StochVolModels (stochvolmodels)

stochvolmodels provides Fourier-transform pricing, Monte Carlo validation, and calibration of European options under stochastic-volatility models in Python.

It is a focused research and practitioner library, not a general derivatives platform: the stable workflows cover European vanilla and related variance analytics under Heston and the Karasinski-Sepp log-normal stochastic-volatility model.

Paper: Sepp, A. and Rakhmonov, P. (2023), Log-normal stochastic volatility model with quadratic drift, International Journal of Theoretical and Applied Finance, 26(8). See Citation for the full BibTeX list.

Documentation: stochvolmodels.readthedocs.io · offline quickstart · LogSV quickstart in Colab · JOSS paper draft

Statement of need

stochvolmodels is the reference implementation of the Karasinski-Sepp log-normal stochastic volatility model, maintained by one of the model's originators, with the Heston model implemented alongside as a benchmark. The design goal is a single generic interface for a stochastic volatility model — a closed-form moment generating function for Fourier-transform pricing on one side, Monte Carlo dynamics on the other — so that analytic prices, simulated prices, and calibrated implied volatilities are directly comparable model to model.

Researchers and quantitative practitioners need more than a standalone pricing formula when they evaluate a stochastic-volatility specification: market quotes must share explicit forward, discount, option-type, and maturity conventions; calibration must fail visibly when constraints are not satisfied; and analytic prices need an independent simulation route. General derivatives libraries provide much broader instrument infrastructure, while model collections provide many formulas. This package deliberately serves the narrower workflow around the quadratic-drift LogSV model, with Heston as a like-for-like benchmark and paper implementations tied to the same code.

The same analytics power the research: the repository's papers/ directory reproduces the computations and figures of five papers, from the quadratic-drift log-normal SV model (IJTAF) to cryptocurrency inverse options (Quantitative Finance), robust stochastic volatility modelling, impermanent-loss hedging in DeFi, and stochastic volatility for the factor HJM framework — see Supporting Illustrations.

When to use it — and when not

Use stochvolmodels for European vanilla pricing and implied-volatility analytics under stochastic volatility, for model calibration to option chains (a calibration example to Bitcoin options data is included), and for replicating the papers above.

It is not a general derivatives platform: no American or path-dependent payoffs, no local-volatility or term-structure models. Black-Scholes-Merton and absolute-normal Bachelier analytics are provided by the required vanilla-option-pricers package and re-exported from stochvolmodels; for strategy backtesting and reporting, use qis.

Installation

Install using

pip install stochvolmodels

Upgrade using

pip install --upgrade stochvolmodels

Clone using

git clone https://github.com/ArturSepp/StochVolModels.git

Reviewer verification

The first result is offline and needs no credentials, local YAML, qis, OCA, or data download:

python -m pip install stochvolmodels
python examples/getting_started/quickstart.py

From a source checkout, verify the complete public artifact and documentation path with:

python -m pip install -e ".[dev,docs]"
python -m pytest -m "not slow"
python -m pytest -m slow
python -m pytest --cov=stochvolmodels --cov-report=json
python scripts/check_coverage_scopes.py coverage.json
python -m sphinx -W --keep-going -b html docs docs/_build/html
python -m build
python scripts/check_wheel_contents.py dist/*.whl

The v2.2.0 quickstart prints two five-strike slices and deterministic reference values including vanilla_price=0.197331 and six_month_atm_price=0.275202. The first call normally takes seconds because Numba compiles the numerical kernels. See the full verification guide and contribution/support guide.

Core Dependencies

• python >= 3.10
• vanilla-option-pricers >= 2.0.0
• numba >= 0.60.0
• numpy >= 2.0
• scipy >= 1.12.0
• pandas >= 2.2.0
• matplotlib >= 3.8.0
• seaborn >= 0.13.0

Optional extras

Install an extra using

pip install stochvolmodels[research]

The library itself imports none of these: import stochvolmodels needs the core dependencies only. Pinned contributor lint and audit tools live in PEP 735 groups and are installed with uv; see the testing and coverage guide.

