pricing::heston::HestonPricer;

fn main() {
    // Mean-reverting Ornstein-Uhlenbeck path: Ou::new(theta, mu, sigma, n, x0, t, seed)
    let ou = Ou::<f64>::new(2.0, 0.0, 1.0, 1_000, Some(0.0), Some(1.0), Unseeded);
    let path = ou.sample();
    println!("OU path points: {}", path.len());

    // Heston (1993) European option, closed form. The model holds only its own
    // parameters; the pricing query is passed to the call, so one model can
    // price a whole strike/maturity grid.
    // HestonPricer::new args: v0, rho, kappa, theta, sigma, lambda
    let pricer = HestonPricer::new(0.04, -0.5, 2.0, 0.04, 0.3, Some(0.0));
    // price_call/price_put args: s, k, r, q, tau
    let call = pricer.price_call(100.0, 100.0, 0.03, 0.0, 1.0);
    let put = pricer.price_put(100.0, 100.0, 0.03, 0.0, 1.0);
    println!("call={call:.4}, put={put:.4}");
}

import stochastic_rs as srs

# Mean-reverting OU path
p = srs.Ou(theta=2.0, mu=0.0, sigma=1.0, n=1000, x0=0.0, t=1.0)
path = p.sample()                       # numpy.ndarray, shape (1000,)

# Heston European option
pricer = srs.HestonPricer(
    s=100, v0=0.04, k=100, r=0.03, kappa=2.0, theta=0.04, sigma=0.3,
    rho=-0.5, tau=1.0, q=0.0,
)
call, put = pricer.call_put()
print(f"call={call:.4f}, put={put:.4f}")

More end-to-end recipes (Heston calibration, fBM Hurst estimation, vol-surface from quotes, Python interop) live in the tutorials section.

Benchmarks

FGN — CPU vs CUDA native (f32, H = 0.7)

cargo bench --features cuda-native --bench fgn_cuda_native

Single path:

Batch:

CUDA wins for large n (≥ 16 k); CPU rayon dominates for medium n because of the GPU launch / transfer overhead.

Distribution sampling — Normal vs upstream rand_distr

Single-thread fill_slice, median of 7 runs (cargo bench --bench dist_multicore). Comparison column:

• rand_distr + SimdRng — rand_distr::Normal consuming our SimdRng (same uniform stream, only the Normal algorithm differs).
• rand_distr + rand::rng() — the out-of-box upstream pipeline.

Single-sample speedup vs prior release

Criterion dist.sample(rng) loop, vs the wide 1.3.0 baseline (cargo bench --bench distributions -- --baseline before):

Driven by SIMD u64→f64 / u32→f32 magic-number conversion in SimdRng (direct-write fill_uniform_f64 / fill_uniform_f32 APIs that skip the [f64; 8] return-by-value round-trip), fused Exp(λ) scaling inside fill_exp_scaled, and an 8-at-a-time main loop in fill_ziggurat so copy_from_slice inlines to stp stores instead of a memcpy call.

Opt-in: dual-stream RNG (dual-stream-rng feature)

[dependencies]
stochastic-rs = { version = "3.0.0-rc.1", features = ["dual-stream-rng"] }

Unlocks SimdRngDual (two parallel xoshiro engines) and SimdNormalDual (Ziggurat unrolled 2× over the dual streams). Measured against the single-stream SimdNormal::fill_slice on Apple Silicon (cargo bench --bench dual_stream_compare --features dual-stream-rng):

The win comes from hiding the 16 scalar kn / wn table-lookup latencies behind the second engine's xoshiro state update on a modern out-of-order core. Uniform fills are not bottlenecked on the engine so they see no speedup. Trade-off: SimdRngDual::from_seed does not reproduce SimdRng::from_seed's bit-exact sequence (statistical properties are identical and KS-validated).

Contributing

Contributions are welcome — bug reports, feature suggestions, or PRs. Open an issue or start a discussion on GitHub. Per-feature recipes (add-diffusion-process, adding-distribution, calibration-pattern, docs-writing, …) live under .claude/skills/.

License

MIT — see LICENSE.