Bayes Estimation of GLARMA Models With Applications

Ana Julia Alves C\^amara; Guilherme Pumi

REVIEW 4 major objections 6 minor 17 references

Bayes Estimation of GLARMA Models With Applications

T0 review · 4 major / 6 minor · reviewed 2026-07-12 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A Bayesian MCMC procedure estimates GLARMA models reliably for positive, double-bounded and count series, extending inference beyond the exponential family.

desk verdict Solid, usable Bayesian MCMC pipeline for NB/Beta/Gamma GLARMA in nimble; Beta prior collapse is real but does not sink the rest of the work. read the letter →

arxiv 2607.03586 v1 pith:ZBCI5NQ4 submitted 2026-07-03 stat.ME stat.CO

Guilherme Pumi , Ana Julia Alves Câmara This is my paper · ORCID

classification stat.MEstat.CO MSC 62M1062F1562M2062G2062-08

keywords timeseriesanalysisregressionmodelsBayesiannon-gaussianGLARMAMCMCnimble

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Environmental and energy time series are often positive, confined to (0,1), or integer-valued, so Gaussian ARMA models produce impossible forecasts and transformations distort dependence. The paper supplies a single Bayesian MCMC pipeline, coded in nimble, that estimates the full parameter vector of a GLARMA model under Negative-Binomial, Beta or Gamma random components. Extensive Monte Carlo experiments show that point and interval estimates remain accurate even when the autoregressive coefficient sits near the stationarity boundary, while two real applications (U.S. nuclear generation and Brazilian COPD admissions) demonstrate that the same code recovers seasonal harmonics, pollution effects and serial dependence with usable posterior uncertainty. A sympathetic reader therefore obtains a practical, unified tool for non-Gaussian time series that previously required either fragile frequentist optimizers or ad-hoc transformations.

What carries the argument

The observation-driven GLARMA recursion W_t = X_t'β + Z_t, with Z_t an ARMA filter of generalized residuals, together with the resulting conditional likelihood sampled by MCMC under independent normal or uniform priors (implemented in nimble).

What would settle it

Generate 50 independent Beta-GLARMA(1,0) series of length 1000 with true φ = 0.6, fit each under a non-informative U(-0.9,0.9) prior, and check whether the median of the posterior medians of φ remains near 0.6 with coefficient of variation below 10 percent; the paper’s Table 3 predicts catastrophic collapse, so any recovery would falsify the claimed necessity of informative priors.

Watch

Extended reading notes

Core claim

A unified MCMC-based Bayesian procedure recovers the parameters of general GLARMA models accurately and stably for Negative-Binomial, Beta and Gamma responses in finite samples of size 1000, including high-persistence regimes near the unit circle, thereby extending Bayesian inference beyond exponential-family GLARMA and furnishing a practical tool for non-Gaussian environmental and energy series.

Load-bearing premise

For Beta-GLARMA series the autoregressive coefficient is well enough identified by the data alone that a non-informative uniform prior on (-0.9,0.9) can be used; the paper’s own sensitivity table shows this prior collapses the estimate to zero.

Editorial extensions

If this is right

Practitioners can obtain full posterior uncertainty for Gamma-, Beta- and NB-GLARMA models without relying on gradient-based optimizers that break near the stationarity boundary.
Positive energy series and double-bounded or count environmental-health series can be modelled on their natural scale, avoiding transformations that alter serial dependence.
Prior-sensitivity diagnostics for the AR coefficient become a routine part of Beta-GLARMA analysis.
Seasonal harmonics and high-dimensional pollution covariates can be included while still recovering the dispersion and persistence parameters.
Near-unit-root persistence no longer produces routine numerical non-convergence.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same nimble skeleton could accommodate other non-exponential families (inverse-Gaussian, zero-inflated, etc.) with only a change of the random-component density.
Reversible-jump or spike-and-slab model selection, listed as future work, would replace the present residual-based heuristic and make the pipeline fully automatic.
Because non-informative priors fail for the Beta AR coefficient, agencies analysing double-bounded series will need historical or expert information; pure default priors are insufficient.
The nuclear-generation and COPD applications suggest immediate operational value for any forecasting group that already uses GLARMA but currently lacks uncertainty quantification.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The new piece is a single nimble MCMC setup that covers Negative-Binomial, Beta and Gamma GLARMA under the same residual recursion, plus two carefully diagnosed applications. Franco et al. already did the count case; this is the natural extension to the continuous families that environmental people actually use. That is useful, not revolutionary.