API stability

The names listed by stochvolmodels.__all__ are the stable high-level API. Historical package-root names remain available lazily for compatibility, but names not in __all__ should be treated as advanced interfaces.

The rough-LogSV Monte Carlo and Factor HJM implementations are experimental research surfaces. Their characterized pricing paths are tested, but deep imports under stochvolmodels.pricers.rough_logsv and stochvolmodels.pricers.factor_hjm may evolve between minor releases. The legacy Gaussian_interval quadrature path requires unsupported orthopy and quadpy packages and now raises a precise ImportError; the incomplete rough-Heston kernel raises NotImplementedError until a characterized Mittag-Leffler backend is provided.

Table of contents

1. Model Interface
   1. Adding a new model engine
   2. Log-normal stochastic volatility model
   3. Heston stochastic volatility model
2. Running log-normal SV pricer
   1. Computing model prices and vols
   2. Running model calibration to sample Bitcoin options data
   3. Comparison of model prices vs MC
   4. Analysis and figures for the paper
3. Running Heston SV pricer
4. Supporting Illustrations for Public Papers

Implemented Stochastic Volatility models

The package provides interfaces for a generic volatility model with the following features.

1. Interface for analytical pricing of vanilla options using Fourier transform with closed-form solution for moment generating function
2. Interface for Monte-Carlo simulations of model dynamics
3. Interface for visualization of model implied volatilities

The model interface is in src/stochvolmodels/pricers/model_pricer.py.

Adding a new model engine

ModelPricer separates what a model must supply from what every model inherits. To add a model engine, subclass ModelPricer, define the model's ModelParams dataclass, and implement three model-specific pieces:

1. price_chain — analytic pricing of an OptionChain, typically a wrapper over a Numba-compiled moment-generating-function transform;
2. model_mc_price_chain with simulate_vol_paths and simulate_terminal_values — Monte Carlo simulation of the model dynamics and chain pricing from the simulated paths;
3. calibrate_model_params_to_chain — constrained calibration to an option chain, reporting failures through CalibrationError.

Everything downstream is inherited and works unchanged for the new model: single-option and slice pricing (price_vanilla, price_slice), model implied vols from chain prices (compute_chain_prices_with_vols), Monte Carlo implied vols with confidence bounds (compute_mc_chain_implied_vols), and the visualisation layer (plot_model_ivols, plot_model_ivols_vs_bid_ask, plot_model_ivols_vs_mc). LogSVPricer and HestonPricer are the two stable reference implementations of this contract. Requiring both the analytic and the Monte Carlo route from each model is deliberate: it means every new engine can be cross-validated analytic-versus-simulation through the same interface, at the cost of implementing two pricing paths rather than one.

Log-normal stochastic volatility model

The analytics for Karasinski-Sepp log-normal stochastic volatility model is based on the paper

Log-normal Stochastic Volatility Model with Quadratic Drift by Artur Sepp and Parviz Rakhmonov

The dynamics of the log-normal stochastic volatility model:

$$dS_{t}=r(t)S_{t}dt+\sigma_{t}S_{t}dW^{(0)}_{t}$$

$$d\sigma_{t}=\left(\kappa_{1} + \kappa_{2}\sigma_{t} \right)(\theta - \sigma_{t})dt+ \beta \sigma_{t}dW^{(0)}{t} + \varepsilon \sigma{t} dW^{(1)}_{t}$$

$$dI_{t}=\sigma^{2}_{t}dt$$

where $r(t)$ is the deterministic risk-free rate; $W^{(0)}_{t}$ and $W^{(1)}_t$ are uncorrelated Brownian motions, $\beta\in\mathbb{R}$ is the volatility beta which measures the sensitivity of the volatility to changes in the spot price, and $\varepsilon>0$ is the volatility of residual volatility. We denote by $\vartheta^{2}$, $\vartheta^{2}=\beta^{2}+\varepsilon^{2}$, the total instantaneous variance of the volatility process.

