PaperlayerView exact revision on arXiv

Source-linked AI summary

Vine Copula VAR:From Recursive Margins to Joint Forecast Inference

Hunter Ng, Yubo Tao

arXiv:2610.07589v1econ.EMmath.STstat.MLstat.TH

TL;DR

Joint-event forecasts combine dependence estimated from historical errors with new marginal fits, but issued transforms carry overlapping estimation errors. The paper derives their joint influence with terminal forecast estimates and finds terminal-margin uncertainty larger than the archive term, while logit intervals improve lower-tail coverage in the studied designs.

  • Problem

    Issued historical transforms use date-specific marginal fits, so their overlapping estimation errors must be incorporated when forecasting joint events.

    Source: §1 · pp. 2–3

  • Method

    The paper derives a joint influence and feasible covariance estimator for sequential vine estimates and terminal forecast coefficients.

    Source: §6 · p. 28

  • Results

    In the simulations and real-time application, the archive variance cost is smaller than current marginal-estimation uncertainty, and logit intervals improve lower-tail coverage.

    Source: §6 · p. 28

  • Takeaways & Limitations

    The estimator supports repeated-sample intervals for conditional event probabilities at a random forecast state.

    Source: §6 · p. 28

    The simulations do not establish a sample-size threshold for when the estimation history is sufficiently long.

    Source: §4.2 · p. 21

Abstract

from arXiv · show

Joint-event forecasts often combine a dependence estimate based on past forecast errors with newly estimated marginal distributions. When each historical error retains the marginal fit available at its issue date, inference must account for an overlapping sequence of estimation errors. We derive their joint influence with the terminal forecast estimates in a stable Vine Copula VAR with normal innovation margins and a fixed, correctly specified Gaussian or positive Clayton vine. An intercept identity and the stable VAR filter reduce the historical correction to harmonically weighted innovation moments, while terminal slope uncertainty remains. The resulting covariance estimator gives asymptotically valid repeated-sample intervals for fixed one-sided event probabilities at the realized forecast state. In the Gaussian submodel, retaining issued transforms adds a positive semidefinite covariance term relative to refitting margins on the same observations. Monte Carlo simulations show that terminal-margin uncertainty is quantitatively more important than this additional term and that logit intervals improve lower-tail coverage in the designs studied. A real-time forecasting application to U.S. macroeconomic releases shows how marginal estimation contributes to uncertainty in predicted probabilities of joint contractions and identifies limitations of the stationary marginal model.

1 Introduction

This section develops joint-event inference for a stable VAR with a fixed vine copula when historical transforms retain the marginal fits available at their issue dates, deriving the joint influence of overlapping historical estimates and terminal forecast parameters and evaluating interval performance in simulations. It also applies the framework to real-time forecasts of joint contractions in U.S. employment and economic activity across 450 historical origins, distinguishing changes in marginal estimates from revisions to observations and identifying limitations of the stationary marginal model.

1 Introduction

  • The analysis derives the joint influence of overlapping dated marginal fits, whose errors affect many subsequent PITs and pass through later vine conditional transforms.It combines the sequential vine estimate with terminal coefficients and scales, including their cross-covariances.Source: §1 · p. 3
  • An intercept identity and the stable VAR filter cancel the centered historical lag-regressor contribution, leaving harmonic weights on innovation location and scale moments.Earlier observations receive larger weights because they enter more subsequent regressions, while terminal slopes remain in the forecast gradient.Source: §1 · p. 3
  • Studentized intervals have repeated-sample coverage for the conditional probability at each sample’s realized state under the stated fixed-law, correctly specified model conditions.The theory covers fixed one-sided events with positive limiting variance almost surely, including mixed directions and variable subsets; bounded rectangles require event-specific nondegeneracy.Source: §1 · p. 3
  • In the Gaussian submodel, retaining issued transforms adds a positive semidefinite term to the copula covariance block relative to refitting margins on the same observations.The terminal-margin and cross-covariance blocks coincide across the two procedures; replacing the history with revised observations is a separate comparison.Source: §1 · p. 3
  • For lower-tail events, logit intervals achieved 93.8–96.1% coverage in Gaussian vines and 93.3–95.2% in Clayton vines, versus archive Wald intervals’ 85.8–92.6% and 86.4–93.0%.Some coverage shortfalls remained, particularly with persistent dynamics; in these designs, terminal-margin uncertainty was much more consequential than the additional archive term.Source: §1 · p. 4