What works: the Monte Carlo design is transparent (n=1000, 50 reps, four persistence levels including |φ|=0.8). NB and Gamma recover cleanly under weak priors; intercepts get noisier near the unit circle but slopes and φ stay usable and chains do not break. The nuclear-energy Gamma fit (harmonics + AR(12)) and the COPD NB fit (PCA pollutants + seasonality) both pass residual checks that most applied papers skip. Diagnostics (Geweke, Heidel, ESS, Domínguez–Lobato, quantile residuals) are done properly. Citations look honest; self-cites are background, not load-bearing.

Soft spots, in proportion. The Beta case is the real one: Table 3 shows a flat U(−0.9,0.9) prior collapses φ̂ to ~0 with CVs of hundreds of percent; only priors whose support already contains the truth work. Because ε_t depends on ν, φ and ν are entangled, so the data alone do not identify the AR coefficient. The abstract’s “reliable \ldots across all settings” therefore over-reaches for Beta. Model selection is purely heuristic (Box–Jenkins on Pearson residuals). No code or processed data is released. These are genuine limits, not fatal ones; the NB and Gamma pipelines still stand.

Who it is for: applied people who already run GLARMA on counts, rates or positive series and want posterior intervals without writing their own sampler. A methods reader looking for new theory or asymptotic results will not find them. I would send it to a serious referee; the core MCMC contribution is real and the over-claim is easy to fix with a more precise abstract. Worth engaging if you work in this corner.

Referee Report

4 major / 6 minor

Summary. The paper proposes a Bayesian MCMC estimation framework (implemented in nimble) for GLARMA models whose random component need not belong to the exponential family, covering Negative Binomial (counts), Beta (double-bounded), and Gamma (positive continuous) responses. Priors are discussed, finite-sample performance is studied via Monte Carlo (n=1000, 50 replications, four AR(1) persistence levels including near-unit-root regimes), and two applications are given: U.S. nuclear net generation (Gamma-GLARMA with harmonics) and COPD admissions in Belo Horizonte (NB-GLARMA with pollution PCs). The abstract and conclusion claim reliable, stable estimation “across all settings.”

Significance. If the reliability claim holds, the work fills a genuine gap: Bayesian GLARMA has been largely restricted to counts (Franco et al., 2019), while environmental and energy series routinely require positive or unit-interval responses. A unified nimble implementation, transparent Monte Carlo design for NB and Gamma under weakly informative priors, and two substantive applications are concrete contributions. The authors also document prior sensitivity for Beta rather than hiding it. These strengths make the paper useful to applied researchers once the scope of the reliability claim is aligned with the evidence.

major comments (4)

Abstract and §6 claim “reliable and stable estimation across all settings,” but §4.2 and Table 3 show that for Beta-GLARMA a non-informative prior φ∼U(−0.9,0.9) collapses the posterior median of φ to ≈0 (0.005 / 0.003) with CVs of 470% and 1011%. Even moderately informative uniforms recover φ only when their support already contains the truth; misspecified uniforms pull the median to the prior mean. Because ε_t depends on ν, φ and ν are entangled, so the data alone do not identify the AR coefficient. The central claim must be restated to distinguish NB/Gamma (weak priors work) from Beta (informative, correctly centered priors are required), and the abstract/conclusion language must be revised accordingly.
§4 states that the simulation “assesses point and interval estimation,” yet Tables 1–4 report only median estimates and coefficients of variation. No coverage rates, average lengths, or calibration checks for HPD/quantile credible intervals appear for any of the three DGPs. Without these, the interval-estimation part of the claim is unsupported and should either be supplied (e.g., empirical coverage of 95% HPDs over the 50 replications) or removed from the abstract and §4 framing.
The Monte Carlo study (§4) is restricted to GLARMA(1,0). The paper motivates general GLARMA(p,q) and uses AR(12) and seasonal harmonics in the nuclear application, but never examines MA terms, higher-order AR, or joint (p,q) recovery under the same MCMC scheme. At minimum, a short additional experiment with a non-zero MA coefficient (or a brief justification why MA is “inherently well-behaved” with supporting numbers) is needed before claiming a general GLARMA framework.
There is no comparison, even on a subset of replications, with conditional MLE (the standard frequentist estimator for GLARMA). Given that the paper’s selling point is numerical stability near the stationarity boundary where gradient-based MLE is known to struggle (Dunsmuir, 2016), a head-to-head on bias, RMSE, and failure rate for φ near ±0.8 would substantially strengthen the contribution; its absence leaves the practical advantage of the Bayesian route asserted rather than demonstrated.