Implementation of Lognormal SV model is contained in

src/stochvolmodels/pricers/logsv_pricer.py

Heston stochastic volatility model

The dynamics of Heston stochastic volatility model:

$$dS_{t}=r(t)S_{t}dt+\sqrt{V_{t}}S_{t}dW^{(S)}_{t}$$

$$dV_{t}=\kappa (\theta - V_{t})dt+ \vartheta \sqrt{V_{t}}dW^{(V)}_{t}$$

where $W^{(S)}$ and $W^{(V)}$ are correlated Brownian motions with correlation parameter $\rho$

Implementation of Heston SV model is contained in

src/stochvolmodels/pricers/heston_pricer.py

Running log-normal SV pricer

Basic features are implemented in

examples/calibration/run_lognormal_sv_pricer.py

Option data for examples and calibration

All supported data routes converge to the same lightweight OptionChain, so fitting and pricing code does not depend on the original provider:

The repository does not redistribute CBOE, ThetaData, or Tardis records. OCA owns provider access, normalization, and caches; SVM receives only the strikes, option types, forwards, discounts, and bid/ask quotes required for an illustration or calibration.

Ready chain: no credentials and no OCA runtime dependency

The package includes one two-maturity chain captured from OCA's deterministic simulator. It is generated data rather than a vendor snapshot and is suitable for documentation, notebooks, and tests:

import numpy as np

from stochvolmodels.data.sample_option_chains import get_oca_simulated_chain_data
from stochvolmodels.fitters import calc_logsv_ivols, fit_logsv_ivols

chain = get_oca_simulated_chain_data()
idx = 1  # one-month slice
log_strikes = np.log(chain.strikes_ttms[idx] / chain.forwards[idx])
mid_vols = 0.5 * (chain.bid_ivs[idx] + chain.ask_ivs[idx])
fit = fit_logsv_ivols(log_strikes, mid_vols, chain.ttms[idx])
fitted_vols = calc_logsv_ivols(log_strikes, **fit)

Run the complete plotting example with:

python examples/calibration/run_logsv_smile_fitter.py --maturity 1m

The same chain can be passed to LogSVPricer.calibrate_model_params_to_chain() for a full analytic model calibration. The approximate fitter is a fast three-parameter smile illustration; it is not a substitute for calibrating the full term-structure model.

Loading cached SPX/VIX chains for experiments

Empirical CBOE chains remain owned and normalized by OptionChainAnalytics. Install the optional packages separately, configure RESOURCE_PATH in src/stochvolmodels/settings.yaml, place the normalized cache under its cboe_options/ subdirectory, and request only the observation window needed by the experiment:

pip install "stochvolmodels[research]" "option-chain-analytics[cboe]>=5.0.0"

import pandas as pd

from stochvolmodels.data.fetch_option_chain import load_cboe_option_chain

option_chain = load_cboe_option_chain(
    ticker='SPX',
    value_time=pd.Timestamp('2023-11-08 22:00:00+00:00'),
    days_map={'1w': 7, '1m': 21, '3m': 63},
    delta_bounds=(None, None),
)

The adapter reads OCA's ignored per-underlying Parquet cache and returns SVM's existing lightweight OptionChain; it does not copy the dataset or add provider metadata to the calibration object. See examples/calibration/load_cboe_option_chain.py for SPX and VIX cases.

For a credential-free end-to-end OCA 5 conversion and LogSV calibration, run:

python examples/calibration/run_oca_logsv_calibration.py

The example uses OCA's deterministic simulated panel. Replace its loader with any normalized OCA OptionsDataDFs source while keeping the same load_option_chain and LogSV calibration calls. Select LocalTests.CONVERT_CHAIN or LocalTests.CALIBRATE_LOGSV in the script's main guard.