Real-time forecasting application

  • The application studies joint contractions in U.S. employment and economic activity at 450 historical origins, separating changes in marginal estimates from revisions to underlying observations.Source: §1 · pp. 4–5
  • The archived vine showed no consistent Brier-loss advantage over benchmarks, while current marginal estimation accounted for most adjustment to reported standard errors.Source: §1 · p. 5
  • Marginal diagnostics questioned the stationary, constant-scale law even before the pandemic.Source: §1 · p. 5
  • The application quantifies parameter uncertainty in real-time forecasts and identifies features that richer marginal models would need to address.Source: §1 · p. 5

2 Model, forecast target, and estimation

This section specifies a stable VAR with contemporaneous dependence represented by a fixed vine, and defines the joint-event forecast using historical PITs that retain issue-date margins and terminal margins at the realized state. Its sequential-fit analysis focuses on marginal estimation errors, leaving time-varying coefficients and latent volatility for separate analysis.

2.1 A stationary VAR and a conditional event

  • A causal stationary stable VAR has innovations independent across dates but potentially dependent across coordinates, and its fixed-threshold one-step event probability varies with the random terminal state.The innovation coordinates have Gaussian margins, while their joint dependence may be non-Gaussian; estimating the threshold itself would add another influence term.Source: §2.1 · pp. 5–6

2.2 Recursive margins and the archive

  • Historical PITs retain the marginal fits available at each issue date, while the next forecast uses margins estimated through the terminal date.The archived PITs form the copula-estimation inputs; their finite-sample margins may not be exactly uniform, and inference accounts for their generated estimation error. In the Gaussian submodel, refitting margins on the same observations weakly reduces asymptotic variance, so retaining dated transforms follows the specified training protocol rather than an efficiency claim.Source: §2.2 · p. 7

2.3 Fixed vine and sequential fit

  • The vine structure, depth, and pair families are fixed and correctly specified, with omitted edges treated as independence copulas; pair parameters are fit sequentially in tree order.Earlier-edge estimates determine later conditional transforms, so their sampling error enters later scores; the theory does not account for uncertainty from selecting the vine structure or pair families.Source: §2.3 · p. 9
  • Time-varying coefficients and latent volatility require separate analysis of marginal estimation errors.Source: §2.3 · p. 9

3 Inference for conditional joint forecasts

The section develops inference for conditional joint-event probabilities at the realized forecast state, combining recursively estimated marginal and vine-score errors with terminal VAR uncertainty. It establishes a joint covariance limit and conditions for valid event inference, compares archive-based covariance with dynamic refitting, and specifies logit intervals and implementation requirements.

3 Inference for conditional joint forecasts

  • The inferential problem has two parts: propagate recursive marginal error through the sequential vine fit and terminal VAR estimates, then evaluate the joint limit at the random next-forecast state.The section treats these in order to derive the covariance matrix and justify the event interval.Source: §3 · p. 10

3.1 Scores and regularity

  • Stable VAR filtering and growing training prefixes establish a joint CLT without an additional mixing-rate condition, but the bounds do not ensure uniform coverage near unit roots or pair-parameter boundaries.Correct copula specification is needed for the conditional event probability.Source: §3.1 · p. 11

3.2 The joint dynamic limit

  • The additional covariance term captures variation from dated marginal fits and acts only on the copula block, while the common outer-product term retains cross-covariances.Innovations reused across successive fits receive harmonic weights, with earlier innovations weighted more heavily.Source: §3.2 · pp. 12–13
  • The sequential estimator is consistent, and its covariance estimator is positive semidefinite and consistent under Assumptions 1–4.The joint limit separates the common score-and-terminal-estimate variation from an additional historical marginal-fit correction affecting only the copula block.Source: §3.2 · p. 13

