Inference¶
Two-stage standard errors¶
Two-stage QML is consistent but the stage-2 information matrix alone understates uncertainty: the correlation parameters are estimated on residuals that were themselves standardized by estimated marginals. pymgarch implements the Engle and Sheppard (2001) correction by stacking the per-observation scores of every marginal likelihood and the correlation likelihood into one M-estimator moment vector \(g_t(\theta)\), \(\theta = (\phi_1, \dots, \phi_N, \psi)\), and computing
\(\hat A\) is block lower-triangular because marginal scores do not depend on \(\psi\). All derivatives are central finite differences; marginal likelihoods are re-evaluated through arch’s own variance recursions with the backcast and variance bounds arch used during fitting held fixed, so the perturbed evaluations are consistent with the objective arch actually maximized.
summary() labels the estimates two-stage-robust. If the stacked bread
matrix is numerically singular, the library falls back to a stage-2-only
sandwich (labelled stage2-robust) and warns; those standard errors ignore
marginal estimation error.
Two further labels exist. Composite-likelihood fits report
composite-godambe: the same stacked sandwich, but the stage-2 scores are
those of the pairwise composite objective (Godambe information), which is
the correct covariance for the estimator that was actually maximized. BEKK
reports qml-robust: a single-stage QML sandwich (there is no marginal
stage to stack) that holds the estimated mean and the targeting covariance
fixed – the same fixed-targets approximation the two-stage models make,
which can understate SEs for very persistent processes. On boundary or
singular optima BEKK SEs degrade to a pseudo-inverse sandwich
(qml-robust-pinv) or NaN (qml-robust (failed)) with a warning; the fit
itself always survives.
Known approximations¶
Correlation targets \(\bar S\) and \(\bar N\) are held fixed at their point estimates, so targeting uncertainty is ignored. rmgarch makes the same approximation.
Set
compute_se=Falseonfit()to skip the sandwich entirely (useful in simulation studies where only point estimates matter).
Replication against rmgarch¶
scripts/make_fixtures.R fits rmgarch on a pinned dataset (percent returns
of five dji30ret constituents) and exports parameters, likelihoods, sigma and
residual paths, correlation targets, and one-step forecasts to JSON. The
test suite then:
evaluates pymgarch’s likelihood at R’s fitted parameters on R’s own volatility paths and correlation targets, isolating the correlation math from optimizer and univariate-initialization differences; and
runs the full pipeline on the same CSV and requires parameter agreement and a log-likelihood no worse than R’s optimum.
The tests skip automatically when the fixture file is absent.