Directional-Shift Dirichlet ARMA Models for Compositional Time Series with Structural Break Intervention

Harrison Katz

arxiv: 2601.16821 · v2 · submitted 2026-01-23 · 📊 stat.ME · q-fin.ST· stat.AP

Directional-Shift Dirichlet ARMA Models for Compositional Time Series with Structural Break Intervention

Harrison Katz This is my paper

Pith reviewed 2026-05-16 11:36 UTC · model grok-4.3

classification 📊 stat.ME q-fin.STstat.AP

keywords compositional time seriesstructural breaksDirichlet ARMAdirectional shift interventionBayesian time seriessimplex constraintsmonotone transitionsprobabilistic forecasting

0 comments

The pith

A directional-shift intervention in Dirichlet ARMA models captures gradual structural breaks in compositional time series while preserving simplex constraints and short-run dynamics.

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

Compositional time series data, which must sum to one at every time point, often undergo structural breaks from shocks or policy changes that standard models handle poorly by either ignoring them or using inflexible fixed effects. The paper introduces a Bayesian extension to Dirichlet ARMA models that incorporates a directional-shift intervention defined by a direction vector indicating which components increase or decrease, an amplitude scaling the size of the redistribution, and a logistic gate controlling the timing and rate of the transition. This setup ensures that the break follows a geodesic path on the simplex and remains consistent regardless of the coordinate basis chosen for the data. Simulations with 400 model fits demonstrate that when the direction is correctly specified, the model recovers the amplitude with near-zero bias and achieves approximately 80 percent coverage for its credible intervals, outperforming fixed-effect versions that drop to 66 percent coverage for monotone ongoing breaks. Applications to Airbnb occupancy shares during the COVID period confirm that the advantages appear specifically in cases with roughly one-directional transitions.

Core claim

The directional-shift Dirichlet ARMA model augments the base process with an intervention mechanism whose trajectory corresponds to geodesic motion on the simplex, parameterized by a direction vector, amplitude, and logistic gate, thereby producing coherent probabilistic forecasts that maintain compositional constraints through and after structural breaks while retaining the short-run autoregressive and moving-average dependence structure.

What carries the argument

The directional-shift intervention mechanism, which redistributes component shares along a specified direction vector scaled by amplitude and timed by a logistic function, ensuring invariance to ILR basis choice.

If this is right

The model generates coherent probabilistic forecasts that remain valid during the transition period and beyond the sample.
When the break direction is known, amplitude estimates exhibit near-zero bias across simulated scenarios.
Credible intervals attain near-nominal coverage rates specifically for monotone structural breaks.
The intervention separates the structural shift from the short-term DARMA dynamics, improving interpretability of parameters.
In non-monotone post-break regimes, the model remains calibrated though point forecasts may not outperform fixed effects.

Where Pith is reading between the lines

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

If the direction vector must be estimated from data rather than pre-specified, additional uncertainty quantification around that choice would be needed for reliable inference.
The geodesic property on the simplex may allow direct comparison or combination with other manifold-valued time series approaches.
Applying the framework to other sum-constrained series such as budget allocations or species abundances could test its generality beyond the Airbnb examples.

Load-bearing premise

The structural break is roughly monotone with its direction either correctly identified or specified in advance, and the logistic gate plus geodesic motion adequately represent the transition.

What would settle it

A dataset or simulation featuring a non-monotone break or a misspecified direction vector in which the intervention model's credible interval coverage falls substantially below nominal levels or amplitude bias becomes large.

read the original abstract

Compositional time series frequently exhibit structural breaks due to external shocks, policy changes, or market disruptions. Standard methods either ignore such breaks or handle them through fixed effects that cannot extrapolate beyond the sample, or step-function dummies that impose instantaneous adjustment. We develop a Bayesian Dirichlet ARMA model augmented with a directional-shift intervention mechanism that captures structural breaks through three interpretable parameters: a direction vector specifying which components gain or lose share, an amplitude controlling redistribution magnitude, and a logistic gate governing transition timing and speed. The model preserves compositional constraints by construction, maintains DARMA dynamics for short-run dependence, and produces coherent probabilistic forecasts through and after structural breaks. The intervention trajectory corresponds to geodesic motion on the simplex and is invariant to the choice of ILR basis. A simulation study with 400 fits across 8 scenarios shows near-zero amplitude bias and nominal 80\% credible interval coverage when the shift direction is correctly identified (77.5\% of cases); supplementary studies confirm robustness across extreme transition speeds and non-monotone DGPs. Two empirical applications to COVID-era Airbnb data characterize performance relative to simpler alternatives. Where the break is monotone and ongoing, the intervention model achieves near-nominal calibration (79.6\%) while the fixed effect substantially under-covers (66.1\%). Where post-break dynamics are non-monotone, both models are acceptably calibrated and the fixed effect outperforms on point accuracy. The intervention model's advantages are thus specific to settings with roughly monotone structural transitions.