3.3 A confidence interval at the random forecast state

  • For a fixed event, the forecast-state probability map must be twice continuously differentiable, with positive conditional variance almost surely under the stationary lag-state law.The event gradient includes the entire forecast state.Source: §3.3 · p. 14
  • Studentized intervals have repeated-sample coverage for conditional event probabilities evaluated at a random realized forecast state.The forecast error has a state-dependent Gaussian projection variance, so feasible standard errors adapt to the realized state. This does not guarantee coverage conditional on every training history or provide a prediction set for the next binary outcome.Source: §3.3 · p. 15
  • Asymptotic independence of estimation error and the terminal state follows by separating recent innovations from the stable VAR estimation influence, not from finite-sample independence.A slowly growing block of recent innovations approximates the terminal state while contributing negligibly to normalized estimation influence.Source: §3.3 · p. 15

3.4 Dynamic refitting and finite training weights

  • Finite-training-weight covariance adjustment is positive semidefinite and consistent, but its linearization omits an exact finite-sample variance or coverage-improvement guarantee.The adjustment retains the archive’s initial training fraction in the covariance of the leading array and can replace the limiting-weight estimator in the event interval.Source: §3.4 · p. 17

3.5 Probability boundaries and the cost of the archive

  • Logit intervals yield asymmetric probability endpoints while preserving first-order repeated-sample coverage under the event-specific positivity condition.The proof localizes away from probability boundaries and small projected variances, without requiring a global lower bound over lag states.Source: §3.5 · p. 18
  • The archive correction measures the event- and state-dependent variance cost of preserving issued margins, and using refit variance at archived estimates can cause limiting undercoverage.The comparisons hold the probability estimate fixed to isolate covariance omission from changes in the fitted forecast.Source: §3.5 · p. 18

3.6 Numerical implementation

  • Implementation fits the vine to dated transforms, forms covariance from stored normal scores and lag rows, then evaluates the event gradient at terminal margins.The resulting standard error supports either probability-scale or logit intervals.Source: §3.6 · p. 18
  • The covariance calculation requires derivatives with respect to dated marginal inputs and terminal forecast parameters, and permits analytical or numerical derivatives.These derivative choices must satisfy the stated accuracy conditions.Source: §3.6 · pp. 18–19
  • A local numerical solution needs an average score residual and probability error of op(T^-1/2), plus consistency or a vanishing stagewise objective gap to select the correct stationary branch.A small residual alone is insufficient; the resulting gradient and covariance errors must be op(1).Source: §3.6 · p. 19
  • A forward-difference step sT = T^-1/4 is justified when probability-evaluation error is op(T^-1/2), with evaluation bounds uniform on compact lag-state sets and a neighborhood of the true parameters.Stationarity permits evaluation at the random terminal state.Source: §3.6 · p. 19

4 Monte Carlo Simulations

The simulations assess interval coverage for conditional event probabilities at each sample’s realized terminal state, comparing Wald and logit intervals across sample sizes and dependence settings. Logit intervals improve lower-tail coverage in the shorter histories studied, while coverage remains sensitive to sample size, persistence, and event.

4.1 Simulation design

  • The design uses a stable VAR(1) with normal margins and a fixed, correctly specified Gaussian or positive Clayton D-vine, under baseline and persistent dynamics.The coefficient matrices have spectral radii 0.45 and 0.85; sample sizes include T = 200–800 and larger benchmarks at 2000 and 8000.Source: §4.1 · pp. 19–20

4.2 Simulation results

  • Across T = 200–800, logit coverage ranges from 93.3% to 96.1%, versus 85.8–93.0% for Wald, with paired gains of 1.6–9.2 percentage points.The largest reported mean-length increase is 0.0032 in the baseline Gaussian design at T = 200, from 0.0402 to 0.0434.Source: Table 1 · p. 21, §4.2 · p. 21
  • At T = 8000, lower-tail Wald coverage ranges from 94.1% to 95.4%, serving as a large-sample benchmark beyond application histories.The benchmark histories are much longer than the 327–776 observations available in the application.Source: §4.2 · p. 21
  • The results support the asymptotic approximation for sufficiently long histories, but establish neither an accurate-coverage threshold nor comparable accuracy at shorter application-sized histories.Finite-sample coverage depends on sample size, persistence, and event; the transformation also slightly lowers coverage for some other events.Source: §4.2 · pp. 21–22
  • At T = 8000, omitting terminal-margin gradients lowers Wald coverage by 33.6–48.6 points, versus 6.7–17.3 for terminal lag gradients and 0.1–0.9 for the additional historical term.The authors connect the importance of terminal slope uncertainty to its role in the forecast gradient despite cancellation of centered historical lag terms.Source: §4.2 · p. 22

5 Real-time forecasting of joint contractions

The section reconstructs real-time forecasts of joint contractions from locked release histories, retaining issued marginal transforms for dependence estimation and comparing archived forecasts with refits. It examines forecast losses and model-based parameter uncertainty, including limitations of the stationary marginal specification.

