💰 quantitative finance

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

This paper introduces a centered-innovation modification to Bayesian Dirichlet ARMA models that, while predictively equivalent to the standard specification, significantly improves computational stability by reducing Hamiltonian Monte Carlo divergences caused by pathological posterior geometries in production workflows.

Original authors: Harrison Katz

Published 2026-05-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Harrison Katz

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Predicting the "Pie"

Imagine you are trying to predict how a bank's total money (its assets) is split up next week. The money is divided into four buckets: Cash, Securities (investments), Loans, and Other.

Because the total amount of money is fixed, these four buckets always have to add up to 100%. If the "Loans" bucket gets bigger, one of the others must get smaller. In statistics, this is called a compositional time series. It's like a pie chart that changes shape every week.

The author, Harrison Katz, is studying a specific math recipe (a model) used to predict how these pie slices will shift. He found a tiny tweak to the recipe that makes the computer much less likely to crash, without changing the final prediction at all.

The Problem: The "Raw" Recipe is a Bit Clumsy

The standard way to predict these pie slices involves a step called a "Moving Average" (MA). Think of this as looking at the surprise from last week to guess what will happen next week.

In the standard ("Raw") recipe, the computer calculates a "surprise" by simply subtracting the predicted pie slice from the actual pie slice.

The Flaw: Because of the math behind the "pie" (the Dirichlet distribution), this simple subtraction leaves a tiny, invisible bias. It's like trying to measure a distance with a ruler that is slightly bent. The ruler works fine for small distances, but if you use it too many times, the errors add up and confuse the computer.
The Consequence: When the computer tries to learn from the data (a process called "sampling"), this tiny bend in the ruler makes the math landscape look like a jagged, rocky mountain. The computer's "hiker" (the algorithm) keeps tripping over rocks, getting stuck, or falling off cliffs. In technical terms, this causes divergent transitions (crashes) and makes the computer take a very long time to find the answer.

The Solution: The "Centered" Recipe

The author proposes a tiny fix: Center the Innovation.

Instead of just subtracting the prediction from the actual result, the new recipe first calculates exactly what the "average surprise" should be according to the math, and then subtracts that from the actual result.

The Analogy: Imagine you are throwing darts at a board. The "Raw" method assumes your average throw is dead center. But if you know your arm naturally pulls slightly to the left, the "Centered" method adjusts for that pull before you throw.
The Result: This adjustment straightens out the ruler. The math landscape becomes a smooth, flat plain. The computer's "hiker" can now walk across it easily without tripping.

What the Paper Found

The author tested this on real data: 10 years of weekly bank asset data (from 2015 to 2025). He ran the prediction model 104 times, each time updating the data by one week.

Here are the three main takeaways:

1. The Predictions are Identical
Does the new recipe predict the future better? No.
The paper shows that the "Raw" and "Centered" recipes produce statistically identical predictions. If you asked both models, "What will the loan share be next week?", they would give you the exact same answer. The "Centered" recipe doesn't make the forecast more accurate; it just makes the journey to get there smoother.

2. The Computer Runs Much Better
Does the new recipe make the computer happier? Yes, dramatically.

The Raw Recipe: The computer stumbled and crashed (diverged) 446 times across the 104 tests. In one specific week (February 2025), it crashed 211 times in a single run, essentially giving up.
The Centered Recipe: The computer stumbled only 34 times total. In that same bad week in February 2025, it didn't crash at all.
The Ratio: The new recipe was about 13 times more stable than the old one.

3. It Works No Matter How You Slice the Pie
The author tested the model using different "reference" buckets (e.g., predicting based on Cash instead of Loans).

Predictions: Still identical across all versions.
Stability: The new recipe was still more stable, though the amount of improvement varied. When the old recipe was prone to crashing (like with the "Loans" reference), the new recipe saved the day. When the old recipe was already behaving, they were tied.

Why Should You Care? (The "So What?")

The paper concludes that this change is operational, not predictive.

For a Bank: Banks use these predictions to plan for emergencies (stress tests). If the computer crashes or gets stuck during a calculation, the bank might get a bad or missing report.
The Benefit: By using the "Centered" recipe, banks can run these complex simulations without the computer freezing up or needing a human to manually fix the errors. It's like switching from a car with a sticky transmission to one with a smooth automatic gear shift. You get to the same destination, but the ride is much less likely to break down.

Summary

The paper introduces a simple, plug-and-play math fix for predicting how money is split in a bank.

Old Way: Works, but the computer trips over itself often.
New Way: Works just as well, but the computer glides smoothly.
Verdict: If you are building a system that needs to run reliably without constant human supervision, you should use the "Centered" version.

Technical Summary: Centered-Innovation MA for Bayesian Dirichlet ARMA

1. Problem Statement

Compositional time series, where observations sum to a fixed total (e.g., bank asset shares), are often modeled using the Dirichlet distribution mapped to Euclidean space via the additive log-ratio (ALR) transformation. In observation-driven Bayesian Dirichlet ARMA (B–DARMA) models, the moving-average (MA) block typically utilizes the raw ALR residual, defined as $\epsilon_t^{raw} = \text{alr}(y_t) - \eta_t$ , where $\eta_t$ is the linear predictor.