minor comments (6)

The opening of §1 repeats essentially the same paragraph on environmental/energy series and the Grande et al. (2022) influenza example twice in succession; one copy should be deleted.
Notation: the residual scaling λ∈(0,1] is introduced in (1) but never varied or discussed in the simulations or applications; either fix λ (and state the choice) or drop it from the general definition if only Pearson residuals (λ=0.5) are used.
Table 2 (Beta, informative priors): ν is estimated near 3.00 when |φ|=0.8 despite true ν=4 and prior U(3,7); the text should note that the posterior piles up at the lower prior boundary, which is itself a warning about prior influence.
§5.1 residual MDS test reports p-values 0.050 and 0.056; at the stated 5% level these are borderline. A short remark on sensitivity to lag choice or bootstrap seed would help.
References: Maia & Franco (2026) and Prass & Pumi (2026) are cited with future years; confirm status or use “forthcoming” consistently.
Code/data availability is not stated. Given the nimble implementation is a selling point, a repository link (or supplementary scripts) would aid reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No derivation circularity: standard likelihood-plus-prior MCMC; self-citations are background only; Beta prior sensitivity is a robustness issue, not a definitional loop.

full rationale

The paper proposes a Bayesian MCMC procedure for GLARMA models (likelihood (2) times priors (4)–(5), sampled in nimble). All simulation and empirical results are ordinary posterior summaries under that procedure. No quantity is defined from a fitted parameter and then re-used as a “prediction,” no uniqueness theorem is imported from the authors’ prior work to force the model form, and no ansatz is smuggled in via self-citation. Self-citations (Camara et al. 2025a,b; Grande et al. 2022, 2025; Franco et al. 2019) supply applications or related models and are not load-bearing for any derivation step. The only soft point is the Beta-GLARMA prior sensitivity (Table 3): a non-informative prior collapses φ, so reliable recovery requires an informative prior whose support already contains the truth. That is a genuine limitation of the claim “reliable \ldots across all settings,” but it is an empirical/robustness failure, not circularity by construction. Score 1 reflects the absence of definitional loops plus the minor, non-load-bearing self-citations.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the classical GLARMA recursion, standard conditional-likelihood factorisation, and conventional weakly informative priors; no new mathematical objects are postulated. Free parameters are the usual hyper-parameters of those priors and the residual scaling λ, all chosen by the analyst rather than estimated from first principles.

free parameters (3)

prior variances σ²_β, σ²_φ, σ²_θ
Set to large constants (or 1) to obtain weakly informative normals; their precise values affect posterior concentration near stationarity boundaries.
uniform bounds for ν / r and for φ
Chosen by hand (e.g. U(0,50), U(-1,1), or tighter intervals for Beta); the Beta sensitivity study shows that the choice is load-bearing.
residual scaling λ ∈ (0,1]
Fixed at 0.5 (Pearson) or 1 (score) without estimation; alters the filter recursion and therefore the likelihood surface.