5.1 The release target and information set

  • Forecasts target the first common release, whose event can differ from the subsequently revised outcome; evaluation covers 390 primary forecasts and a separate 60-month stress period.The release target is constructed from the common employment, production, and unemployment data available at that release.Source: §5.1 · p. 23

5.2 Forecast construction and predictive performance

  • For the primary labor event, archived-Gaussian Brier loss is 0.0882 versus 0.0872 for the locked-history refit, 0.0834 for independence, and 0.0832 for the latest-vintage refit.For employment–production contraction, independence has Brier loss 0.0747 versus 0.0819 for the archive and 0.0777 for the latest-vintage refit. These descriptive comparisons provide little support for an archive advantage in the primary period; no significance ranking is asserted.Source: Table 2 · p. 25, §5.2 · p. 25
  • Positive Clayton dependence lowers loss relative to the archived Gaussian fit for all three primary events, but independence has lower loss than either archive.The study retains issued PITs to train the dependence model on simultaneous marginal forecast surprises, while updating margins with each new release.Source: §5.2 · pp. 24–25

5.3 Sources of estimation uncertainty

  • The uncertainty intervals measure parameter uncertainty under the maintained continuous model, but their coverage under the observed release process is not established.Before 2020, standardized employment errors have standard deviation 0.563; production errors have lag-one correlation −0.304 and squared-error lag-one correlation 0.404, while unemployment changes have ties.Source: §5.3 · p. 25
  • Dropping only terminal slopes leaves 93.0% and 87.2% of complete median standard errors, while dropping all terminal-margin gradients leaves 37.7% and 32.1%.The reported interpretation is that current marginal estimation and its covariance with the dependence fit account for most of the adjustment in these real-time forecasts.Source: Table 3 · p. 26, §5.3 · p. 26
  • Removing the extra archive term changes exclusion of the 0.20 reference probability for a 95% symmetric labor interval on one of 390 primary dates and for two employment–production dates.These fixed reference probabilities illustrate interval-exclusion sensitivity, not a policy threshold.Source: §5.3 · p. 26

5.4 Calibration, revisions, and the stress period

  • In the primary period, labor had 47 events at frequency 0.1205 versus an average archived Gaussian forecast of 0.1996; employment–production had frequency 0.1179.Source: §5.4 · p. 26
  • The employment–production gap interval’s upper endpoint is −0.0014, detecting narrow aggregate overprediction, while the simultaneous labor-gap interval includes zero.The bounds are simultaneous at 95% across eight event–period pairs and concern fixed-period averages, not individual forecast origins.Source: §5.4 · p. 27
  • In the stress period, labor occurred twice against an average archived forecast of 0.3315, and aggregate overprediction was detected.The stress-period Gaussian labor interval had median width 0.989, one numerically degenerate date, and Clayton estimates at the parameter boundary for 56 of 60 origins, so interior-theory intervals were withheld.Source: §5.4 · p. 27
  • September 2026 vintages reclassified 21 primary labor dates, adding nine events and removing twelve, and 25 employment–production dates despite a net count change from 46 to 45.The release target must be fixed before comparing forecasts because latest-vintage histories cannot reproduce information available at the original forecast dates.Source: §5.4 · p. 27
  • Within the working law, current marginal estimation contributes much more uncertainty than the extra archive term, but release diagnostics reveal limitations it does not resolve.Valid inference at an individual origin under an adapting marginal model would require a new first-stage analysis.Source: §5.4 · p. 27

6 Conclusion

The conclusion presents joint inference that accounts for archived marginal fits and terminal forecast estimates, then compares archived transforms with terminal refitting and discusses the stationary model’s limits.

