arch

What it is

arch fits conditional volatility models to a return series. arch_model(returns).fit() gets you a GARCH(1,1) with a constant mean and normal errors in one line, and every part of that is swappable: the mean model (constant, zero, AR, HAR, least squares, with or without exogenous regressors), the volatility process (ARCH, GARCH, EGARCH, APARCH, FIGARCH, HARCH, MIDAS hyperbolic, EWMA/RiskMetrics, RiskMetrics 2006, fixed variance) and the error distribution (normal, Student's t, skew-t, generalised error).

It is written by Kevin Sheppard, an econometrician at Oxford, and it shows: the estimators are the ones from the literature, the standard errors are the robust ones, and the parameter tables print like a paper rather than like a dashboard. Cython and optional Numba do the recursion.

Three other modules travel with it and are worth the install on their own. Unit root and cointegration: augmented Dickey-Fuller, DF-GLS, Phillips-Perron, KPSS, Zivot-Andrews and variance ratio tests, Engle-Granger and Phillips-Ouliaris cointegration tests, and CCR, dynamic OLS and fully modified OLS estimators. Bootstrap: IID, stationary, circular block and moving block, with confidence intervals for any statistic you can write as a function — a Sharpe ratio, for instance. Multiple comparison: Hansen's SPA and the Reality Check, StepM and the Model Confidence Set, which is the correct answer to "I tested two hundred strategies and one looked great".

Repository state on 19 September 2026. Version 8.0.0 was released on 21 October 2025 and is still the current one; the last commit to main was 14 September 2026, there were 27 open issues, and the repository is not archived. That September date flatters it, though: it is Dependabot bumping cibuildwheel, as is everything else on main since mid-June 2026, and the last commit that touched the library was a division-warning fix on 16 June. Kevin Sheppard is the sole maintainer. 8.0.0 was largely a compatibility release — the build moved to Meson, the Python floor rose from 3.9 to 3.10, Python 3.14 wheels appeared, and the code was made ready for NumPy 2.4 and pandas 3.

Pricing

Free under the NCSA licence, with nothing to buy. Thirty-one files ship per release: wheels for CPython 3.10 through 3.14 across the usual platforms, plus a source distribution. conda-forge carries it as arch-py, which is worth knowing because the import is still arch.

Data & coverage

No market data, but unusually for a library in this category it bundles example datasets — S&P 500 and NASDAQ prices, the VIX, WTI and crude, Fama-French factors, core CPI and two cross-sectional sets — so the documentation examples run offline. Everything else is a returns series you supply. The models are indifferent to what produced it.

Limitations

• Univariate only. No DCC, no BEKK, no multivariate volatility of any kind.
• Estimation is maximum likelihood with no Bayesian option, and long or awkward series can hit convergence problems that the package reports honestly and does not solve for you.
• The README's quick-start example still fetches FTSE prices through pandas_datareader with an end date of 2014. pandas_datareader is not a dependency, so the first example a new user copies does not run on a clean install.
• Forecasting has real footguns — analytic versus simulation versus bootstrap forecasts differ, and only simulation and bootstrap work for the asymmetric models. The documentation covers this; the API will not stop you.
• Releases are roughly yearly, so a NumPy or pandas break can sit unreleased for months even while main has the fix.

Alternatives

statsmodels covers the time-series ground around it — ARIMA, VAR, state space — and is a dependency of arch rather than a competitor, so the two are installed together anyway. For pricing models rather than volatility estimation, QuantLib. Anyone who only wants a rolling standard deviation should stop at pandas; arch earns its place when the question is whether volatility clusters, persists or reacts asymmetrically to losses.