Speed dashboard¶

Accuracy before speed. This page is generated from one run of the benchmark harness, whose rule is that an estimate must match an independent reference (statsmodels, arch, SciPy, scikit-learn) at a stated tolerance before it is timed. The parity matrix therefore comes first and is the deliverable; the timings below it are indicative, single-machine numbers and are labelled with the build that produced them.


Parity	65/65 metrics across 25 operations: ALL PASS
Faster than the reference	22/25 timed operations (RELEASE build)
Slower than the reference	3/25 — published below, not dropped
Measured	2026-09-12, tsecon 0.10.0, Intel(R) Xeon(R) Processor @ 2.10GHz (4 cores)

1. Parity matrix — the deliverable¶

max abs diff is the largest absolute disagreement between tsecon and the reference on that metric; tol is the tolerance the harness asserts, set from the known source of numerical difference (machine precision for closed-form linear algebra, optimiser tolerance for iterative QMLE fits) and never widened to pass. This table is machine-independent: a debug build produces the identical matrix.

Operation	Metric	max abs diff	tol	Result
ADF test (regression='c', fixed lag=4) vs `statsmodels.tsa.stattools.adfuller`	statistic	2.22e-16	1e-06	PASS
	p_value	1.11e-16	1e-06	PASS
	crit values	0.00e+00	1e-06	PASS
VAR(2) coefficients (2 vars, trend='c') vs `statsmodels.tsa.api.VAR`	coef matrix (5x2)	6.11e-16	1e-08	PASS
	log-likelihood	1.14e-13	1e-06	PASS
OLS + HAC (Newey-West) SEs (maxlags=4, corrected) vs `statsmodels OLS cov_type='HAC'`	params	3.11e-15	1e-08	PASS
	HAC bse	2.08e-16	1e-08	PASS
GARCH(1,1) QMLE (constant mean, normal) vs `arch.arch_model`	log-likelihood (rtol 1e-5)	1.46e-07	5e-02	PASS
	params (atol 1e-3)	7.84e-06	1e-03	PASS
GJR-GARCH(1,1,1) QMLE (constant mean, normal) vs `arch.arch_model (o=1)`	log-likelihood (rtol 1e-5)	2.46e-09	4e-02	PASS
	params (atol 1e-3)	1.53e-06	1e-03	PASS
EGARCH(1,1,1) QMLE (constant mean, normal) vs `arch.arch_model (vol='EGARCH')`	log-likelihood (rtol 1e-5)	3.03e-08	2e-02	PASS
	params (atol 1e-3)	1.49e-05	1e-03	PASS
KPSS test (regression='c', auto lags) vs `statsmodels.tsa.stattools.kpss`	statistic	5.55e-17	1e-10	PASS
	p_value (clipped)	0.00e+00	1e-10	PASS
	auto lags	0.00e+00	exact (0)	PASS
ACF (20 lags) + Bartlett SEs vs `statsmodels.tsa.stattools.acf`	acf (adjusted=False)	4.44e-16	1e-12	PASS
	Bartlett SE	4.16e-17	1e-12	PASS
	acf (adjusted=True)	3.33e-16	1e-12	PASS
PACF (15 lags, Yule-Walker + OLS) vs `statsmodels.tsa.stattools.pacf`	pacf (yw / ywm)	1.09e-15	1e-10	PASS
	pacf (ols)	2.60e-15	1e-10	PASS
Ljung-Box + Box-Pierce (lags 1..10) vs `statsmodels.stats.diagnostic.acorr_ljungbox`	lb_stat	6.22e-15	1e-10	PASS
	lb_pvalue	4.25e-15	1e-12	PASS
	bp_stat	3.55e-15	1e-10	PASS
	bp_pvalue	4.30e-15	1e-12	PASS
Jarque-Bera normality test vs `statsmodels.stats.stattools.jarque_bera`	statistic	4.26e-14	1e-10	PASS
	p_value	2.11e-29	1e-12	PASS
	skewness	8.33e-16	1e-12	PASS
	kurtosis	8.88e-16	1e-12	PASS
Engle ARCH-LM test (4 lags) vs `statsmodels.stats.diagnostic.het_arch`	LM statistic	7.07e-13	1e-09	PASS
	LM p_value	1.20e-13	1e-12	PASS
White heteroskedasticity test vs `statsmodels.stats.diagnostic.het_white`	LM statistic	3.55e-13	1e-09	PASS
	LM p_value	7.26e-15	1e-12	PASS
	F statistic	7.11e-14	1e-09	PASS
	F p_value	5.52e-15	1e-12	PASS
Breusch-Pagan test (Koenker studentised) vs `statsmodels.stats.diagnostic.het_breuschpagan`	LM statistic	0.00e+00	1e-09	PASS
	LM p_value	1.39e-17	1e-12	PASS
	F statistic	2.66e-15	1e-09	PASS
	F p_value	6.73e-15	1e-12	PASS
Ramsey RESET (powers of yhat up to 3) vs `statsmodels.stats.diagnostic.linear_reset`	F statistic	1.81e-13	1e-09	PASS
	p_value	2.04e-13	1e-10	PASS
	df (num, den)	0.00e+00	exact (0)	PASS
Johansen cointegration (3 vars, k_ar_diff=1) vs `statsmodels.tsa.vector_ar.vecm.coint_johansen`	eigenvalues	9.89e-14	1e-10	PASS
	trace stat	6.37e-11	1e-08	PASS
	max-eig stat	6.37e-11	1e-08	PASS
	trace crit (90/95/99)	0.00e+00	1e-12	PASS
	max-eig crit	0.00e+00	1e-12	PASS
VAR(2) orthogonalised IRF + FEVD (h=10) vs `statsmodels VARResults.irf/.fevd`	orth IRF (11x2x2)	1.67e-16	1e-10	PASS
	FEVD (10x2x2)	4.44e-16	1e-10	PASS
VAR(2) Granger causality F-test vs `statsmodels VARResults.test_causality(kind='f')`	F statistic	9.77e-15	1e-10	PASS
	p_value	3.83e-17	1e-12	PASS
	df (num, den)	0.00e+00	exact (0)	PASS
HP filter (lambda=1600, two-sided) vs `statsmodels.tsa.filters.hp_filter.hpfilter`	trend	5.41e-12	1e-08	PASS
	cycle	5.41e-12	1e-08	PASS
Baxter-King band-pass (low=6, high=32, k=12) vs `statsmodels.tsa.filters.bk_filter.bkfilter`	first_index (== K)	0.00e+00	exact (0)	PASS
	cycle	1.22e-15	1e-12	PASS
Christiano-Fitzgerald band-pass (low=6, high=32) vs `statsmodels.tsa.filters.cf_filter.cffilter`	cycle	2.55e-15	1e-12	PASS
	trend	3.55e-15	1e-12	PASS
Periodogram PSD (boxcar, n=4096) vs `scipy.signal.periodogram`	freqs	0.00e+00	1e-15	PASS
	psd	4.26e-14	1e-12	PASS
Welch PSD (Hann, nperseg=256, 50% overlap) vs `scipy.signal.welch`	freqs	0.00e+00	1e-15	PASS
	psd	1.07e-14	1e-12	PASS
Ridge regression (alpha=1.0, no intercept) vs `sklearn.linear_model.Ridge`	coef	2.66e-15	1e-10	PASS
Elastic net / lasso (coordinate descent) vs `sklearn.linear_model.ElasticNet`	coef (l1_ratio=1.0)	3.09e-13	1e-08	PASS
	coef (l1_ratio=0.5)	2.19e-12	1e-08	PASS

