Bayesian Seasonal Adjustment for Survey Time Series

Siu-Ming Tam

This paper claims that seasonal adjustment should treat published survey standard errors as a first-class input, and that doing so can collapse the estimated stochastic seasonal variance to zero.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 18:38 UTC pith:GUYNLEKM

load-bearing objection A genuinely useful state-space method for incorporating survey SEs into seasonal adjustment, but the X-11-equivalence anchor is not established and the paper overstates its simulation.

arxiv 2607.17226 v1 pith:GUYNLEKM submitted 2026-07-19 stat.ME

Bayesian Seasonal Adjustment for Survey Time Series

Siu-Ming Tam This is my paper

classification stat.ME MSC 62M1062F1562D05

keywords seasonal adjustmentBasic Structural Modelsurvey sampling errorKalman filterGibbs samplercredible intervalX-11state space model

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

National statistical offices seasonally adjust survey series with X-11-style filters that treat each monthly estimate as exact, throwing away the standard errors that come with it. This paper embeds those time-varying sampling variances into a Basic Structural Model, so the Kalman filter automatically down-weights high-variance months. The paper shows that with zero measurement variance the model reduces to an X-11-equivalent benchmark, making the comparison clean. In simulations, credible intervals that use the survey variances achieve near-nominal coverage where the X-11-equivalent under-covers badly. On 120 months of a large and a small labour-force domain, the maximum-likelihood estimate of stochastic seasonal variance collapses to zero once sampling error is modelled, suggesting the apparent seasonal fluctuation is largely measurement noise.

Core claim

The central claim is that a Basic Structural Model whose observation equation carries the design-based sampling variance V_t^d — rather than zero — is a principled Bayesian generalisation of X-11, and that the extra uncertainty quantification this provides is large. The paper proves the generalisation through the Harvey–Todd equivalence chain: BSM with V_t=0 has the airline-model reduced form, which Cleveland–Tiao showed is asymptotically X-11. It then supplies a two-block Gibbs sampler (forward-filter backward-sampler for the state trajectory, conjugate inverse-gamma draws for variances) that yields exact joint smoothing posteriors, so trend-level credible intervals, k-step trend-movement c

What carries the argument

The carrying mechanism is the time-varying measurement variance V_t^d placed in the Kalman filter observation equation, together with the two-block Gibbs sampler: a Forward Filter Backward Sampler (FFBS) draws the full 12-dimensional BSM state trajectory jointly from the smoothing posterior, and conjugate inverse-gamma updates draw the variance hyperparameters conditional on that trajectory. The Kalman gain automatically reduces weight on high-CV months, and the joint trajectory draws make any k-step movement CBI a simple quantile computation.

Load-bearing premise

The benchmark equivalence between BSM(V_t=0) and X-11 rests on the Harvey–Todd airline reduced form, but the BSM used here drops the observation irregular term, and the paper does not derive the constrained reduced form that follows; if the constrained model is not X-11-equivalent, the benchmark and the coverage comparison change meaning.

What would settle it

Derive the reduced form of the BSM with no observation irregular and compare its autocovariance and ARIMA representation with the airline model; if they differ, or if on a series generated from an airline model the X-11 filter and BSM(V_t=0) signal extraction diverge by more than simulation noise, the equivalence claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

National statistical offices could publish exact credible intervals for trend, trend movements, and posterior probabilities of directional change, instead of point-only X-11 output.
Series with high-CV months will receive wider, honest intervals, with the filter relying more on the structural forecast in precisely those months.
The empirical collapse of stochastic seasonal variance suggests some published seasonal patterns may be artefacts of sampling noise, making deterministic seasonality a viable alternative.
The machinery extends to multivariate variables via the full sampling covariance matrix, enabling joint inference across multiple labour-force characteristics.
The coverage advantage is largest for small domains, where the X-11-equivalent intervals are most misleading.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the constrained-BSM reduced form is not genuinely airline-equivalent, the claimed X-11 anchor weakens; a direct check of the constrained reduced-form autocovariance structure would settle this.
The under-coverage in small domains (86.7%/82.1%) suggests a heavier-tailed prior on the trend-variance hyperparameter or Rao-Blackwellisation might close the gap; the paper flags this itself.
The zero seasonal-variance result, if it replicates across other survey series, implies a practical guideline: test for deterministic seasonality before applying X-11, or report seasonality as model-selected.
The one-step backward conditional could support real-time filtered publication, giving statistical offices an online movement significance test.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

A genuinely useful state-space method for incorporating survey SEs into seasonal adjustment, but the X-11-equivalence anchor is not established and the paper overstates its simulation.

read the letter

Here's the short version. The method itself — a BSM with time-varying known sampling variances, sampled with a two-block FFBS Gibbs — is coherent and useful for NSOs that want uncertainty intervals around seasonally adjusted trends. The Gibbs full conditionals are standard, the simulation design with within-replication parameter estimation is honest in intent, and the empirical illustration on ABS data is a nice demonstration. Give credit for that.