assumptions (3)

domain assumption The observation-driven recursion (1) with initial conditions Z_t=ε_t=0 for t≤0 yields a well-defined conditional likelihood that can be evaluated for any (β,φ,θ,ν).
Standard GLARMA construction (Davis et al. 1999, 2003); invoked throughout Sections 2–3.
domain assumption Independent normal or uniform priors on the regression and AR/MA coefficients are compatible with the link function and produce proper posteriors.
Stated in Section 3; used for all MCMC runs.
standard math Pearson residuals form a martingale-difference sequence under correct specification, so Domínguez–Lobato and Ljung–Box tests are valid residual diagnostics.
Used in both empirical applications (Sections 5.1–5.2).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Bayes Estimation of GLARMA Models With Applications." pith.science (2026). https://pith.science/paper/ZBCI5NQ4

@misc{pith2026260703586,
  author       = {Pith},
  title        = {Pith review of: Bayes Estimation of GLARMA Models With Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZBCI5NQ4}},
  note         = {Machine review of arXiv:2607.03586}
}

read the original abstract

This work presents a Bayesian approach for parameter estimation in the class of Generalized Linear Autoregressive Moving Average (GLARMA) models, extending the methodology beyond the common exponential family setting. The proposed framework accommodates positive, double-bounded, and count time series through a unified MCMC-based estimation procedure implemented in \texttt{nimble}. We discuss prior specifications for the model parameters and conduct an extensive Monte Carlo simulation study to evaluate the finite-sample performance of the approach under three distinct data-generating mechanisms: Negative Binomial-GLARMA for count data, Beta-GLARMA for double-bounded outcomes, and Gamma-GLARMA for positive continuous time series. The simulation study assesses point and interval estimation, prior sensitivity, and the behaviour of the estimators under varying levels of temporal persistence, including challenging scenarios near the boundaries of the stationarity region. The practical utility of the proposed framework is illustrated through two empirical applications: analysing monthly net electricity generation by nuclear plants in the United States using a Gamma-GLARMA model with harmonic seasonal components; and modelling monthly hospital admissions due to chronic obstructive pulmonary disease in Belo Horizonte, Brazil, using a Negative Binomial-GLARMA model with principal components derived from air pollution covariates. The results demonstrate that the proposed Bayesian framework provides reliable and stable estimation across all settings, offering a flexible and practical tool for analysing non-Gaussian time series in a wide range of applications.

Figures

Figures reproduced from arXiv: 2607.03586 by the authors.

**Figure 2.** Time series of monthly COPD cases 5.3.1 Dimensionality Reduction: Principal Component Analysis Initial exploratory analysis revealed significant correlation among air pollutants (PM10, NO, NO2, CO, NOX, and O3), suggesting that a direct regression approach would suffer from multicollinearity. To overcome this, Principal Component Analysis (PCA) was applied to the standardized pollutants [PITH_FULL_IMAGE:figures/ful… view at source ↗

**Figure 3.** Monthly concentrations of air pollutants in Belo Horizonte from January 2007 to [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Diagnostic plots for Pearson residuals. To overcome the inherent limitations of Pearson residuals in discrete settings, we implemented randomized quantile residuals. Once the original data consists of counts, these residuals map the discrete probabilities into a conti…

**Figure 5.** Figure 5: Diagnostic plots for Quantile residuals. [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 3 canonical work pages

[1]

Grande, A. F. and Pumi, G. and Cybis, G. B. , title =. Journal of Statistical Computation and Simulation , year =
[2]

2022 , author =

Granger causality and time series regression for modeling the migratory dynamics of. 2022 , author =

2022
[3]

Statistical Inference for Stochastic Processes , author =

Memory properties of transformations of linear processes , volume =. Statistical Inference for Stochastic Processes , author =. 2017 , pages =. doi:10.1007/s11203-016-9134-4 , number =

work page doi:10.1007/s11203-016-9134-4 2017
[4]

and Dunsmuir, William T

Davis, Richard A. and Dunsmuir, William T. M. and Wang, Ying , title =. Asymptotics, Nonparametrics, and Time Series , editor =. 1999 , publisher =

1999
[5]

and Dunsmuir, William T

Davis, Richard A. and Dunsmuir, William T. M. and Streett, Sarah B. , title =. Biometrika , volume =. 2003 , publisher =

