Paper: arXiv 2510.18903
Authors: Harrison Katz
Abstract
Observation-driven Dirichlet models for compositional time series commonly use the additive log-ratio (ALR) link and include a moving-average (MA) term based on ALR residuals. In the standard Bayesian Dirichlet Auto-Regressive Moving-Average (B-DARMA) recursion, this MA regressor has a nonzero conditional mean under the Dirichlet likelihood, which biases the mean path and complicates interpretation of the MA coefficients. We propose a minimal change: replace the raw regressor with a centered innovation equal to the ALR residual minus its conditional expectation, computable in closed form using digamma functions. Centering restores mean-zero innovations for the MA block without altering either the likelihood or the ALR link. We provide closed-form identities for the conditional mean and forecast recursion, show first-order equivalence to a digamma-link DARMA while retaining a simple inverse back to the mean composition, and supply ready-to-use code. In a weekly application to the Federal Reserve H.8 bank-asset composition, the centered specification improves log predictive scores with virtually identical point accuracy and markedly cleaner Hamiltonian Monte Carlo diagnostics.
Complexity vs Empirical Score
- Math Complexity: 8.5/10
- Empirical Rigor: 9.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper presents advanced mathematical derivations using Dirichlet distributions, digamma functions, and closed-form identities, while also demonstrating strong empirical rigor with a real-world dataset (Federal Reserve H.8), predictive evaluation metrics (ELPD, coverage), code availability, and robust out-of-sample testing.
Research Flowchart
flowchart TD
A["Research Goal<br>Resolve MA Bias in<br>Dirichlet Compositional Models"] --> B{"Methodology"}
B --> C["Standard B-DARMA<br>Use Raw ALR Residuals"]
B --> D["Centered MA Dirichlet ARMA<br>Use Centered Innovations"]
C --> E["Computational Process<br>Bayesian MCMC / HMC<br>Convergence Issues"]
D --> E
subgraph Inputs ["Data & Theory"]
F["Weekly Fed H.8<br>Bank Asset Composition"]
G["Dirichlet Likelihood<br>ALR Link Function"]
end
F --> E
G --> E
E --> H["Key Outcomes"]
H --> I["Point Accuracy<br>Virtually Identical"]
H --> J["Predictive Score<br>Improved Log Score"]
H --> K["Diagnostics<br>Cleaner HMC Convergence"]
style A fill:#e1f5fe
style H fill:#e8f5e8
style D fill:#fff3e0
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes replacing the raw MA regressor with a centered innovation equal to the ALR residual minus its conditional expectation, restoring mean-zero innovations without altering the likelihood or link.
Source sentence
We propose a minimal change: replace the raw regressor with a centered innovation equal to the ALR residual minus its conditional expectation, computable in closed form using digamma functions.
Assumes 5
- Weak stationarity of the ALR-transformed series. statistical
Source sentence
To verify weak stationarity on the ALR scale, we applied Augmented Dickey–Fuller (ADF) unit-root tests with a drift term and six lags to each of the three ALR coordinates (cash/loans, securities/loans, other/loans).
- Dirichlet likelihood with time‑varying precision driven by a log‑linear function of lagged ALR volatility. model
Source sentence
The Dirichlet precision is time‑varying and driven by a two‑regressor design, log ϕt = γ0 + γ1zt, where zt is a lagged
- Weekly frequency and seasonally adjusted H.8 bank‑asset composition data. data
Source sentence
We use weekly, seasonally adjusted (SA) H.8 series from the Federal Reserve Bank of St. Louis (FRED).
- Identical priors, covariates, and sampler settings across raw and centered models. model
Source sentence
All diagnostics (divergences, treedepth hits, bR, bulk ESS, auto-refits) are logged per fit.
Show 1 more
- Evaluation designs: fixed one‑step holdout (last 104 weeks) and rolling one‑step evaluation (104 weekly origins). other
Source sentence
We report results for two standard designs: (i) a fixed one‑step holdout using the last 104 weeks and (ii) a rolling one‑step evaluation over the most recent 104 weeks with weekly re‑estimation at each origin.
Data
- Federal Reserve H.8 weekly bank‑asset composition macro public
Source sentence
We use weekly, seasonally adjusted (SA) H.8 series from the Federal Reserve Bank of St. Louis (FRED).
Costs
Not applicable
Finds 3
- Centering improves one‑step density forecasts: ELPD 785.913 vs 785.745 and 95% coverage 0.962 vs 0.952, while RMSE/MAE remain essentially tied. predictive-accuracy
Source sentence
Across both designs, centering improves one‑step density forecasts while leaving point accuracy unchanged. In the fixed holdout (104 weeks), ELPD is 785.913 vs.
- Sampler diagnostics are cleaner for the centered model: no divergent transitions versus several divergences and an auto‑refit for the raw‑MA model. statistical
Source sentence
With identical settings, the centered fit completes without divergent transitions, while the raw‑MA fit exhibits several divergences and triggers a single auto‑refit with doubled iterations and warmup.
- In the rolling evaluation, cumulative ELPD gains sum to +0.424, coverage improves to 0.9529, and total divergences drop from 119 to 16. predictive-accuracy
Source sentence
In the rolling experiment (104 weekly origins), cumulative ELPD gains sum to +0.424 with 66:38 per‑origin wins, coverage is closer to nominal (0.9529 vs. 0.9505), and total divergences drop from 119 to 16.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 18 related papers). Coverage is partial, so this is neither replication nor refutation.
4 papers in the archive
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.