cultivars 1.0.0a3

Research-grade time series econometrics for Python: ARIMA to state space, VAR to structural identification, Bayesian and spectral methods, on one estimation surface.

CI / Quality
Security
Coverage
Packaging
Distribution
License
Usage
Research / Index

Pre-release. cultivars is at 1.0.0a3. The modelling surface is complete and tested; the public API is settling toward 1.0.0 and module paths may still move between alphas. Install with --pre and pin the exact version in anything that must not change under you.

Used by & Featured In

Note: Listing does not imply endorsement or affiliation.

Institutions / Organizations

What it is

cultivars is a single, typed, dependency-light library for the econometric time-series toolkit, written so that a research result can be reproduced from the summary it prints. Every model is a class holding a specification; fit() (or identify()) returns an immutable record that renders the same way everywhere (print, a notebook cell, summary()), carries its own diagnostics, and refuses to report what its assumptions do not support. Models that need an input the data cannot supply take it as an argument you chose, and the result restates that choice in its own text.

Dependencies are numpy, scipy and fedfred (FRED data access). pandas and polars are optional and only used for the to_pandas() / to_polars() conveniences on results.

Installation

Requires Python 3.12 or later.

pip install --pre cultivars

With the DataFrame extras:

pip install --pre "cultivars[pandas]"      # or [polars]

With uv:

uv add --prerelease allow cultivars

For a reproducible research environment, pin the alpha explicitly: cultivars==1.0.0a2. A conda-forge package will follow the first stable release.

Quick start

A univariate model

import numpy as np
from cultivars.univariate.box_jenkins import ARIMA

rng = np.random.default_rng(0)
y = np.zeros(300)
for t in range(1, 300):
    y[t] = 0.6 * y[t - 1] + rng.standard_normal()

res = ARIMA(y, order=(1, 0, 0)).fit()
print(res)

                             ARIMA(1, 0, 0) Results
================================================================================
Model:                  ARIMA(1, 0, 0)  Log-likelihood:               -430.692
Trend:                               c  AIC:                           867.383
Exog regressors:                     0  BIC:                           878.495
Observations:                      300  HQIC:                          871.830
--------------------------------------------------------------------------------
              coef
const      -0.0961
ar.L1       0.6245
sigma2      1.0323
================================================================================
Stationary: True   max |AR root| = 0.6245

Every result is a frozen dataclass: res.params, res.resid, res.information_criteria, res.is_stationary are attributes, res.compare(other) and res.likelihood_ratio_test(other) rank and test nested specifications, and res.to_pandas() hands the per-observation series to a DataFrame.

A VAR, its diagnostics, and a structural view

from cultivars.multivariate.reduced_form.vector_autoregression import VAR
from cultivars.multivariate.structural.zero_restrictions import RecursiveSVAR

Y = np.zeros((300, 2))
for t in range(1, 300):
    Y[t] = np.array([[0.5, 0.1], [0.0, 0.4]]) @ Y[t - 1] + rng.standard_normal(2)

var = VAR(Y, order=1, names=("gdp", "infl")).fit()
var.forecast(3)                              # (3, 2) point forecasts
var.irf(20)                                  # orthogonalized responses by default
print(var.granger_causality("infl", "gdp"))  # a Wald test with its verdict

                                   Wald Test
================================================================================
Verdict at 5%:                    keep
--------------------------------------------------------------------------------
null                                statistic   df   p-value
infl does not Granger-cause gdp        2.6420    1    0.1041
================================================================================

Identification is a separate step that takes the fitted reduced form and the restriction you are prepared to defend:

svar = RecursiveSVAR(var, order=("gdp", "infl")).identify()
svar.impact                 # (k, k) impact matrix, lower triangular in the declared order
svar.irf(20)                # structural responses
svar.fevd(20)               # variance shares, summing to one across identified shocks
svar.structural_shocks()    # recovered unit-variance shocks
print(svar.restriction)

Recursive ordering gdp -> infl: each variable responds on impact only to shocks at or before its own position. Permuting the ordering changes the answer; that is the identifying assumption, not a numerical artifact.