2003
[6]

Dunsmuir, William T. M. , title =. Handbook of Discrete-Valued Time Series , editor =. 2016 , pages =

2016
[7]

Maia, G. O. and Franco, G. C. , title =. Environmental and Ecological Statistics , year =. doi:10.1007/s10651-025-00692-4 , url =

work page doi:10.1007/s10651-025-00692-4
[8]

Maia, G. O. and Franco, G. C. , title =. Journal of Statistical Computation and Simulation , volume =. 2025 , doi =

2025

Show all 17 references

[9]

and Migon, Helio S

Franco, Glaura C. and Migon, Helio S. and Prates, Marcos O. , title =. Brazilian Journal of Probability and Statistics , volume =. 2019 , month =. doi:10.1214/19-BJPS437 , url =

2019 doi
[10]

Combining Generalized Linear Autoregressive Moving Average and Bootstrap Models for Analyzing Time Series of Respiratory Diseases and Air Pollutants , journal =

Camara, Ana J. Combining Generalized Linear Autoregressive Moving Average and Bootstrap Models for Analyzing Time Series of Respiratory Diseases and Air Pollutants , journal =. 2025 , publisher =

2025
[11]

Applied Mathematical Modelling , volume =

Camara, Ana Julia Alves and Reisen, Valderio Anselmo and Bondon, Pascal , title =. Applied Mathematical Modelling , volume =. 2025 , doi =

2025
[12]

Gil-Alana and James E

Carlos Pestana Barros and Luis A. Gil-Alana and James E. Payne , title =. Energy Systems , volume =. 2013 , doi =

2013
[13]

Taiane Schaedler Prass and Guilherme Pumi , year =
[14]

Dom\'inguez and Ignacio N

Manuel A. Dom\'inguez and Ignacio N. Lobato , title =. Econometric Reviews , volume =
[15]

CODA: Convergence Diagnosis and Output Analysis for

Martyn Plummer and Nicky Best and Kate Cowles and Karen Vines , journal =. CODA: Convergence Diagnosis and Output Analysis for. 2006 , volume =

2006
[16]

2024 , note =

TeachingDemos: Demonstrations for Teaching and Learning , author =. 2024 , note =

2024
[17]

and Pederson, Gregory T

Peitzsch, Erich H. and Pederson, Gregory T. and Birkeland, Karl W. and Hendrikx, Jordy and Fagre, Daniel B. , title =. Scientific Reports , volume =. 2021 , doi =

2021

Pith tools

Reviewed July 12, 2026 · model on record in the stance chip above.

[1] [1]

Grande, A. F. and Pumi, G. and Cybis, G. B. , title =. Journal of Statistical Computation and Simulation , year =

[2] [2]

2022 , author =

Granger causality and time series regression for modeling the migratory dynamics of. 2022 , author =

2022

[3] [3]

Statistical Inference for Stochastic Processes , author =

Memory properties of transformations of linear processes , volume =. Statistical Inference for Stochastic Processes , author =. 2017 , pages =. doi:10.1007/s11203-016-9134-4 , number =

work page doi:10.1007/s11203-016-9134-4 2017

[4] [4]

and Dunsmuir, William T

Davis, Richard A. and Dunsmuir, William T. M. and Wang, Ying , title =. Asymptotics, Nonparametrics, and Time Series , editor =. 1999 , publisher =

1999

[5] [5]

and Dunsmuir, William T

Davis, Richard A. and Dunsmuir, William T. M. and Streett, Sarah B. , title =. Biometrika , volume =. 2003 , publisher =

2003

[6] [6]

Dunsmuir, William T. M. , title =. Handbook of Discrete-Valued Time Series , editor =. 2016 , pages =

2016

[7] [7]

Maia, G. O. and Franco, G. C. , title =. Environmental and Ecological Statistics , year =. doi:10.1007/s10651-025-00692-4 , url =

work page doi:10.1007/s10651-025-00692-4

[8] [8]

Maia, G. O. and Franco, G. C. , title =. Journal of Statistical Computation and Simulation , volume =. 2025 , doi =

2025