tsecon

High-performance time series econometrics: a Rust core with a Python-first API.

Pre-1.0 software under active development. The name is settled — tsecon is what you install and import — but the API may still change before 1.0.

tsecon brings the scattered toolkit of modern time series econometrics — diagnostics and specification tests, ARIMA, GARCH, VARs and structural identification, local projections, Bayesian VARs, nowcasting, predictive regressions, panels, the term structure — into one fast, validated, Python-first library. The compute core is written from scratch in Rust (no BLAS, no heavy dependencies) so bootstrap inference and Monte Carlo work that is painfully slow elsewhere runs in seconds, with results bit-reproducible at any thread count.

Every estimator is validation-gated: its numbers are checked against a reference implementation (statsmodels, SciPy, NumPy, arch, linearmodels, scikit-learn, ArviZ) or a documented closed form before it ships, and a cross-library parity gate re-verifies the agreement in CI on every push. The library ships no data loaders and makes no network calls — the only runtime dependency is NumPy.

Install

pip install tsecon

A single self-contained wheel whose only runtime dependency is NumPy — no Rust toolchain, no system BLAS. Prebuilt wheels cover Linux (x86_64, aarch64), macOS on Apple Silicon (arm64), and Windows (x64), for every Python >= 3.9. Plotting is an optional extra:

pip install 'tsecon[plots]'    # adds matplotlib for the .plot_*() methods

Then check what you got:

import tsecon
print(tsecon.__version__)   # 0.2.0

There is no prebuilt wheel for Intel macOS (x86_64); on that platform, and for building from a checkout, pip compiles from the source distribution, which needs a Rust toolchain and Python >= 3.9. Contributors build with maturin:

pip install maturin
maturin develop --release -m bindings/python/Cargo.toml   # builds + installs into the active venv

A taste

import numpy as np
import tsecon

rng = np.random.default_rng(0)
y = np.cumsum(rng.standard_normal(300))          # a random walk

tsecon.check_stationarity(y)["recommendation"]   # -> "Difference"

# Fit a VAR, read off impulse responses, decompose the variance
data = rng.standard_normal((200, 3))
irf  = tsecon.var_irf(data, lags=2, horizon=16)          # [h][response][shock]
fevd = tsecon.var_fevd(data, lags=2, horizon=16)   # [h][variable][shock]

# Robust SEs, exact-MLE ARIMA with a forecast fan, GARCH with robust SEs,
# instrumented local projections, a fiscal multiplier, a Bayesian VAR...
tsecon.ols(y, X, se_type="hac")
tsecon.arima_fit(y, p=1, d=1, q=1, forecast_steps=12, conf_alpha=0.1)
tsecon.garch_fit(returns, vol="gjr", dist="t")
tsecon.lp_multiplier(y_out, spending, instrument, horizons=20)
tsecon.bvar_irf_draws(data, lags=2, horizon=12, n_draws=800)

Every function takes plain NumPy arrays and returns arrays or dicts of documented keys. An opt-in tsecon.results layer wraps the same calls in objects that render themselves (.summary(), .plot_irf()) without changing the dict contract.

What's here today

126 functions across the full applied workflow: the diagnostic battery and unit-root workflow (ADF, KPSS, check_stationarity, and the one-call check_series battery with model recommendations); specification and stability tests (White, Breusch-Pagan, RESET, Chow, CUSUM); robust and HAC standard errors; the bootstrap family; an exact-diffuse Kalman filter; ARIMA, GARCH/GJR/EGARCH, and GAS score-driven volatility; VAR/SVAR with sign-restricted identification, FAVAR, and Diebold-Yilmaz connectedness; local projections (lag-augmented, LP-IV, state-dependent, and the Ramey-Zubairy integral multiplier); a Minnesota-NIW Bayesian VAR; GMM and IV-GMM; predictive regressions with IVX inference; heterogeneous panels (mean group, CCE-MG, PMG); cointegration and Markov switching; MIDAS and DFM nowcasting with a news decomposition; multivariate GARCH; realized volatility; spectral analysis; long memory; recession-probability models; survey-expectations tools; the Nelson-Siegel/Svensson term structure with the arbitrage-free (AFNS) adjustment; and a linear rational-expectations (DSGE-lite) solver.

The library ships with complete type stubs (py.typed), so autocomplete and type checking work out of the box.

Learn more

The full documentation — a 15-chapter guide to time series econometrics, model cards for every estimator family, a worked figure gallery, two replications of published results, a Monte Carlo validation suite, and an honest benchmark harness — lives in the project repository.

License

MIT OR Apache-2.0.