A fundamental theoretical issue arises under finite precision ( $\phi_t$ ): the conditional expectation of the ALR-transformed observation, $E[\text{alr}(Y_t) | \mu_t, \phi_t]$ , is not equal to the linear predictor $\eta_t$ . Instead, it is a digamma function of the concentration parameters. Consequently, the raw residual $\epsilon_t^{raw}$ possesses a non-zero conditional mean of order $O(\phi_t^{-1})$ . This bias distorts the conditional mean path, obscures the interpretation of MA coefficients, and introduces a deterministic shift that the priors do not anticipate. While frequentist approaches sidestep this by using digamma-based links that lack closed-form inverses, the Bayesian framework often retains the ALR link for tractability, leaving this bias unaddressed.

2. Methodology

The paper proposes a minimal analytic modification to the B–DARMA framework: replacing the raw MA regressor with a centered innovation.

2.1 The Centered Innovation

Instead of using $\epsilon_t^{raw}$ , the authors define the centered innovation $\epsilon_t^\circ$ as:
$\epsilon_t^\circ = \text{alr}(y_t) - E[\text{alr}(Y_t) | \mu_t, \phi_t]$
The conditional mean term $E[\text{alr}(Y_t) | \mu_t, \phi_t]$ is available in closed form via digamma identities:
$g(\mu_t, \phi_t)_j = \psi(\phi_t \mu_{tj}) - \psi(\phi_t \mu_{tj^\star})$
where $\psi(\cdot)$ is the digamma function and $j^\star$ is the reference component. This adjustment ensures that the MA innovation has a conditional mean of zero, restoring the standard interpretation of MA shocks without altering the Dirichlet likelihood or requiring numerical inversion.

2.2 Theoretical Equivalence

The paper establishes a recursion-level first-order equivalence between the centered specification and a hypothetical digamma-link DARMA in the high-precision regime ( $\phi_t \to \infty$ ).

Theorem 1: Under explicit interior (bounded away from simplex boundary) and lag-stability conditions, the centered-innovation recursion and the digamma-link recursion differ by $O(r_T)$ , where $r_T = \max \phi_t^{-1}$ .
Implication: In high-precision regimes, the two specifications should be predictively indistinguishable at the recursion level. However, the paper clarifies that this theoretical equivalence does not automatically govern the geometry of the Bayesian posteriors produced during parameter re-estimation.

2.3 Forecast Recursion

A key practical advantage of the centered specification is the restoration of the standard ARMA forecast recursion.

Centered: Future shocks integrate out to zero. The one-step-ahead conditional mean depends only on observed shocks.
Raw: Future raw residuals contribute via their non-zero conditional mean, introducing a bias term that complicates multi-step forecasting logic.

3. Empirical Application

The methodology is evaluated on weekly Federal Reserve H.8 bank-asset shares (Cash, Securities, Loans, Other) from October 2015 to October 2025 ( $T=522$ ).

Protocol: A rolling one-step evaluation over 104 weekly origins.
Comparison: Identical covariates, priors, and sampler settings (Hamiltonian Monte Carlo via Stan) are used for both the "Raw-MA" and "Centered-MA" specifications.
Sensitivity: A four-reference analysis is conducted, using Cash, Securities, Loans, and "Other" as the ALR reference component in turn.

4. Key Results

4.1 Predictive Performance

Indistinguishability: The predictive performance of the two specifications is statistically indistinguishable across all accuracy metrics (ELPD, RMSE, MAE, and coverage).
Reference Invariance: This equivalence holds regardless of the ALR reference component chosen. The cumulative difference in Expected Log Predictive Density (ELPD) is negligible (e.g., +0.37 nats for the Loans reference, with a mean per-origin difference of +0.0035 nats).
Consistency with Theory: These results align with the high-precision theoretical equivalence, as the realized precision $\phi_t$ remained well above unity (often >80) throughout the test window.

4.2 Sampling Geometry (Hamiltonian Monte Carlo)

While predictive accuracy is identical, the geometric properties of the posterior differ sharply:

Divergent Transitions: The Raw specification suffers from significantly more HMC divergent transitions, indicating pathological posterior geometry (narrow, curved, or funnel-like regions).
- Loans Reference: Raw MA produced 446 divergences across 104 origins, compared to only 34 for Centered MA (a ratio of ~13:1).
- Extreme Cases: A single origin in February 2025 generated 211 divergences for the Raw specification (with a potential scale reduction factor $\hat{R}$ of 1.58), whereas the Centered specification remained clean.
Reference Dependence of Geometry: The magnitude of the geometric advantage varies by reference. It is most dramatic for the Loans reference (where Raw fits occasionally fail catastrophically) and diminishes to parity for the Cash reference, where both specifications behave well.
ESS and $\hat{R}$ : The Centered specification consistently yields higher bulk Effective Sample Sizes (ESS) and lower maximum $\hat{R}$ values, indicating more reliable convergence.

5. Significance and Claims

The paper claims that the primary value of centering is operational rather than predictive.

Operational Robustness: In production workflows where posterior simulation feeds downstream stress tests and asset-liability management, the "catastrophic raw-MA divergence spikes" that occur at isolated rolling origins under the raw specification are a critical failure mode. The centered specification avoids these spikes without operator intervention (e.g., auto-refits).
Minimal Cost: The adjustment is analytic, plug-in, and requires only a local change to the MA innovation calculation. It preserves the predictive content of the model while significantly improving the conditioning of the posterior for the MA coefficients.
Theoretical Clarification: The work clarifies that while the raw and centered specifications are predictively equivalent in high-precision regimes, the raw specification introduces a deterministic shift that creates unnecessary tension in the joint posterior, leading to sampling pathologies that the centered formulation resolves.

The authors conclude that centering should be the natural default for MA terms in observation-driven Dirichlet ARMA models, particularly when posterior simulation quality is a priority for downstream applications.