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

The directional-shift intervention adds a workable way to model monotone breaks in compositional series on top of Dirichlet ARMA, but the calibration numbers rest on simulations whose details are not fully checkable from the provided material.

read the letter

The paper introduces a directional-shift mechanism on Dirichlet ARMA models that uses a direction vector, amplitude, and logistic gate to capture how compositional shares redistribute after a shock. The trajectory follows geodesic motion on the simplex and stays invariant to the ILR basis. This is genuinely new relative to fixed effects or simple step dummies, because it allows extrapolation past the break when the shift is roughly monotone and the direction can be identified or specified. The simulation with 400 fits across eight scenarios reports near-zero amplitude bias and nominal 80% coverage when direction is correct (77.5% of cases), and the Airbnb COVID application shows 79.6% calibration for ongoing monotone breaks versus 66.1% for fixed effects. That is the useful part: a concrete forecasting tool for market-share or resource data where breaks are gradual rather than instantaneous. The soft spots are straightforward. The abstract and stress-test note that full equations for the logistic gate, its integration with the ARMA component, the invariance proof, exact simulation design, data rules, and fitting code are not available for inspection. Without those, it is difficult to verify whether the reported coverage is robust to prior choices or specific DGPs, or whether the geodesic implementation is correct in practice. The assumption that the break is monotone and direction identifiable is load-bearing, and the paper itself shows no advantage in non-monotone cases. This work is aimed at applied statisticians who already use Dirichlet or compositional time-series models and need to handle policy or market shocks. A reader who wants a practical extension for extrapolation will find the framing clear and the empirical comparison helpful. It deserves peer review because the core modeling idea addresses a real gap and the simulation design is at least outlined, even though the authors will need to supply code and derivations before the calibration claims can be taken as settled.

Referee Report

3 major / 2 minor

Summary. The paper proposes a Bayesian Dirichlet ARMA model augmented with a directional-shift intervention for structural breaks in compositional time series. The intervention is defined via three parameters—a direction vector, amplitude, and logistic gate—with the break trajectory following geodesic motion on the simplex and claimed to be invariant to ILR basis choice. It maintains short-run DARMA dynamics and produces coherent forecasts. A simulation study (400 fits, 8 scenarios) reports near-zero amplitude bias and nominal 80% credible interval coverage conditional on correct direction identification (77.5% of cases). Empirical applications to COVID-era Airbnb data show the intervention achieving 79.6% calibration versus 66.1% for fixed effects under monotone ongoing breaks, with advantages specific to roughly monotone transitions.

Significance. If the central claims hold after verification, the work would offer a valuable, interpretable extension for compositional time series that allows extrapolation beyond observed breaks while preserving simplex constraints and probabilistic coherence—addressing clear limitations of fixed effects and instantaneous dummies. The parameter interpretability, basis invariance, and reported robustness to transition speeds represent potential strengths for applied work in economics, ecology, and market-share modeling.

major comments (3)

[Model specification] Model specification section: the explicit functional forms for the directional-shift intervention (logistic gate, amplitude scaling, and embedding into the Dirichlet ARMA mean or precision parameters) are absent, preventing verification of the claimed geodesic motion on the simplex and ILR-basis invariance.
[Simulation study] Simulation study section: the data-generating processes for the 8 scenarios, exact prior choices, fitting algorithm, and data exclusion rules are not provided, so the reported near-zero bias and 79.6% vs. 66.1% coverage figures cannot be independently assessed or reproduced.
[Empirical application] Empirical application section: the Airbnb COVID-era dataset definition, variable construction, and precise model specifications used for the calibration comparison are missing, undermining the claim that the intervention outperforms fixed effects specifically under monotone breaks.

minor comments (2)

[Abstract] The abstract refers to 'supplementary studies' confirming robustness but does not indicate their location or summarize the key robustness checks beyond transition speed and non-monotonicity.
Notation for the direction vector and logistic gate parameters should be introduced with explicit symbols in the main text to improve readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed review. The comments correctly identify omissions that affect reproducibility and verifiability. We will revise the manuscript to supply the missing specifications, data-generating processes, and dataset details while preserving the original claims and results.

read point-by-point responses

Referee: [Model specification] Model specification section: the explicit functional forms for the directional-shift intervention (logistic gate, amplitude scaling, and embedding into the Dirichlet ARMA mean or precision parameters) are absent, preventing verification of the claimed geodesic motion on the simplex and ILR-basis invariance.