6 Conclusion

  • The historical regression correction reduces to harmonically weighted innovation moments, while the terminal forecast retains uncertainty in every coefficient.Their joint influence yields a feasible covariance estimator and repeated-sample intervals for conditional event probabilities at a random forecast state.Source: §6 · p. 28
  • In simulations and the real-time application, the variance cost of retaining issued transforms is smaller than the contribution of current marginal estimation.The Gaussian refit comparison assigns a positive semidefinite covariance term to retaining issued transforms.Source: §6 · p. 28
  • Logit intervals improve lower-tail coverage, while coverage shortfalls remain at shorter histories relevant to the application.The larger-sample results support the asymptotic approximation, but do not establish comparable accuracy at shorter histories.Source: §6 · pp. 28–29
  • Changes in marginal scales and serial dependence in issued errors limit the stationary benchmark even when its estimation uncertainty is fully accounted for.Allowing volatility to adapt requires deriving the historical correction and terminal influence for that marginal estimator.Source: §6 · p. 29
  • Discrete release distributions and explicit revision models pose related inference problems.Inference must follow both the forecasts used to construct the historical record and the estimates used at the current origin.Source: §6 · p. 29

A Joint inference for the copula and terminal VAR estimates

The joint limit combines historical score corrections with terminal regression errors on the same probability space. A finite-lag approximation supports the joint central limit argument.

A Joint inference for the copula and terminal VAR estimates

  • The joint limit requires expanding historical score corrections and terminal regression errors on the same probability space.The proof first derives the recursive score expansion and then calculates covariance with terminal influences.Source: A Joint inference for the copula and… · p. 30
  • A finite-lag approximation establishes the joint central limit theorem, with prefix bounds and moment-consistency results supplied in the online appendix.The derivation includes cancellation of the centered lag design.Source: A Joint inference for the copula and… · p. 30
  • The coefficients hT,t encode the repeated use of observations in the dated regression fits.Those repeated uses are part of the historical correction to the joint influence.Source: A Joint inference for the copula and… · p. 30

A.1 Recursive score expansion

The recursive score expansion controls estimation error accumulated across historical fits. Centered lag-design cancellation leaves a weighted innovation-score sum, yielding the sequential estimator’s limit and supporting covariance estimation.

A.1 Recursive score expansion

  • Aggregating the historical location error makes the centered lag-design component negligible, leaving innovation moments in the historical correction.The lag row and prefix regression error both depend on past innovations, so the replacement is justified after aggregation rather than term by term.Source: §A.1 · p. 30
  • The centered lag-design cancellation is needed because the design is correlated with accumulated regression error; the stable filter separates dependence into a current-innovation term and a telescoping difference.This decomposition underlies the aggregation argument that leaves innovation moments in the correction.Source: §A.1 · p. 30
  • Under the stated model and estimator conditions, the sequential estimator is consistent and its leading score term converges to a normal limit.The resulting covariance consistently estimates the sequential covariance block, and fixed contrasts with positive limiting variance have asymptotically standard normal studentized statistics.Source: §A.1 · pp. 32, 35
  • The historical score expansion yields a weighted innovation-score sum with harmonic weights hT,s.The result identifies the leading term after the generated-score remainder is controlled.Source: §A.1 · pp. 32, 34, 35

A.2 Terminal influences and the joint array

Terminal coefficient and scale influences are derived from the OLS equations and combined with the historical score array. The resulting joint array has a normal limit under the theorem’s conditions.

A.2 Terminal influences and the joint array

  • Terminal OLS equations and the minimized residual sum of squares give coefficient and scale influences, whose covariance with the innovation score uses the intercept identity.Mixed third moments are retained because normal coordinate margins do not eliminate them under asymmetric copula dependence.Source: §A.2 · p. 36
  • The terminal influence covariance includes every coefficient–variance cross block without assuming sign symmetry of the copula.Predictability of the lag design and the intercept identity determine its covariance with the innovation score.Source: §A.2 · pp. 36–37
  • The joint array covariance converges to (3.5), with harmonic identities giving coefficient limits 1, 1, and 2.The burn-in error tends to zero under the stated condition.Source: §A.2 · pp. 36–37
  • The joint influence array converges to a normal distribution with covariance WVAR under the conditions of Theorem 3.1.Truncating the stable filter separates dependent coefficient influences into independent blocks while bounding the discarded part.Source: §A.2 · p. 37

A.3 Completion of the joint inference theorem

The theorem combines the joint limit for the fitted parameters with inference at the random terminal VAR state. Stability and temporal independence separate the leading estimation error from that state.

