Centered-Innovation MA for Dirichlet ARMA in Compositional Time-Series Forecasting: Theory and Evidence from Bank-Asset Shares
Harrison Katz · 2025
Read Centered-Innovation MA for Dirichlet ARMA in Compositional Time-Series Forecasting: Theory and Evidence from Bank-Asset Shares by Harrison Katz (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-18903.txt.
Introduction
Harrison Katz, writing alone, studies a small change to a Bayesian time-series model for data that are shares of a whole, such as the asset mix of banks. The model, a Dirichlet ARMA, normally feeds a raw transformed residual into its moving-average part. Katz replaces that residual with a centered innovation, subtracting the exact conditional mean, which is available in closed form through a classic special function identity.
He proves the centered and digamma-link versions agree to first order at fixed parameters under stated conditions, clarifying why predictions should match when precision is high. On weekly Federal Reserve data covering bank asset shares across a decade and five hundred twenty-two weeks, predictions were indeed statistically indistinguishable across many rolling tests. Yet the centered version's Bayesian computation ran far more cleanly, with roughly ten times fewer pathological sampling transitions, a geometric advantage confirmed across reference choices.
Catalog record
| CREATOR | Harrison Katz |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 38 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-18903
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Centered-Innovation MA for Dirichlet ARMA in Compositional Time-Series Forecasting: Theory and Evidence from Bank-Asset Shares Harrison Katz Abstract We study a minimal change to an observation–driven Dirichlet ARMA used for compositional time-series forecasting : replace the raw ALR residual with a centered innovation that subtracts the Dirichlet conditional ALR mean, available in closed form via digamma identities. Centering yields mean–zero MA shocks and an ARMA-consistent forecast recursion, while keeping the Dirichlet likelihood and the ALR link (with its closed-form inverse) unchanged. On weekly Federal Reserve H.8 bank-asset shares, the centered specification improves density forecasts while leaving point error essentially unchanged. In a fixed 104-week holdout, one-step ELPD is 785.913 785.913 vs. 785.745 785.745 and empirical 95 % 95\% coverage is 0.962 0.962 vs. 0.952 0.952 (Raw), with cleaner HMC diagnostics. In a 104-origin rolling one-step experiment, cumulative ELPD gains sum to + 0.424 +0.424 ( 66 66 vs. 38 38 wins), coverage is closer to nominal ( 0.9529 0.9529 vs. 0.9505 0.9505 ), and divergences drop from 119 119 to 16 16 across fits, at effectively tied RMSE/MAE. Because the adjustment is analytic and plug-in, it provides a practical recipe for density forecasting on the simplex without changing likelihood, link, priors, or code structure. Keywords: Compositional time series; probabilistic forecasting; density forecasting; Dirichlet ARMA (B–DARMA); additive log–ratio (ALR); centered innovations; expected log predictive density (ELPD); H.8 bank assets. 1 Introduction Compositional time series arise whenever a fixed total is allocated across categories through time. Finance and data science provide many examples: allocations of earned fees into future recognition buckets for planning and staffing, evolving market or sector shares in portfolio analytics, the distribution of transactions across settlement currencies that drives treasury, hedging, and consolidated reporting, and bank balance sheet shares such as cash, securities, loans, and other assets in the Federal Reserve’s H.8 release. Valid forecasts must respect the simplex constraints, so they must remain nonnegative and sum to one. Classical work maps compositions to Euclidean space with log ratios. Additive, centered, and isometric log-ratio transformations enable standard multivariate tools while preserving the subcompositional coherence that practitioners care about ( Aitchison 1982 ; Egozcue et al. 2003 ) . These mappings motivate transformed VARMA and state space approaches across marketing, demographics, ecology, environmental science, and forecasting ( Cargnoni et al. 1997 ; Ravishanker et al. 2001 ; Silva and Smith 2001 ; Brunsdon and Smith 1998 ; Mills 2010 ; Barceló-Vidal et al. 2011 ; Koehler et al. 2010 ; Kynčlová et al. 2015 ; Snyder et al. 2017 ; AL-Dhurafi et al. 2018 ) . Modeling directly on the simplex is an alternative that avoids ad hoc renormalization and yields coherent predictive distributions. For shares and market fractions, Dirichlet regression and its variants are widely used, and there is a growing literature on Dirichlet time series in both state space and observation-driven forms ( Hijazi and Jernigan 2009 ; Grunwald et al. 1993 ; da Silva et al. 2011 ; da Silva and Rodrigues 2015 ; Zheng and Chen 2017 ; Morais et al. 2018 ; Giller 2020 ; Creus-Martí et al. 2021 ; Tsagris and Stewart 2018 ) . Within this class, the Bayesian Dirichlet ARMA framework (B–DARMA) evolves the Dirichlet mean on the additive log-ratio (ALR) scale with a VARMA process and has been used for forecasting lead times, investigating prior sensitivity, and modeling energy mixes ( Katz et al. 2024 ; Katz et al. 2025 ; Katz and Maierhofer 2025 ) . Time-varying precision accommodates volatility clustering on the simplex in a Dirichlet–ARCH spirit ( Katz and Weiss 2025 ) . These ideas connect to broader Bayesian time series references ( Prado and West 2010 ; West 1996 ) and to Bayesian VAR and VARMA models with shrinkage or stochastic volatility ( Bańbura et al. 2010 ; Karlsson 2013 ; Huber and Feldkircher 2019 ; Kastner and Huber 2020 ) . They also sit alongside generalized linear time-series designs for non-Gaussian data ( Brandt and Sandler 2012 ; Roberts and Penny 2002 ; Chen and Lee 2016 ; McCabe and Martin 2005 ; Berry and West 2020 ; Fukumoto et al. 2019 ; Silveira de Andrade et al. 2015 ) and the volatility literature that motivates precision dynamics ( Engle 1982 ; Bollerslev 1986 ; Nelson 1991 ; Bauwens et al. 2006 ; Engle 2001 ; Francq and Zakoïan 2019 ; Silvennoinen and Teräsvirta 2009 ; Tsay 2005 ) . A practical issue arises for moving-average terms under a Dirichlet likelihood. With finite precision, the conditional expectation of \operatorname{alr}(\mathbf{Y}_{t}) is a digamma function of the concentration parameters and is not equal to the linear predictor. The commonly used regressor \operatorname{alr}(\mathbf{y}_{t})-\bm{\eta}_{t} therefore has nonzero conditional mean, which biases the conditional mean path and obscures the interpretation of MA coefficients. Frequentist Dirichlet ARMA designs sidestep this by using a digamma-based link whose inverse depends on precision and is not available in closed form ( Zheng and Chen 2017 ) . We study a minimal fix that keeps the Dirichlet likelihood and the ALR link: replace the raw regressor with a centered innovation e_{t}^{\circ}=\operatorname{alr}(\mathbf{y}_{t})-\mathbb{E}\{\operatorname{alr}(\mathbf{Y}_{t})\mid\bm{\mu}_{t},\phi_{t}\} . The expectation has a closed form via digamma functions, so the centering is straightforward to compute, restores mean-zero innovations for the MA block, and delivers mean-consistent forecasts without changing the likelihood or requiring numerical inversion. We evaluate with predictive tools standard in Bayesian time series. We summarize out-of-sample fit with expected log predictive density and approximate leave-future-out cross-validation ( Vehtari et al. 2017 ; Vehtari et al. 2015 ; Vehtari and Ojanen 2012 ; Bürkner et al. 2020 ) , and we report interpretable point-error summaries for compositions. The empirical study focuses on public weekly H.8 bank asset shares, compares Raw–MA and Centered–MA under identical covariates and priors, uses a fixed 104-week holdout, and then runs a rolling one-step evaluation over the last two years to accumulate differences in expected log predictive density and track coverage. The code follows modern Bayesian forecasting workflows ( R Core Team 2022 ; Stan Development Team 2022 ; Tsay et al. 2022 ; Hyndman and Athanasopoulos 2018 ) . For banking decisions this matters directly. Calibrated density forecasts of H.8 shares feed liquidity planning and asset and liability management by quantifying the probability that cash and securities buffers will breach internal limits and by shaping the expected cost of near‑term funding, while better calibrated one‑step densities for the asset mix improve hedge sizing and risk budgets for interest‑rate risk when reallocations are volatile. Contribution to forecasting. This paper makes three forecasting contributions. (i) It shows that the usual ALR residual in Dirichlet ARMA carries a nonzero conditional mean under finite precision and gives an analytic centering that restores mean–zero MA shocks without altering the likelihood or link. (ii) It derives the forecast recursion that results (already-available shocks enter means; future ones do not), clarifying the interpretation of MA coefficients for one-step forecasting. (iii) Empirically, on weekly H.8 shares, centering improves density forecasts, higher ELPD and closer-to-nominal coverage, while point accuracy remains tied and Hamiltonian Monte Carlo diagnostics improve (Figures 2–4; Table 2). 2 Model recap and centered innovations Let \mathbf{y}_{t}=(y_{t1},\dots,y_{tJ})^{\prime} be a J –part composition. Conditional on a mean \bm{\mu}_{t}\in\Delta^{J-1} and precision \phi_{t}>0 , \mathbf{y}_{t}\mid\bm{\mu}_{t},\phi_{t}\sim\mathrm{Dir}(\phi_{t}\bm{\mu}_{t}). Fix a reference