Authors: We agree that the explicit functional forms must be stated for independent verification. In the revised manuscript we will add the precise definitions: the logistic gate as a scaled logistic function of time with location and scale parameters, the amplitude applied as a scalar multiplier to the direction vector on the simplex, and the additive embedding of the resulting shift into the mean vector of the Dirichlet distribution (with renormalization to maintain the simplex constraint). These additions will also include a short derivation confirming geodesic motion under the Aitchison geometry and explicit invariance under any ILR transformation. revision: yes
Referee: [Simulation study] Simulation study section: the data-generating processes for the 8 scenarios, exact prior choices, fitting algorithm, and data exclusion rules are not provided, so the reported near-zero bias and 79.6% vs. 66.1% coverage figures cannot be independently assessed or reproduced.

Authors: We accept that the current text lacks sufficient detail for reproduction. The revision will expand the Simulation study section (or add an appendix) with the exact data-generating processes for each of the 8 scenarios, the full prior specifications (including hyperparameters), the MCMC algorithm settings (chains, iterations, burn-in, thinning), convergence diagnostics, and any observation exclusion rules applied before fitting. These additions will enable direct replication of the bias, coverage, and direction-identification results. revision: yes
Referee: [Empirical application] Empirical application section: the Airbnb COVID-era dataset definition, variable construction, and precise model specifications used for the calibration comparison are missing, undermining the claim that the intervention outperforms fixed effects specifically under monotone breaks.

Authors: We agree that the dataset and model details are required to substantiate the empirical claims. In the revised manuscript we will insert a dedicated subsection describing the Airbnb data source, the exact compositional variables constructed (including any aggregation or transformation steps), the precise DARMA orders and intervention parameterizations used for both the directional-shift model and the fixed-effects comparator, and the calibration metric computation. This will allow readers to evaluate the reported 79.6% versus 66.1% calibration difference under monotone transitions. revision: yes

Circularity Check

0 steps flagged

No circularity detected in derivation chain

full rationale

The paper introduces a directional-shift intervention for Dirichlet ARMA models via three new parameters (direction vector, amplitude, logistic gate) whose properties (geodesic motion on the simplex, ILR invariance) are asserted as model features rather than derived from prior fitted quantities. The simulation study (400 fits, 8 scenarios) and empirical calibration results (79.6% vs 66.1% coverage) are presented as external validation outcomes, not as quantities forced by construction from the same inputs. No self-citations, uniqueness theorems, or ansatzes from prior author work appear as load-bearing steps in the provided text; the derivation remains self-contained against the stated assumptions without reducing predictions to tautological reparameterizations.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 1 invented entities

The central claim rests on standard compositional and time-series assumptions plus the new intervention parameters; no independent evidence for the invented mechanism is supplied beyond simulation performance.

free parameters (3)

direction vector
Specifies which components gain or lose share during the structural break
amplitude
Controls the magnitude of the redistribution
logistic gate parameters
Govern transition timing and speed

axioms (3)

domain assumption Compositional observations lie on the simplex and follow a Dirichlet distribution
Standard modeling choice for proportions that sum to one
domain assumption Short-run dependence follows DARMA dynamics
Maintains existing time-series structure outside the intervention
ad hoc to paper Structural break trajectory corresponds to geodesic motion on the simplex
Invoked to ensure basis invariance of the intervention

invented entities (1)

directional-shift intervention mechanism no independent evidence
purpose: Captures structural breaks via direction, amplitude, and logistic gate while preserving compositional constraints
New construct introduced to address limitations of fixed effects and step dummies

pith-pipeline@v0.9.0 · 5562 in / 1674 out tokens · 46315 ms · 2026-05-16T11:36:50.149784+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/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

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

The intervention trajectory corresponds to geodesic motion on the simplex and is invariant to the choice of ILR basis.

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.