A.3 Completion of the joint inference theorem

  • The covariance estimator is consistent and samplewise positive semidefinite, with the fallback used with probability tending to zero.The proof establishes convergence of the additional block to C and verifies the positive semidefinite property of both covariance terms.Source: §A.3 · p. 40
  • The joint theorem covers fixed Gaussian and positive-parameter Clayton vines under the stated stable VAR assumptions, including estimated higher-tree ancestors for Clayton.Normal coordinate margins provide the moments; the innovation copula need not be Gaussian.Source: §A.3 · p. 40

B Inference for the conditional event probability at the terminal VAR state

  • The event forecast is evaluated at the last observed lag state, so parameter asymptotics are supplemented by a joint convergence argument.Localization on compact state sets and stationarity support the expansion and studentization.Source: B Inference for the conditional event… · p. 40

B.1 Separation from the terminal state

  • The leading parameter errors and terminal state have an independent joint limit, enabling inference at the realized forecast state.The result follows by deleting recent terms from the influence sum and approximating the terminal state using recent innovations.Source: §B.1 · pp. 40, 42

B.2 Proof of Theorem 3.2

The proof establishes studentized inference for a conditional event probability evaluated at the terminal state, then extends it to fixed rectangles. The arguments use localization to handle the random state and fitted gradients.

B.2 Proof of Theorem 3.2

  • The random terminal state is handled by bounding derivatives on compact sets and then letting the compact-set radius increase.Stationarity makes the terminal state tight, allowing the Taylor remainder bound to extend beyond each fixed compact set.Source: §B.2 · p. 42, §B.3 · p. 45
  • The conditional event probability admits a studentized normal limit at the random terminal state under the stated positive-variance condition.The proof uses a Taylor expansion and joint convergence, with localization to control the fitted gradient.Source: §B.2 · pp. 43–44

B.3 Fixed rectangles and mixed directions

  • For fixed rectangles, the corresponding studentized limit and coverage hold when the projected variance is almost surely positive.The result applies to the fixed rectangle covered by Lemma S2.2.Source: §B.3 · pp. 44–45

B.4 Nondegeneracy for one-sided events

For one-sided events, the paper establishes almost-sure positive archive variance in the verified Gaussian and positive Clayton families. It also justifies logit event intervals while limiting the claim to first-order validity.

B.4 Nondegeneracy for one-sided events

  • The archive-event variance is positive almost surely for one-sided events in both verified vine families, without requiring sign symmetry of the innovation copula.The result does not assert positivity at the mean state in the Clayton case or a uniform lower bound over states.Source: §B.4 · pp. 46–47

B.5 Intervals on the log-odds scale

  • Logit intervals have first-order validity at a random terminal state without a uniform probability bound away from zero and one.The argument localizes away from the probability boundaries and then lets that localization shrink.Source: §B.5 · pp. 47, 49
  • The interval result does not establish higher-order accuracy or improved finite-sample coverage.Source: §B.5 · p. 48

C Gaussian refitting and the archive variance cost

The Gaussian full-sample refit recomputes historical residuals on the archive's observations and derives their generated-margin correction using an exact residual identity. Its covariance estimator supports repeated-sample inference and differs from the archive covariance by a positive semidefinite term at common population parameters.

C Gaussian refitting and the archive variance cost

  • The Gaussian full-sample refit yields a positive semidefinite covariance difference relative to the archive at common population parameters.The comparison follows from comparing the joint influence of the refit and archive estimators.Source: C Gaussian refitting and the archive… · p. 49
  • The refit's exact residual cross-product identity includes the estimated intercept and lag coefficients and accounts for dependence between the lag design and past innovations.The identity avoids treating the lag design as independent of past innovations.Source: C Gaussian refitting and the archive… · pp. 49–50
  • The full-sample Gaussian covariance estimator is positive semidefinite on every sample and converges in probability to Wfs.The identity fallback is used in the estimator.Source: C Gaussian refitting and the archive… · p. 49
  • For a fixed rectangle, the resulting normal interval has repeated-sample limiting coverage 1 −α for each fixed 0 < α < 1.The event argument uses the joint limit of the full-sample estimator and terminal state.Source: C Gaussian refitting and the archive… · pp. 50, 52
Loading 2610.07589v1…