Result: 65/65 parity checks passed — the harness exited 0.

Tolerance notes recorded by the harness:

GARCH(1,1) QMLE (constant mean, normal) — QMLE optimisers differ; parity is asserted at optimiser tolerance (loglik rtol 1e-5, params atol 1e-3), not machine precision.
GJR-GARCH(1,1,1) QMLE (constant mean, normal) — Leverage-term QMLE; two different optimisers, so parity is at optimiser tolerance (loglik rtol 1e-5, params atol 1e-3), not machine precision.
EGARCH(1,1,1) QMLE (constant mean, normal) — Same parameterisation (mu, omega, alpha, gamma, beta) on both sides; parity at optimiser tolerance (loglik rtol 1e-5, params atol 1e-3).
KPSS test (regression='c', auto lags) — p-value is interpolated and clipped to [0.01, 0.10] by BOTH sides (Kwiatkowski table); parity is on the clipped value.
PACF (15 lags, Yule-Walker + OLS) — tsecon method='yw' is statsmodels method='ywm' (Yule-Walker, no mean adjustment).
Baxter-King band-pass (low=6, high=32, k=12) — bkfilter drops k observations at each end; tsecon returns the same trimmed series plus first_index = k, so the two align element-wise.
Elastic net / lasso (coordinate descent) — Both minimise (1/2n)||y-Xb||^2 + al1||b||_1 + (a/2)(1-l1)||b||^2. Tolerance reflects coordinate-descent stopping rules, not a formula difference.

2. Timings — indicative, subordinate to parity¶

Read the caveats before the numbers:

Build: RELEASE — detected via <repo>/bindings/python/python/tsecon/_core.abi3.so (19.4 MB) == <target>/release/lib_core.so.
Machine: Intel(R) Xeon(R) Processor @ 2.10GHz, 4 cores, Linux-6.18.44-fc-v24-x86_64-with-glibc2.39 (x86_64); Python 3.11.15 (CPython).
Versions: tsecon 0.10.0, numpy 2.4.6, scipy 1.17.1, statsmodels 0.15.0, arch 8.0.0, scikit-learn 1.9.0.
Method: best (minimum) of 20 wall-clock runs after a warm-up, on the harness's small synthetic inputs; the three QMLE volatility fits use at most 3 repeats. ratio is reference time / tsecon time, so values above 1 mean tsecon was faster. Order-of-magnitude information only.
Date: 2026-09-12T12:49:58+0000.
Per-call overhead: the public tsecon.kpss took 0.009 ms against 0.006 ms for the raw extension entry point tsecon._core.kpss on the same array — a fixed ~0.003 ms of Python-side argument validation per call. It is included in every tsecon timing below and dominates the rows whose compute is far below a millisecond; the ratios on those rows measure that wrapper, not the Rust core.