The soft spot is load-bearing. The paper claims BSM(V_t=0) is equivalent to X-11 via Harvey–Todd/Cleveland–Tiao, but the BSM in Section 2.3 has no observation irregular. The standard reduced-form airline result requires that irregular term. The paper's footnote 1 half-admits this and asserts a 'constrained airline model with sigma_epsilon^2=0' without derivation. The stress-test note is right: for the model as written, (1-B)(1-B^12) on y_t = L_t + gamma_t gives nonzero autocov at lag 2 and zero at lag 13, which is not the airline MA(1,12). So the benchmark in Table 1, labeled 'X-11 equivalent', is not what it claims to be. That doesn't sink the core idea — you can still compare DMM-BSM against BSM(V_t=0) as a no-measurement-error baseline — but it punctures the 'principled generalisation of X-11' framing and the abstract's headline claim.

Other soft spots are smaller. The abstract promises 'exact credible intervals', but the paper's own Table 1 shows 86.7% and 82.1% coverage for ACT. The paper does discuss this, but the word 'exact' should be dropped. The MLE-boundary finding is interesting, but the abstract interprets it as 'measurement noise rather than genuine seasonal drift' — a zero stochastic seasonal variance is really saying the seasonal component is deterministic, and the paper's own Remark 4 is more careful. No code or data are provided, so the simulation is not independently reproducible.

Who is this for? People working on seasonal adjustment of survey series or state-space methods for official statistics. It deserves a serious referee, because the method is potentially practical and the flaws are fixable — redefine the benchmark, add the irregular or drop the X-11 equivalence claim, tone down the coverage language. I wouldn't cite it in its current form.

Circularity Check

0 steps flagged

No load-bearing circularity: the posterior, Gibbs inference, coverage simulation, and MLE-boundary finding are genuine outputs; only a minor non-load-bearing self-citation to Tam (2026) appears.

full rationale

The central derivation is not circular. The posterior intervals and movement probabilities are computed from the specified linear-Gaussian BSM likelihood, the published sampling variances, and declared priors via a valid two-block FFBS Gibbs sampler; no inferential output is defined in terms of the quantity it is supposed to estimate. The coverage simulation re-estimates hyperparameters within each replication, so the reported coverage is a genuine frequentist property of the posterior intervals, not an algebraic identity. The MLE boundary sigma_omega = 0 is a fitted parameter and is presented as an empirical diagnostic, not as a prediction of the model. The only self-citation is Tam (2026) for the inherited DMM/SHBU components, but those components are not used in the empirical illustration: the paper explicitly reduces to the direct-estimate limiting case because stratum-level SHBU outputs are unavailable, and the novel BSM-with-sampling-variance contribution does not rest on their validity. The possible weakness identified in Section 3 and footnote 1, namely that the BSM omits the observation irregular while claiming the Harvey–Todd airline equivalence, is an unsupported or potentially incorrect external premise rather than a circular reduction: the paper does not define X-11 equivalence in terms of its own fitted outputs, nor does it fit a parameter to the quantity it then 'predicts.' Thus there is no load-bearing circularity; at most a minor non-load-bearing self-citation, captured by the score of 2.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The model introduces no new physical or structural entities. Its free parameters are the usual state-space variance hyperparameters, plus the hand-set inverse-gamma prior scale. The load-bearing background facts are the Harvey-Todd/Cleveland-Tiao equivalence chain and the treatment of survey variances as known and independent; the former is applied outside its standard domain, which is the main technical risk.

free parameters (3)

sigma_xi (trend innovation std) = AU/ACT MLE ~1,825 persons; posterior mean 1,491 (SD 210)
Estimated via MLE/Gibbs; controls trend smoothness and is central to the coverage simulation and to the posterior intervals.
sigma_omega (seasonal innovation std) = MLE ~0; posterior mean 48 (SD 14)
Estimated via MLE/Gibbs; the MLE collapse is the paper's headline empirical finding, and the posterior value is partly prior-influenced (Remark 4).
IG(0.01,0.01) prior hyperparameters on variances = a=b=0.01
Chosen by hand as diffuse priors; the paper admits the posterior mean of sigma_omega is partially influenced by this choice because the prior has support only on (0,infinity).