The same VAR result feeds every scheme in cultivars.multivariate.structural: sign restrictions return the accepted set rather than a point, external instruments return exactly the one column they identify, and the statistical schemes test their own identifying condition and print the verdict.

Volatility

from cultivars.univariate.conditional_variance import GARCH

r = rng.standard_normal(500) * np.sqrt(np.r_[np.ones(250), 3 * np.ones(250)])
garch = GARCH(r, p=1, q=1).fit()
print(garch)

                              GARCH(1, 1) Results
================================================================================
Model:                     GARCH(1, 1)  Log-likelihood:               -863.345
Mean:                         constant  AIC:                          1734.690
AR lags:                             0  BIC:                          1751.548
Observations:                      500  HQIC:                         1741.305
--------------------------------------------------------------------------------
                coef
const        -0.0239
omega         0.0171
alpha[1]      0.0627
beta[1]       0.9359
================================================================================
Persistence: 0.9985   Covariance stationary: True   Half-life: 465.5
Unconditional variance: 11.4758

A diagnostic

from cultivars.diagnostics.unit_roots import adf

walk = np.cumsum(rng.standard_normal(300))
print(adf(walk, trend="c"))

                                    ADF Test
================================================================================
Null:                        unit root  Trend:                               c
Lags:                                0  Method:                   aic (max 15)
Observations:                      299  Verdict at 5%:          keep unit root
--------------------------------------------------------------------------------
statistic       value   p-value       1%       5%      10%
ADF           -0.7203    0.8414   -3.452   -2.871   -2.572
================================================================================
Rejection lies in the lower tail.

Design

Three commitments run through the package and are worth knowing before you read any docstring.

Estimation and identification are separate acts. A reduced form is fitted once and reports only what the data determine. Anything that needs an assumption (a causal ordering, a sign pattern, an instrument, a prior) is constructed from that result, declares what it adds, and restates the declaration in its summary, so an impulse response can always be traced to the assumption that signed it.

Results refuse what they cannot support. A partially identified scheme carries only the columns it identified. A set-identified scheme returns the set. A mixture model refuses the regime-count likelihood-ratio test its likelihood cannot answer. A long-run restriction refuses a unit root. A particle-filter likelihood says it is an estimate. Where a method is a known approximation, the summary names it.

Failures say what kind of fix is needed. Every exception the package raises derives from cultivars.exceptions.CultivarsError and from one of three subclasses: DimensionError (reshape the data), SpecificationError (change the choice), NumericalError (look at the numbers). Each also derives from the matching builtin, so code written against ValueError and RuntimeError keeps working.

Packages re-export modules rather than classes; import a model from the module that defines it, as in the examples above.

Documentation

• Stable: https://nikhilxsunder.github.io/cultivars/stable/
• Development: https://nikhilxsunder.github.io/cultivars/dev/

The API reference is generated from the docstrings, every one of which carries a runnable, deterministic example; the examples double as the doctest suite.

Development

git clone https://github.com/nikhilxsunder/cultivars.git
cd cultivars
uv sync --all-groups
uv run pytest                                   # unit tests
uv run pytest --doctest-modules src/cultivars   # docstring examples
uv run ruff check . && uv run ruff format --check .
uv run mypy src
uv run make -C docs html                        # documentation

Contributions are welcome; see CONTRIBUTING.md for the conventions (Google-style docstrings with doctests, strict typing, no silent failure) and SECURITY.md for reporting vulnerabilities.

Citation

If cultivars contributes to published work, cite the concept DOI, which resolves to the latest release:

@software{sunder_cultivars,
  author  = {Sunder, Nikhil},
  title   = {cultivars: Research-grade time series econometrics for Python},
  year    = {2026},
  doi     = {10.5281/zenodo.23028870},
  url     = {https://github.com/nikhilxsunder/cultivars}
}

License

MIT. See LICENSE.