Running the cache-first SPY monthly prototype

OCA's ThetaData cache can drive time-series plots, approximate smile fitting, and full LogSV calibration from one example. Build the cache in the OptionChainAnalytics checkout, then run:

# From the OptionChainAnalytics checkout:
python examples/build_thetadata_eod_cache.py --ticker SPY \
    --start-date 2026-07-01 --end-date 2026-07-31

# From the StochVolModels checkout:
python examples/calibration/run_spy_thetadata_month.py --case all \
    --output-dir outputs/spy_thetadata_july_2026

The default window is July 2026 and the calibration observation is 17 July. Use --cache-root when the cache is not under <RESOURCE_PATH>/thetadata_options/spy. The empirical Parquet files and generated figures remain ignored local artifacts. The approximate smile utilities are available from stochvolmodels.fitters; all Black prices used by their synthetic grid helper come from vanilla-option-pricers.

To run only the approximate smile fit or only the full LogSV calibration:

python examples/calibration/run_spy_thetadata_month.py --case smile
python examples/calibration/run_spy_thetadata_month.py --case calibrate

Programmatically, load the observation once and use it exactly like the bundled chain:

from pathlib import Path

import pandas as pd

from stochvolmodels import local_path as lp
from stochvolmodels.data.fetch_option_chain import load_thetadata_option_chain

cache_root = Path(lp.get_resource_path()) / "thetadata_options" / "spy"
chain = load_thetadata_option_chain(
    cache_root=cache_root,
    value_time=pd.Timestamp("2026-07-17 23:59:00", tz="America/New_York").tz_convert("UTC"),
    days_map={"1w": 7, "3w": 21, "6w": 42},
    delta_bounds=(-0.05, 0.05),
)

run_spy_thetadata_month.py shows both downstream paths: fit_logsv_ivols() for a single-slice approximation and LogSVPricer.calibrate_model_params_to_chain() for the full model.

Choosing hourly or EOD Tardis data for paper workflows

The cryptocurrency paper code uses two explicit conventions rather than resampling silently:

• load_tardis_hourly_options_data() and load_tardis_hourly_option_chain() preserve the raw hourly BTC/ETH observation grid for intraday studies and historical paper calculations.
• load_tardis_eod_options_data() and load_tardis_eod_option_chain() read OCA's standardized exact-08:00 UTC cache for daily chain reports and calibrations. The chain loader requests a bounded lookback and selects the latest observation at or before the timezone-aware valuation time, so it does not look ahead.

Both routes default to <RESOURCE_PATH>/tardis; configure the ignored local settings file rather than hardcoding a machine path. See papers/jump_risk_premia_clustered_jumps/README.md for the script-by-script data policy.

Imports:

import numpy as np 
import stochvolmodels as sv
from stochvolmodels import LogSVPricer, LogSvParams, OptionChain

Computing model prices and vols

# instance of pricer
logsv_pricer = LogSVPricer()

# define model params    
params = LogSvParams(sigma0=1.0, theta=1.0, kappa1=5.0, kappa2=5.0, beta=0.2, volvol=2.0)

# 1. compute the price
model_price, vol = logsv_pricer.price_vanilla(params=params,
                                             ttm=0.25,
                                             forward=1.0,
                                             strike=1.0,
                                             optiontype='C')
print(f"price={model_price:0.4f}, implied vol={vol: 0.2%}")

# 2. prices for slices
model_prices, vols = logsv_pricer.price_slice(params=params,
                                             ttm=0.25,
                                             forward=1.0,
                                             strikes=np.array([0.9, 1.0, 1.1]),
                                             optiontypes=np.array(['P', 'C', 'C']))
print([f"{p:0.4f}, implied vol={v: 0.2%}" for p, v in zip(model_prices, vols)])

# 3. prices for option chain with uniform strikes
option_chain = OptionChain.get_uniform_chain(ttms=np.array([0.083, 0.25]),
                                            ids=np.array(['1m', '3m']),
                                            strikes=np.linspace(0.9, 1.1, 3))
model_prices, vols = logsv_pricer.compute_chain_prices_with_vols(option_chain=option_chain, params=params)
print(model_prices)
print(vols)

