Reading Returned Objects¶
The main UniBM entrypoints return lightweight dataclasses rather than raw tuples. Start with the fits created in Worked Examples; this page explains their fields and inference limits.
EVI fits¶
Read the result in this order:
fit.slopeis the headline UniBMxiestimatefit.confidence_intervalgives the uncertainty interval forxifit.regression_policyrecords the requestedOLS,FGLS, orAUTOpolicyfit.regressionrecords theOLSorFGLSestimator actually usedfit.ci_variantishc0for OLS orbootstrap_covfor FGLSfit.plateau_boundsshows which block-size window supported the fitfit.bootstrapstores the bootstrap metadata and covariance inputs used by FGLS fittingdesign_lifecontains the design-life-level estimates on the original data scale
The remaining fields such as curve and plateau are mainly for plotting,
diagnostics, and workflow-side reuse.
OLS does not bootstrap. FGLS and AUTO use adaptive bootstrap repetitions by
default unless a labeled bootstrap_result is supplied. bootstrap_reps="adaptive"
is explicit; an integer such as bootstrap_reps=480 uses exactly that fixed budget.
The checkpoints are 128, 256, 512, 768, and 1024, with fixed diagonal shrinkage 0.73.
FGLS fails closed when
covariance is unavailable or invalid; only AUTO can fall back to OLS, and
malformed caller-supplied covariance is always an error.
EI fits¶
Read the result in this order:
fit.theta_hatis the headline extremal-index estimatefit.confidence_intervalgives the uncertainty interval forthetafit.stable_windowshows which block-size region was pooledfit.base_pathandfit.regressionrecord which BM path and pooling rule produced the estimatefit.ci_variantidentifies the interval construction actually usedfit.standard_erroris on thethetascale;fit.z_standard_erroris the pooled regression-scale SE, before transforming back fromz = log(1 / theta)
The path-level fields are supporting diagnostics:
fit.path_levelrecords the observed block sizes retained on the finite pathfit.path_thetaandfit.path_eirretain the observed path values for plotting and method audits
For native EI prepared with a single block_sizes=[b], selected_level is that
fixed size and stable_window is None: no stability search was performed.
Threshold-only bundles prepared with path_keys=() contain no BM paths or grid.
FGLS versus OLS¶
Pooled EI fits always pool the observed stable-window path. If you switch from
regression="OLS" to regression="FGLS", the observed path is still what
gets pooled. The bootstrap result only contributes the cross-block covariance
matrix used for FGLS weighting.
bootstrap_bm_ei_path likewise defaults to reps="adaptive"; passing an integer
retains fixed-R sampling. It checks pooled theta and CI endpoints, plus the
unconstrained z fit and endpoints so clipping at theta=1 cannot conceal
Monte Carlo error. For adaptive monitoring, supply the same
covariance_shrinkage to bootstrap and fit when overriding the default 0.37.
Fixed-R bootstrap rejects explicit shrinkage; both modes return raw sample
covariance, which is regularized only when fitting. The pooled estimator still requires explicit
usable covariance for FGLS and never silently falls back to OLS.
Reading adaptive precision¶
The following fields are present on both EVI and EI fit objects:
| Field | Meaning |
|---|---|
bootstrap_reps_policy |
adaptive, fixed, or no bootstrap policy |
bootstrap_reps_used |
Number of bootstrap path draws used for covariance |
bootstrap_precision_met |
True: tolerance met; False: cap reached without meeting it; None: not assessed for this fit |
bootstrap_mcse_targets, bootstrap_mcse |
Matching target names and estimated Monte Carlo standard errors |
bootstrap_mcse_max_ratio |
Largest MCSE divided by its statistical-SE denominator |
covariance_shrinkage |
Shrinkage actually used for FGLS |
Adaptive sampling uses eight equal delete groups and two random partitions, refitting the same observed window after deleting bootstrap rows. It stops when every target's estimated MCSE is at most 10% of its statistical-SE denominator. The endpoint checks use the corresponding parameter's statistical SE as the denominator, not a separately estimated endpoint SE.
EVI checks xi and its two CI endpoints, not extrapolated design-life levels.
EI checks theta and its endpoints as well as the unconstrained z fit and its
endpoints. At R=1024, an unmet tolerance produces a warning and retains the fit
with bootstrap_precision_met=False. Fixed-R fits are not automatically assessed.
Changing a reused fit's window or shrinkage clears its old precision claim.
This controls conditional numerical error from a finite bootstrap budget. It does not guarantee CI coverage, account for selection uncertainty, or ensure small MCSE for a 50-year design-life level. Increasing R does not add observed extreme events and need not narrow a statistical CI.
Application CSV/JSON exports include these diagnostics. The case-study snapshots document the configuration used to generate them; rebuilding this site alone does not rerun the statistical analysis.