itofin

Idiomatic Python bindings over the libitofin Rust core, a ground-up port of QuantLib for pricing, risk, and calibration. The heavy numerics run in Rust; Python drives them. The pricing snippets below reproduce the cached Rust test oracles exactly; the abbreviated calibration snippets converge to the same targets within the tests' calibration tolerance (the full matrices live in tests/).

The API mirrors QuantLib's ql/ layout: types live in submodules (itofin.time, itofin.instruments, itofin.models, ...), while Settings and ItofinError stay at the top level.

Install

pip install itofin

Python versions

itofin ships a single abi3 wheel that runs on CPython 3.10, 3.11, 3.12, and 3.13+. Python 3.13 is the primary, tested target; 3.10 through 3.12 are supported through the stable ABI.

Price a European option and its greeks

from itofin import Settings
from itofin.instruments import OptionType, VanillaOption
from itofin.processes import BlackScholesProcess
from itofin.time import Date, DayCounter

s = Settings()
s.set_evaluation_date(Date(15, 6, 2026))
dc = DayCounter.actual360()

process = BlackScholesProcess(60.0, 0.08, 0.0, 0.30, Date(15, 6, 2026), dc)
option = VanillaOption(OptionType.Call, 65.0, Date(15, 6, 2026) + 90, s)
option.set_engine(process)

print(f"NPV   {option.npv():.10f}")
print(f"delta {option.delta():.10f}")
print(f"vega  {option.vega():.10f}")
print(f"rho   {option.rho():.10f}")

NPV   2.1333684449
delta 0.3724827980
vega  11.3515440535
rho   5.0538998582

Price under Heston and calibrate to a flat-vol surface

Semi-analytic Heston pricing:

from itofin import Settings
from itofin.instruments import OptionType, VanillaOption
from itofin.models import HestonModel
from itofin.processes import HestonProcess
from itofin.time import Date, DayCounter

s = Settings()
s.set_evaluation_date(Date(27, 12, 2004))
dc = DayCounter.actual_actual_isda()

# HestonProcess(risk_free, dividend, spot, v0, kappa, theta, sigma, rho, reference_date, day_counter)
process = HestonProcess(
    0.0225, 0.02, 1.0, 0.1, 3.16, 0.09, 0.4, -0.2, Date(27, 12, 2004), dc
)
model = HestonModel(process)
option = VanillaOption(OptionType.Call, 1.05, Date(28, 3, 2005), s)
option.set_heston_engine(model, 64)

print(f"Heston NPV {option.npv():.10f}")

Heston NPV 0.0404774515

Calibrating the model to a flat 10% vol surface drives the vol-of-vol sigma toward zero and both v0 and theta toward the flat variance (0.10 squared = 0.01). The snippet below fits a 3-maturity by 3-strike grid; the full 21-helper matrix and its tighter tolerances live in tests/test_heston_calibration.py.

import math

from itofin import Settings
from itofin.models import CalibrationErrorType, HestonModel, HestonModelHelper
from itofin.optimization import EndCriteria, LevenbergMarquardt
from itofin.processes import HestonProcess
from itofin.termstructures import FlatForward
from itofin.time import Calendar, Date, DayCounter, Period

s = Settings()
s.set_evaluation_date(Date(15, 1, 2026))
dc = DayCounter.actual360()
ref = Date(15, 1, 2026)
risk_free = FlatForward(ref, 0.04, dc)
dividend = FlatForward(ref, 0.50, dc)
VOL = 0.10

helpers = []
for n in (1, 2, 3):
    tau = float(n)
    forward = dividend.discount(tau) / risk_free.discount(tau)  # spot = 1.0
    for moneyness in (-1.0, 0.0, 1.0):
        strike = forward * math.exp(-moneyness * VOL * math.sqrt(tau))
        helpers.append(HestonModelHelper(
            Period(n, "Years"),
            Calendar.null_calendar(),
            1.0,        # spot
            strike,
            VOL,        # 10% flat vol
            0.04,       # risk-free
            0.50,       # dividend yield
            CalibrationErrorType.RelativePriceError,
            ref,
            dc,
            s,
        ))

process = HestonProcess(0.04, 0.50, 1.0, 0.01, 0.2, 0.02, 0.5, -0.75, ref, dc)
model = HestonModel(process)
method = LevenbergMarquardt(1e-8, 1e-8, 1e-8, False)
end_criteria = EndCriteria(400, 40, 1e-8, 1e-8, 1e-8)
model.calibrate(helpers, method, end_criteria, 96)

print(f"sigma {model.sigma():.6f}")
print(f"v0    {model.v0():.6f}")
print(f"theta {model.theta():.6f}")

sigma 0.000216
v0    0.010000
theta 0.010000

Calibrate Hull-White to a swaption matrix

from itofin import Settings
from itofin.indexes import Euribor
from itofin.models import CalibrationErrorType, HullWhite, SwaptionHelper
from itofin.optimization import EndCriteria, LevenbergMarquardt
from itofin.termstructures import FlatForward
from itofin.time import Date, DayCounter, Period

s = Settings()
s.set_evaluation_date(Date(15, 2, 2002))
curve = FlatForward(Date(19, 2, 2002), 0.04875825, DayCounter.actual365_fixed())

model = HullWhite(curve, 0.1, 0.01)
index = Euribor.six_months(curve, s)

