kuant

GPU-accelerated quantitative research kernels. kernel × quant.

Vision

Standard quant libraries stop at technical indicators and portfolio math. kuant additionally exposes:

• Sparse Identification of Nonlinear Dynamics (SINDy) — null-testing tools for feature discovery, distilled from real production research
• Quantum-inspired regime discovery — HMM state machines, Bell- inequality-style aggregation tests, retrain-frequency (Zeno) scans
• Topological Data Analysis (TDA) — persistent homology, Mapper algorithm (planned)

Combined with GPU-batched primitives (Black-Scholes, rolling statistics, implied-vol solvers), kuant is a research-grade toolkit for signal discovery — not just an implementation of textbook indicators.

Status

Alpha. 955 tests across 5 shipped subpackages:

Subpackage	Kernels	Highlights
kuant.core	16	BS pricing (bscall/bsput) + full Gaussian family (normcdf/normpdf/normppf + log-tails) + Student-t (tcdf/tpdf/tppf + log-tails) + Generalized Pareto (gpdcdf/gpdpdf/gpdppf) + logsumexp
kuant.options	21	First-order Greeks (delta/gamma/vega/rho/theta/charm) + second-order (vanna/volga/speed/zomma/color) + payoffs + chain filters + Newton and bisection implied-vol solvers
kuant.stats	19	Rolling primitives with strict-window NaN, plus Hurst (R/S) and rolling Hurst
kuant.qm	5 + hmm/ghmm subpackages	HMM/GHMM inference (forward/backward/viterbi/posterior) + belltest, zenoscan, posteriorentropy, nocloningscan, decoherencescan
kuant.sindy	6	permtest, grangerscan, sindylasso, pinnscan, symbolicscan, accelerationscan

Each kernel has:

• Full API doc under docs/kernels/
• CPU fallback (numpy path — works on any machine)
• GPU path (cupy — same math, verified for parity)
• Cross-checked test suite (golden values, library reference, cross-kernel identities, machine-precision fd tolerances)

Install

# CPU-only
pip install kuant

# With GPU
pip install kuant[gpu]

# Some tools have optional heavy dependencies (scikit-learn for belltest,
# statsmodels for grangerscan). Install those separately when needed:
pip install scikit-learn statsmodels

Documentation

• docs/kernels/ — per-kernel API docs, one per kernel, organized into core/, options/, stats/, qm/, sindy/
• docs/design/ — cross-cutting design decisions
• docs/examples/ — worked examples

Start at docs/README.md.

Quick start

import numpy as np
from kuant.core import bsput, bsputdelta, normcdf
from kuant.stats import rollmean, rollstd, zscore, rollcorr
from kuant.options import impvol
from kuant.qm.hmm import forward, viterbi

# Vectorized Black-Scholes put pricing
strikes = np.linspace(80, 120, 41)
prices = bsput(S=100.0, K=strikes, T=1.0, r=0.05, sigma=0.20)

# Invert market prices to implied vol
sigma_iv = impvol(prices, S=100.0, K=strikes, T=1.0, r=0.05, is_call=False)

# Rolling z-score of returns
z = zscore(returns, window=252)

# HMM state decoding
states, log_prob = viterbi(observations, pi, A, B)

Repository layout

kuant/
├── core/         Mathematical primitives (BS family, normal CDF/PDF)
├── options/      Options analytics (impvol solver)
├── stats/        Rolling and windowed statistics (18 kernels)
├── qm/           Quantum-inspired tools (HMM, belltest, zenoscan)
├── sindy/        SINDy-adjacent null-testing tools (permtest, grangerscan)
├── portfolio/    P&L, drawdown, Sharpe (skeleton)
├── backtest/     Simulation engine (skeleton)
├── signals/      Signal computation (skeleton)
├── topology/     TDA (skeleton)
├── text/         Text parsing (skeleton)
├── data/         Bar aggregation (skeleton)
├── edgecases/    Edge case utilities (skeleton)
└── queueing/     Hardware throttle + coordination layer

docs/
├── kernels/      One doc per kernel (grouped by subpackage)
├── design/       Cross-cutting design docs
└── examples/     Worked examples

tests/            1:1 with kernel files; 887 tests total

Design principles

1. CPU-first, GPU-second — every kernel has a numpy fallback so development works on any machine
2. Batched by default — kernels operate on tensors, not scalars; single scalar case is a special call
3. Explicit edge cases — NaN, zero denominators, empty arrays, past-expiry options all handled in-kernel; callers don't need defensive wrappers
4. Composable primitives — each kernel does one thing; complex operations build up from atoms
5. Reproducible — kernels don't touch global state; same inputs guarantee same outputs
6. No underscores in the API surface — bsput, not bs_put; belltest, not bell_test. Improves typing flow

Contributing

Contribution guidelines coming as project matures.

License

Apache 2.0. See LICENSE.