axioms (4)

standard math Harvey-Todd reduced form of BSM with local level and stochastic seasonal is airline ARIMA(0,1,1)(0,1,1)_12; Cleveland-Tiao gives X-11 asymptotic equivalence.
Invoked in Section 3 and the abstract; however, the paper applies it to a BSM without the observation irregular, where the equivalence is not established.
standard math Kalman filter and FFBS produce exact joint smoothing draws for linear Gaussian state-space models.
Section 4.4 relies on Carter-Kohn and Fruhwirth-Schnatter; this standard result is used for the state trajectory draws.
domain assumption Survey sampling variances V_t are known and independent across waves.
Section 2.1 and Remark 2: the model treats V_t as known and sets inter-wave sampling covariances to zero, acknowledging that this makes movement intervals conservative.
domain assumption In the coverage simulation, the true DGP is the same BSM with the observed SE_t sequence; coverage is evaluated under the model.
Section 6.5 design: this tests parameter-estimation uncertainty but not model misspecification, so the coverage numbers do not validate the model against real-world departures.

pith-pipeline@v1.3.0-alltime-deepseek · 19171 in / 16849 out tokens · 160530 ms · 2026-08-01T18:38:59.715954+00:00 · methodology

0 comments

read the original abstract

Seasonal adjustment procedures used by national statistical offices -- X-11 and X-12-ARIMA -- treat each survey estimate as an exact observation, discarding the accompanying standard errors that survey methodologists routinely compute. This paper closes that gap by embedding time-varying sampling error variances into a Basic Structural Model (BSM), extending a recently proposed Dynamic Mini-Max (DMM) Bayesian framework for survey estimation. Via the Harvey-Todd equivalence, BSM with zero measurement error variance reduces to X-11-style seasonal adjustment, so DMM-BSM is a principled Bayesian generalisation of existing practice rather than a departure from it. A two-block Gibbs sampler delivers the full joint smoothing posterior of the latent state trajectory. Exact credible intervals for the trend level, k-step trend movements (k=1,2,...), and seasonally adjusted estimates follow directly, together with posterior probabilities of directional change -- outputs that X-11 cannot provide. Simulation studies with within-replication parameter estimation confirm substantially higher credible interval coverage than the X-11-equivalent model across both large and small survey domains. Applied to 120 months of Australian Bureau of Statistics Labour Force Survey data, once sampling variance is modelled the maximum likelihood estimate of stochastic seasonal variance collapses to zero -- evidence that apparent seasonal fluctuations in the published series are largely attributable to measurement noise rather than genuine seasonal drift, a distinction X-11 cannot make.

Figures

Figures reproduced from arXiv: 2607.17226 by Siu-Ming Tam.

**Figure 3.** Figure 3: Empirical coverage of 95% credible intervals by month ( [PITH_FULL_IMAGE:figures/full_fig_p021_3.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

15 extracted references · 1 linked inside Pith

[1]

Bell, W. R. and Hillmer, S. C. (1984). Issues involved with the seasonal adjustment of economic time series. Journal of Business & Economic Statistics, 2(4), 291–320

1984
[2]

Carter, C. K. and Kohn, R. (1994). On Gibbs sampling for state space models. Biometrika, 81(3), 541–553

1994
[3]

Cleveland, W. P. and Tiao, G. C. (1976). Decomposition of seasonal time series: A model for the Census X-11 program. Journal of the American Statistical Association, 71(355), 581–587. 22

1976
[4]

Commandeur, J. J. F. and Koopman, S. J. (2007). An Introduction to State Space Time Series Analysis. Oxford University Press

2007
[5]

and Koopman, S

Durbin, J. and Koopman, S. J. (2012). Time Series Analysis by State Space Methods, 2nd ed. Oxford University Press

2012
[6]

F., Monsell, B

Findley, D. F., Monsell, B. C., Bell, W. R., Otto, M. C., and Chen, B.-C. (1998). New capabilities and methods of the X-12-ARIMA seasonal-adjustment program. Journal of Business & Economic Statistics, 16(2), 127–152. Fr¨ uhwirth-Schnatter, S. (1994). Data augmentation and dynamic linear models. Journal of Time Series Analysis, 15(2), 183–202

1998
[7]

Harvey, A. C. (1989). Forecasting, Structural Time Series Models and the Kalman Filter. Cambridge University Press

1989
[8]

Harvey, A. C. and Todd, P. H. J. (1983). Forecasting economic time series with structural and Box–Jenkins models: A case study. Journal of Business & Economic Statistics, 1(4), 299–307