# (option tenor, swap tenor, Black vol); the full co-terminal matrix is in the tests.
swaptions = [(1, 5, 0.1148), (2, 4, 0.1108), (3, 3, 0.1070), (4, 2, 0.1021), (5, 1, 0.1000)]
helpers = [
    SwaptionHelper(
        Period(maturity, "Years"),
        Period(length, "Years"),
        vol,
        index,
        Period(1, "Years"),
        DayCounter.thirty360_bond_basis(),
        DayCounter.actual360(),
        curve,
        CalibrationErrorType.RelativePriceError,
        1.0,
    )
    for maturity, length, vol in swaptions
]

method = LevenbergMarquardt(1e-8, 1e-8, 1e-8, False)
end_criteria = EndCriteria(10000, 100, 1e-6, 1e-8, 1e-8)
model.calibrate(helpers, method, end_criteria, False)  # fix_reversion=False, fit a and sigma

print(f"a     {model.a():.7f}")
print(f"sigma {model.sigma():.7f}")

a     0.0464113
sigma 0.0057992

Price a European swaption with the Jamshidian engine

from itofin import Settings
from itofin.indexes import Euribor
from itofin.instruments import (
    EuropeanExercise,
    SettlementMethod,
    SettlementType,
    Swaption,
    SwapType,
    VanillaSwap,
)
from itofin.models import HullWhite
from itofin.termstructures import FlatForward
from itofin.time import BusinessDayConvention, Calendar, Date, DayCounter, Frequency, Schedule

s = Settings()
s.set_evaluation_date(Date(15, 1, 2026))
curve = FlatForward(Date(15, 1, 2026), 0.03, DayCounter.actual365_fixed())

fixed = Schedule(
    Date(15, 1, 2028), Date(15, 1, 2033),
    Frequency.Annual, Calendar.target(),
    BusinessDayConvention.Unadjusted,
)
floating = Schedule(
    Date(15, 1, 2028), Date(15, 1, 2033),
    Frequency.Semiannual, Calendar.target(),
    BusinessDayConvention.Unadjusted,
)
index = Euribor.six_months(curve, s)

swap = VanillaSwap(
    SwapType.Payer, 100.0,
    fixed, 0.03, DayCounter.thirty360_bond_basis(),
    floating, index, 0.0, DayCounter.actual360(), s,
)
swaption = Swaption(
    swap, EuropeanExercise(Date(15, 1, 2027)),
    SettlementType.Physical, SettlementMethod.PhysicalOTC, s,
)
swaption.set_jamshidian_engine(HullWhite(curve, 0.05, 0.01))

print(f"payer swaption NPV {swaption.npv():.10f}")

payer swaption NPV 1.5666103956

Draw random numbers as NumPy arrays

The generators under itofin.randomnumbers carry QuantLib's Python names. Sequence generators return NumPy arrays, and next_sequences(count) draws a whole (count, dimension) matrix in one call, so a path loop never crosses the binding once per path.

from itofin.randomnumbers import (
    GaussianRandomSequenceGenerator,
    UniformRandomGenerator,
    UniformRandomSequenceGenerator,
)

rng = UniformRandomGenerator(42)                       # MT19937, seeded
usg = UniformRandomSequenceGenerator(3, rng)           # 3 uniforms per sequence
gsg = GaussianRandomSequenceGenerator(usg)             # inverse-cumulative normals
print(usg.next_sequence())                             # ndarray, shape (3,)
print(gsg.next_sequences(1000).shape)                  # (1000, 3) standard normals

SobolRsg, HaltonRsg and GaussianLowDiscrepancySequenceGenerator provide the low-discrepancy counterparts with the same next_sequence / next_sequences calls.

Global bootstrap callbacks

PiecewiseYieldCurve(..., bootstrap="global") accepts keyword-only additional_penalties(times, data), additional_dates(), and additional_variables=SimpleQuoteVariables(quotes, guesses, lower_bounds). Penalties return finite residuals with a constant length during each calculation; times and data are copied lists including the reference node. Extra helpers can be repriced through helper.quote_error() inside a penalty. Additional dates and variables add unknowns and need corresponding residuals. Iterative bootstrap rejects these options.

For a futures convexity quote being fitted, pass register_conv_adj=False to its FuturesRateHelper constructor. This prevents optimizer updates from recursively invalidating the curve. The default remains observable for ordinary market quotes. See the joint futures-convexity example, checked against an independently compiled QuantLib oracle.

Callbacks run synchronously on the owning thread. Python exceptions become ItofinError from fallible pricing/bootstrap queries, including the original exception type and message. Failed queries can be retried after fixing the callback. max_date() retains its existing fallback on bootstrap failure; it neither proves a successful solve nor hides the error from a subsequent pricing query. Do not mutate quotes inside callbacks; such mutations are rejected. Callback state changes alone do not invalidate a successfully cached curve; observed market-input changes do.

Native consumers retain callbacks after the Python curve wrapper is dropped. Avoid callbacks that strongly capture the curve or an engine/model/index that retains it. Those mixed native/Python cycles need the user to break the capture; use weakref.ref(curve) for trial-curve queries. Capturing an additional helper is supported. Full mixed-graph garbage collection is tracked in #1028.

Typing checks

See typing acceptance for the pinned, bidirectional Pyright gate.

License

BSD-3-Clause, matching the libitofin core. See the repository for the layer status table, the divergences-from-QuantLib catalogue, and the Rust test oracles behind the numbers above.