Running model calibration to sample Bitcoin options data

btc_option_chain = sv.get_btc_test_chain_data()
params0 = LogSvParams(sigma0=0.8, theta=1.0, kappa1=5.0, kappa2=None, beta=0.15, volvol=2.0)
btc_calibrated_params = logsv_pricer.calibrate_model_params_to_chain(option_chain=btc_option_chain,
                                                                    params0=params0,
                                                                    constraints_type=sv.ConstraintsType.INVERSE_MARTINGALE)
print(btc_calibrated_params)

logsv_pricer.plot_model_ivols_vs_bid_ask(option_chain=btc_option_chain,
                               params=btc_calibrated_params)

The full fitted-surface figure is generated by the paper workflow and is not committed as a build artifact.

Comparison of model prices vs MC

btc_option_chain = sv.get_btc_test_chain_data()
uniform_chain_data = OptionChain.to_uniform_strikes(obj=btc_option_chain, num_strikes=31)
btc_calibrated_params = LogSvParams(sigma0=0.8327, theta=1.0139, kappa1=4.8609, kappa2=4.7940, beta=0.1988, volvol=2.3694)
logsv_pricer.plot_comp_mma_inverse_options_with_mc(option_chain=uniform_chain_data,
                                                  params=btc_calibrated_params,
                                                  nb_path=400000)
                                           

The full analytic-versus-Monte-Carlo figure is generated by the paper workflow.

Analysis and figures for the paper

The paper figures and equation-mapped analysis live in

papers/logsv_model_with_quadratic_drift

Running Heston SV pricer

Examples are implemented here

examples/pricing/run_heston_sv_pricer.py
examples/pricing/run_heston.py

Content of run_heston.py

import numpy as np
import matplotlib.pyplot as plt
from stochvolmodels import HestonPricer, HestonParams, OptionChain

# define parameters for bootstrap
params_dict = {'rho=0.0': HestonParams(v0=0.2**2, theta=0.2**2, kappa=4.0, volvol=0.75, rho=0.0),
               'rho=-0.4': HestonParams(v0=0.2**2, theta=0.2**2, kappa=4.0, volvol=0.75, rho=-0.4),
               'rho=-0.8': HestonParams(v0=0.2**2, theta=0.2**2, kappa=4.0, volvol=0.75, rho=-0.8)}

# get uniform slice
option_chain = OptionChain.get_uniform_chain(ttms=np.array([0.25]), ids=np.array(['3m']), strikes=np.linspace(0.8, 1.15, 20))
option_slice = option_chain.get_slice(id='3m')

# run pricer
pricer = HestonPricer()
pricer.plot_model_slices_in_params(option_slice=option_slice, params_dict=params_dict)

plt.show()

Supporting Illustrations for Public Papers

As illustrations of different analytics, this repository includes the directory papers/ with codes for computations and visualisations featured in several papers for

1. "Log-normal Stochastic Volatility Model with Quadratic Drift" by Artur Sepp and Parviz Rakhmonov: https://www.worldscientific.com/doi/10.1142/S0219024924500031

papers/logsv_model_with_quadratic_drift

2. "What is a robust stochastic volatility model" by Artur Sepp and Parviz Rakhmonov, SSRN: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4647027

papers/volatility_models

3. "Valuation and Hedging of Cryptocurrency Inverse Options" by Artur Sepp and Vladimir Lucic, SSRN: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4606748

papers/inverse_options

4. "Unified Approach for Hedging Impermanent Loss of Liquidity Provision" by Artur Sepp, Alexander Lipton and Vladimir Lucic, SSRN: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4887298

papers/il_hedging

5. "Stochastic Volatility for Factor Heath-Jarrow-Morton Framework" by Artur Sepp and Parviz Rakhmonov, Review of Derivatives Research, 2025, 28(3), article 12: https://doi.org/10.1007/s11147-025-09217-4 (preprint: http://ssrn.com/abstract=4646925)

papers/sv_for_factor_hjm

6. "Jump risk premia in the presence of clustered jumps" by Francis Liu, Natalie Packham and Artur Sepp, SSRN: https://ssrn.com/abstract=4735365. The repository folder contains related development code and is not an exact replication package.

