Centered-Innovation MA for Bayesian Dirichlet ARMA: Theoretical Equivalence and an Application to Bank-Asset Shares

Harrison Katz

arxiv: 2510.18903 · v3 · pith:2JAJUPC5new · submitted 2025-10-20 · 📊 stat.ME · math.ST· q-fin.ST· stat.TH

Centered-Innovation MA for Bayesian Dirichlet ARMA: Theoretical Equivalence and an Application to Bank-Asset Shares

Harrison Katz This is my paper

Pith reviewed 2026-05-21 20:39 UTC · model grok-4.3

classification 📊 stat.ME math.STq-fin.STstat.TH

keywords Bayesian Dirichlet ARMAcompositional time seriescentered innovationmoving averageHamiltonian Monte Carlopredictive equivalencebank asset shares

0 comments

The pith

Centering the innovation in Bayesian Dirichlet ARMA yields first-order equivalence to a digamma-link specification while reducing divergent transitions in sampling.

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

The paper proves that a minimal change to the moving-average block of a Bayesian Dirichlet ARMA model—replacing the raw additive log-ratio residual with a centered innovation that subtracts the closed-form Dirichlet conditional mean—creates a recursion-level first-order equivalence to a digamma-link DARMA under interior and lag-stability conditions. A sympathetic reader would care because this analytic plug-in adjustment preserves predictive accuracy on compositional time series while addressing practical computational problems in Bayesian fitting. On weekly Federal Reserve bank-asset share data spanning 522 weeks, the two specifications produce statistically indistinguishable forecasts across 104 rolling origins on all examined metrics. At the same time, the raw version triggers roughly ten times more Hamiltonian Monte Carlo divergent transitions at isolated rolling points where localized posterior pathologies appear.

Core claim

We prove a recursion-level first-order equivalence (in 1/φ) between the centered specification and a digamma-link DARMA at fixed parameters, under explicit interior and lag-stability conditions. On the bank-asset data, predictive performance is statistically indistinguishable across 104 rolling origins while Hamiltonian Monte Carlo divergent transitions are approximately an order of magnitude more frequent under the raw specification.

What carries the argument

The centered innovation, which subtracts the Dirichlet conditional ALR mean (available in closed form via digamma identities) from the raw residual inside the moving-average block.

If this is right

Predictive performance remains statistically indistinguishable from the raw specification across all examined accuracy metrics on 104 rolling weekly origins.
Hamiltonian Monte Carlo divergent transitions drop by approximately an order of magnitude, driven by avoidance of isolated raw-posterior pathologies.
The geometric advantage of centering is preserved across four-reference sensitivity analyses but varies in magnitude with the prevalence of raw-fit problems.
The adjustment is analytic and plug-in, requiring only a local change to the MA innovation calculation, and leaves downstream stress-test workflows intact.

Where Pith is reading between the lines

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

Similar centering adjustments could be tested in other observation-driven compositional models to improve sampler reliability without re-deriving full equivalence results.
Production pipelines that refit models weekly on compositional series would gain operational robustness by adopting the plug-in change, especially when occasional divergences halt automated stress testing.
The first-order equivalence result may motivate checking whether higher-order terms in 1/φ or alternative link functions produce comparable stability gains under the same interior conditions.

Load-bearing premise

The first-order equivalence and predictive indistinguishability hold only when interior and lag-stability conditions are satisfied at every rolling origin; violation at any point removes both the theoretical relation and the claimed computational advantage.

What would settle it

At a rolling origin where an interior or lag-stability condition is violated, check whether predictive accuracy metrics or the rate of Hamiltonian Monte Carlo divergent transitions become statistically distinguishable between the centered and raw specifications.

read the original abstract

We study a minimal change to an observation-driven Bayesian Dirichlet ARMA (B--DARMA) for compositional time series: replace the raw additive log-ratio (ALR) residual in the moving-average block with a centered innovation that subtracts the Dirichlet conditional ALR mean, available in closed form via digamma identities. We prove a recursion-level first-order equivalence (in $1/\phi$) between the centered specification and a digamma-link DARMA at fixed parameters, under explicit interior and lag-stability conditions. The result clarifies why the two specifications should be predictively indistinguishable in the high-precision regime but does not by itself govern the geometry of the Bayesian posteriors that re-estimation produces. On weekly Federal Reserve H.8 bank-asset shares (October~2015 through October~2025, $T=522$ weeks), predictive performance is statistically indistinguishable across $104$ rolling weekly origins on every accuracy metric examined, while Hamiltonian Monte Carlo divergent transitions are approximately an order of magnitude more frequent under the raw specification, driven by isolated rolling fits at which the raw posterior exhibits localized pathologies. A four-reference sensitivity analysis confirms that predictive equivalence is reference-invariant and that the geometric advantage of centering is preserved across references but varies with the prevalence of pathological raw fits, from a substantial reduction at the loans reference to parity at the cash reference. The practical implication is operational rather than predictive: centering avoids the catastrophic raw-MA divergence spikes that occur at isolated rolling origins, which matters for production workflows in which posterior simulation feeds downstream stress tests. The adjustment is analytic and plug-in, and requires only a local change to the MA innovation calculation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Centering the MA innovation gives a first-order equivalent Dirichlet ARMA that cuts HMC divergences on the bank data while leaving predictions unchanged, provided the interior and stability conditions hold.