Operation	Reference	tsecon (ms)	reference (ms)	ratio
ADF test (regression='c', fixed lag=4)	`statsmodels.tsa.stattools.adfuller`	0.026	0.449	17.15x	faster
VAR(2) coefficients (2 vars, trend='c')	`statsmodels.tsa.api.VAR`	0.205	1.361	6.65x	faster
OLS + HAC (Newey-West) SEs (maxlags=4, corrected)	`statsmodels OLS cov_type='HAC'`	0.035	0.175	4.96x	faster
GARCH(1,1) QMLE (constant mean, normal)	`arch.arch_model`	49.460	16.765	0.34x	SLOWER
GJR-GARCH(1,1,1) QMLE (constant mean, normal)	`arch.arch_model (o=1)`	45.351	20.955	0.46x	SLOWER
EGARCH(1,1,1) QMLE (constant mean, normal)	`arch.arch_model (vol='EGARCH')`	101.518	11.789	0.12x	SLOWER
KPSS test (regression='c', auto lags)	`statsmodels.tsa.stattools.kpss`	0.009	0.051	5.93x	faster
ACF (20 lags) + Bartlett SEs	`statsmodels.tsa.stattools.acf`	0.030	0.056	1.86x	faster
PACF (15 lags, Yule-Walker + OLS)	`statsmodels.tsa.stattools.pacf`	0.023	0.794	34.05x	faster
Ljung-Box + Box-Pierce (lags 1..10)	`statsmodels.stats.diagnostic.acorr_ljungbox`	0.011	0.264	23.00x	faster
Jarque-Bera normality test	`statsmodels.stats.stattools.jarque_bera`	0.005	0.618	132.96x	faster
Engle ARCH-LM test (4 lags)	`statsmodels.stats.diagnostic.het_arch`	0.018	0.414	23.18x	faster
White heteroskedasticity test	`statsmodels.stats.diagnostic.het_white`	0.028	0.548	19.44x	faster
Breusch-Pagan test (Koenker studentised)	`statsmodels.stats.diagnostic.het_breuschpagan`	0.018	0.423	22.97x	faster
Ramsey RESET (powers of yhat up to 3)	`statsmodels.stats.diagnostic.linear_reset`	0.024	0.583	23.83x	faster
Johansen cointegration (3 vars, k_ar_diff=1)	`statsmodels.tsa.vector_ar.vecm.coint_johansen`	0.057	0.770	13.49x	faster
VAR(2) orthogonalised IRF + FEVD (h=10)	`statsmodels VARResults.irf/.fevd`	0.065	1.716	26.23x	faster
VAR(2) Granger causality F-test	`statsmodels VARResults.test_causality(kind='f')`	0.052	2.153	41.60x	faster
HP filter (lambda=1600, two-sided)	`statsmodels.tsa.filters.hp_filter.hpfilter`	0.034	0.945	28.04x	faster
Baxter-King band-pass (low=6, high=32, k=12)	`statsmodels.tsa.filters.bk_filter.bkfilter`	0.018	0.087	4.94x	faster
Christiano-Fitzgerald band-pass (low=6, high=32)	`statsmodels.tsa.filters.cf_filter.cffilter`	0.417	6.495	15.57x	faster
Periodogram PSD (boxcar, n=4096)	`scipy.signal.periodogram`	0.079	0.442	5.58x	faster
Welch PSD (Hann, nperseg=256, 50% overlap)	`scipy.signal.welch`	0.038	0.691	18.26x	faster
Ridge regression (alpha=1.0, no intercept)	`sklearn.linear_model.Ridge`	0.062	0.520	8.38x	faster
Elastic net / lasso (coordinate descent)	`sklearn.linear_model.ElasticNet`	0.050	0.408	8.18x	faster

tsecon was faster on 22 of 25 timed operations in this run. The losses are published with the wins: GARCH(1,1) QMLE (constant mean, normal) at 0.34x, GJR-GARCH(1,1,1) QMLE (constant mean, normal) at 0.46x, EGARCH(1,1,1) QMLE (constant mean, normal) at 0.12x. The harness README explains what was measured and deliberately not done about each of them.

3. How to reproduce¶

Build a release extension first — maturin develop without --release installs a debug build whose timings the harness refuses to label as speed claims (the parity matrix is identical either way):

pip install numpy scipy statsmodels arch scikit-learn maturin
cd bindings/python && maturin develop --release && cd ../..
python benchmarks/bench.py --json benchmarks/results/latest.json   # exits 0 iff every parity check passes
python benchmarks/render_dashboard.py                              # regenerates this page
python benchmarks/render_dashboard.py --check                      # exits 1 if the page is stale

The committed benchmarks/results/latest.json is the exact input this page was rendered from; the harness's provenance banner (CPU, versions, build mode) is stored in it, so any number quoted from this page can be traced to the machine and build that produced it. Different hardware will move every timing; it must not move a single parity row.