papers/jump_risk_premia_clustered_jumps

Local resource and output paths

Copy src/stochvolmodels/settings.yaml.example to the ignored src/stochvolmodels/settings.yaml and set the two machine-local roots:

RESOURCE_PATH:
  "C:\\Users\\me\\analytics\\resources\\"
OUTPUT_PATH:
  "C:\\Users\\me\\analytics\\outputs\\"

Use the same import in examples and paper workflows:

from stochvolmodels import local_path as lp

local_path = f"{lp.get_resource_path()}bbg_vols\\"

Both getters return absolute strings with a trailing separator, following the qis convention. Convert the result to pathlib.Path when a library or operation benefits from path objects. The local YAML is not included in Git or package distributions; PyYAML is loaded only when it exists.

Project Structure

StochVolModels/
├── src/
│   └── stochvolmodels/
│       ├── data/                 # option-chain containers and sample data
│       ├── fitters/              # approximate LogSV smile fitter and Student-t utilities
│       ├── pricers/              # analytic, transform, and Monte Carlo models
│       ├── utils/                # quadrature, payoff, plotting, and rate helpers
│       └── tests/                # shipped pytest suite and regression data
├── examples/                     # repository-only runnable workflows
│   ├── getting_started/
│   ├── pricing/
│   ├── calibration/
│   └── options_time_series_data/
├── papers/                       # paper replications and labelled development code
├── docs/                         # documentation sources and figures
└── README.md

Ecosystem

This package is part of an open-source Python stack for quantitative finance — full catalogue at github.com/ArturSepp:

Dependency links within the stack: optimalportfolios builds on qis and factorlasso; trendfollowing and privateassets build on qis.

Contributing

See CONTRIBUTING.md for project scope, development commands, numerical-change rules, bug reports, questions/support, pull requests, conduct, and contribution licensing.

License

This project is licensed under the MIT License - see the LICENSE.txt file for details.

Citation

If you use this package in your research, please cite the relevant papers:

@misc{sepp2026stochvolmodels,
  title={StochVolModels: Python Implementation of Stochastic Volatility Models},
  author={Sepp, Artur},
  year={2026},
  version={2.2.0},
  howpublished={\url{https://github.com/ArturSepp/StochVolModels}},
  note={Python package for pricing analytics and Monte Carlo simulations}
}

@article{sepprakhmonov2023,
title={Log-normal stochastic volatility model with quadratic drift},
author={Sepp, Artur and Rakhmonov, Parviz},
journal={International Journal of Theoretical and Applied Finance},
volume={26},
number={8},
year={2023},
url={https://www.worldscientific.com/doi/epdf/10.1142/S0219024924500031}
}

@article{sepprakhmonov2023b,
title={What is a robust stochastic volatility model},
author={Sepp, Artur and Rakhmonov, Parviz},
year={2023},
note={Working paper},
url={http://ssrn.com/abstract=4647027}
}

@article{lucicsepp2024,
title={Valuation and hedging of cryptocurrency inverse options},
author={Lucic, Vladimir and Sepp, Artur},
journal={Quantitative Finance},
volume={24},
number={7},
pages={851--869},
year={2024},
url={https://www.tandfonline.com/doi/full/10.1080/14697688.2024.2364804}
}

@article{sepprakhmonov2025,
title={Stochastic volatility for factor Heath-Jarrow-Morton framework},
author={Sepp, Artur and Rakhmonov, Parviz},
volume={28},
number={3},
pages={12},
year={2025},
journal={Review of Derivatives Research},
doi={10.1007/s11147-025-09217-4},
note={Preprint: http://ssrn.com/abstract=4646925}
}

Acknowledgments

Special thanks to co-authors and collaborators:

• Parviz Rakhmonov
• Vladimir Lucic
• Alexander Lipton

For additional research and advanced analytics, see the companion modules and papers included in this package.