read the letter

This paper's main takeaway is that swapping the raw ALR residual for a centered innovation in the moving-average block of a Bayesian Dirichlet ARMA model produces first-order equivalent predictions but fewer Hamiltonian Monte Carlo problems in practice. The authors prove the recursion-level equivalence in 1/φ to a digamma-link version at fixed parameters, under explicit interior and lag-stability conditions, and then check it on real data. The result is narrow but operational: it targets a known sampler headache without altering the forecasting behavior in the high-precision regime. The bank-asset application uses 104 rolling weekly origins on Federal Reserve H.8 shares from 2015 to 2025, and the predictive metrics line up while divergent transitions fall by roughly an order of magnitude. The sensitivity check across four references is a useful addition, showing the stability gain is not an artifact of one particular ALR reference. The paper does a clean job keeping the theoretical claim qualified and tying the empirical results directly to the rolling-origin design rather than post-hoc tuning. The soft spots are proportionate to the scope. Everything rests on the interior and lag-stability conditions holding at each origin; if those are violated the equivalence and the indistinguishability claim no longer apply. The work is an extension of existing Dirichlet ARMA machinery rather than a new class of models, and the application stays inside one compositional domain. No load-bearing gaps appear in the reported evidence, but a referee would still want to see the full derivation and any edge-case checks. This is for people who already use Dirichlet ARMA or similar compositional time-series models and who care about reliable posterior sampling in repeated fits. A reader facing occasional divergence spikes in production workflows will get a practical fix and a clear explanation of why it works. It deserves a serious referee. The claim is modest, the conditions are stated up front, and the empirical test is direct enough to be worth external scrutiny.

Referee Report

0 major / 3 minor

Summary. The manuscript introduces a centered-innovation moving-average term for Bayesian Dirichlet ARMA (B-DARMA) models of compositional time series by subtracting the closed-form Dirichlet conditional ALR mean from the raw residual. It proves a recursion-level first-order equivalence (in 1/φ) to a digamma-link DARMA at fixed parameters under explicit interior and lag-stability conditions. On weekly Federal Reserve H.8 bank-asset shares (T=522, October 2015–October 2025), rolling-origin evaluation across 104 origins shows statistically indistinguishable predictive performance on all examined metrics, while HMC divergent transitions drop by roughly an order of magnitude; a four-reference sensitivity analysis confirms the geometric advantage is reference-dependent but predictive equivalence is invariant.

Significance. If the stated conditions hold, the work supplies a minimal analytic adjustment that improves posterior sampler reliability for compositional time series without altering high-precision predictive behavior. The rolling-origin design on external data plus reference-sensitivity checks provide concrete evidence of operational value for production pipelines that feed posterior draws into stress tests.

minor comments (3)

[§3] §3 (theoretical development): the interior and lag-stability conditions are stated explicitly, but the empirical section should report how often these conditions are satisfied across the 104 rolling origins (e.g., via a supplementary table of minimum interior probabilities or eigenvalue checks).
[Results] Table 2 or equivalent results table: the claim of 'statistical indistinguishability' would be strengthened by reporting the actual p-values or credible intervals for the pairwise differences in each accuracy metric rather than only stating equivalence.
[Application] The four-reference sensitivity analysis is mentioned in the abstract; the main text should list the four references explicitly and tabulate the divergence counts for each to allow readers to assess the variation in sampler improvement.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive and accurate summary of the manuscript, which correctly identifies the analytic adjustment, the first-order equivalence result under the stated conditions, the rolling-origin predictive equivalence on the H.8 bank-asset shares, and the order-of-magnitude reduction in divergent transitions. We agree with the recommendation for minor revision and have no substantive disagreements with the referee's assessment.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper derives a first-order recursion-level equivalence directly from the definitions of the centered innovation (subtracting the closed-form Dirichlet conditional ALR mean via digamma identities) and the digamma-link DARMA specification, under explicitly stated interior and lag-stability conditions. This is a mathematical identity obtained by expanding the model recursions in 1/φ, not a fitted parameter or self-citation chain. The empirical section applies rolling-origin predictive evaluation to external Federal Reserve H.8 data (T=522 weeks, 104 origins) and reports sampler diagnostics; these results are independent of the theoretical derivation and do not reduce any claim to its own inputs by construction. No load-bearing step matches the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The paper rests on standard Bayesian time-series assumptions plus two explicit conditions for the equivalence proof. No new entities are postulated and no parameters are fitted to produce the equivalence itself.

axioms (1)

domain assumption Interior and lag-stability conditions required for the recursion-level first-order equivalence
Stated explicitly in the abstract as necessary for the equivalence to hold.

pith-pipeline@v0.9.0 · 5835 in / 1456 out tokens · 27538 ms · 2026-05-21T20:39:10.833869+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We prove a recursion-level first-order equivalence (in 1/φ) between the centered specification and a digamma-link DARMA at fixed parameters, under explicit interior and lag-stability conditions.
IndisputableMonolith/Foundation/ArithmeticFromLogic.lean embed_injective unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

ϵ◦t ≡ alr(yt) − E[alr(Yt)|ηt, ϕt] = alr(yt) − g(μt, ϕt) with g via digamma

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.