1983
[9]

Pfeffermann, D. (1991). Estimation and seasonal adjustment of population means using data from repeated surveys. Journal of Business & Economic Statistics, 9(2), 163–175

1991
[10]

Pfeffermann, D., Feder, M., and Signorelli, D. (1998). Estimation of autocorrelations of survey errors with application to trend estimation in small areas. Journal of Business & Economic Statistics, 16(3), 339–348

1998
[11]

Rao, J. N. K. and Molina, I. (2015). Small Area Estimation, 2nd ed. Wiley

2015
[12]

Rubin, D. B. (1984). Bayesianly justifiable and relevant frequency calculations for the applied statistician. The Annals of Statistics, 12(4), 1151–1172

1984
[13]

H., and Musgrave, J

Shiskin, J., Young, A. H., and Musgrave, J. C. (1967). The X-11 variant of the Census method II seasonal adjustment program. Technical Paper 15, US Bureau of the Census

1967
[14]

C., Kennedy, B., and Wu, S

Singh, A. C., Kennedy, B., and Wu, S. (2001). Regression composite estimation for the Canadian Labour Force Survey. Survey Methodology, 27(1), 33–44

2001
[15]

Tam, S.-M. (2026). Dynamic Mini-Max design and sequential HB inference for repeated surveys. arXiv:2606.03702. 23

Pith/arXiv arXiv 2026

[1] [1]

Bell, W. R. and Hillmer, S. C. (1984). Issues involved with the seasonal adjustment of economic time series. Journal of Business & Economic Statistics, 2(4), 291–320

1984

[2] [2]

Carter, C. K. and Kohn, R. (1994). On Gibbs sampling for state space models. Biometrika, 81(3), 541–553

1994

[3] [3]

Cleveland, W. P. and Tiao, G. C. (1976). Decomposition of seasonal time series: A model for the Census X-11 program. Journal of the American Statistical Association, 71(355), 581–587. 22

1976

[4] [4]

Commandeur, J. J. F. and Koopman, S. J. (2007). An Introduction to State Space Time Series Analysis. Oxford University Press

2007

[5] [5]

and Koopman, S

Durbin, J. and Koopman, S. J. (2012). Time Series Analysis by State Space Methods, 2nd ed. Oxford University Press

2012

[6] [6]

F., Monsell, B

Findley, D. F., Monsell, B. C., Bell, W. R., Otto, M. C., and Chen, B.-C. (1998). New capabilities and methods of the X-12-ARIMA seasonal-adjustment program. Journal of Business & Economic Statistics, 16(2), 127–152. Fr¨ uhwirth-Schnatter, S. (1994). Data augmentation and dynamic linear models. Journal of Time Series Analysis, 15(2), 183–202

1998

[7] [7]

Harvey, A. C. (1989). Forecasting, Structural Time Series Models and the Kalman Filter. Cambridge University Press

1989

[8] [8]

Harvey, A. C. and Todd, P. H. J. (1983). Forecasting economic time series with structural and Box–Jenkins models: A case study. Journal of Business & Economic Statistics, 1(4), 299–307

1983

[9] [9]

Pfeffermann, D. (1991). Estimation and seasonal adjustment of population means using data from repeated surveys. Journal of Business & Economic Statistics, 9(2), 163–175

1991

[10] [10]

Pfeffermann, D., Feder, M., and Signorelli, D. (1998). Estimation of autocorrelations of survey errors with application to trend estimation in small areas. Journal of Business & Economic Statistics, 16(3), 339–348

1998

[11] [11]

Rao, J. N. K. and Molina, I. (2015). Small Area Estimation, 2nd ed. Wiley

2015

[12] [12]

Rubin, D. B. (1984). Bayesianly justifiable and relevant frequency calculations for the applied statistician. The Annals of Statistics, 12(4), 1151–1172

1984

[13] [13]

H., and Musgrave, J

Shiskin, J., Young, A. H., and Musgrave, J. C. (1967). The X-11 variant of the Census method II seasonal adjustment program. Technical Paper 15, US Bureau of the Census

1967

[14] [14]

C., Kennedy, B., and Wu, S

Singh, A. C., Kennedy, B., and Wu, S. (2001). Regression composite estimation for the Canadian Labour Force Survey. Survey Methodology, 27(1), 33–44

2001

[15] [15]

Tam, S.-M. (2026). Dynamic Mini-Max design and sequential HB inference for repeated surveys. arXiv:2606.03702. 23

Pith/arXiv arXiv 2026