Stationarity with Occasionally Binding Constraints
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.
Ruth: I'm Ruth.
Warren: And I'm Warren, here with Iris.
Ruth: We're looking at the paper: "Stationarity with Occasionally Binding Constraints".
Warren: Here's the gist of the paper. This paper studies a class of multivariate threshold autoregressive models known as censored and kinked structural vector autoregressions (CKSVAR), which are notably able to accommodate series that are subject to occasionally binding constraints because the CKSVAR model is able to accommodate series that are subject to occasionally binding constraints.
Ruth: Model Definition and Canonical Form The CKSVAR model involves a VAR(k) model in p variables where one series, yt, enters with coefficients that differ according to whether it is above or below a time-invariant threshold b, while the other p−1 series, collected in xt, enter linearly and the canonical CKSVAR is defined by a specific structure of matrices.
Warren: The model can be rewritten compactly as ϕ+(L)y+t + ϕ-(L)y-t + Φx(L)xt = c + ut.
Ruth: For the purposes of this paper, the authors may take b = 0 without loss of generality.
Paper summary: Ruth: So we're wrapping up our look at "Stationarity with Occasionally Binding Constraints" by Duffy, Mavroeidis and Wycherley in this paper. This work introduces the CKSVAR model for time series that hit a binding constraint.
Warren: The authors are basically saying the old rules for checking stationarity with linear vector autoregressions don't work anymore in these cases. They want us to look at how the whole system behaves deterministically when that constraint kicks in >
Iris: They propose using things like the joint spectral radius or related quantities to establish stability because those give us a more reliable estimate of stability for this specific model type. It’s about controlling the dynamics instead of just checking if a simple equation has stable roots >
Ruth: It’s a tool for when you need to get a handle on vector time series where the behavior shifts based on some external limit >
Warren: It sets up a new way of thinking about stability for these kinds of nonlinear vector models that aren't just simple linear ones. That means we need to think about what happens when everyone tries to run this model at once >
Iris: The paper shows that this approach gives us sufficient conditions for the processes generated by a CKSVAR to be stationary, ergodic, and weakly dependent <ref:2307.06190#pg3> It really connects the threshold behavior directly to the stability of the deterministic subsystem >
Ruth: So we have a way to structurally model these complex systems where behavior shifts based on some external limit, but we still need to be careful about the assumptions when applying those stability tests <ref:2307.06190#pg3> The paper is about how this framework allows us to model vector time series subject to occasionally binding constraints >
Conclusion: Ruth: So, to wrap up this look at "Stationarity with Occasionally Binding Constraints" by Duffy, Mavroeidis, and Wycherley, they introduced the CKSVAR model for time series that hit a binding constraint.
Warren: The authors are fundamentally saying that the old rules we use to check stationarity in linear vector autoregressions just don't work anymore when you have these occasional binding constraints in play. They want us to look at how the entire system behaves deterministically when that constraint suddenly kicks in >
Iris: They propose using things like the joint spectral radius or related quantities to establish stability because those give us a more reliable estimate of stability for this specific model type. It’s about controlling the dynamics instead of just checking if a simple equation has stable roots >
Ruth: It turns out it’s a framework for when you need to get a handle on vector time series where the behavior shifts based on some external limit, like that threshold they put in there <ref:2307.06190#pg3>. This lets them model those threshold nonlinearities and occasionally binding constraints in a way that’s structurally sound, even when the usual criteria fall apart >
Warren: And they give us a computational path forward using semidefinite programming and sum of squares programming to approximate those stability measures like the constrained joint spectral radius <ref:2307.06190#pg3>. It's computationally cheap, but they admit it can be too conservative sometimes, which is something you have to watch out for when you’re building real trading strategies or forecasting models >
Iris: So what this means for us is that we get a tool for structural modeling under constraint, but they also flag that their method doesn't cover every single scenario for proving stationarity <ref:2307.06190#pg3>. We need to be careful about the assumptions when we apply those stability tests because they aren't universally sufficient >
Ruth: That’s the balance they strike; they offer a useful way to model these complex systems without overstating the scope of what those specific criteria can actually prove <ref:2307.06190#pg3>. It’s a tool for when you need to model vector time series subject to binding constraints that are more flexible than standard methods allow >
Warren: It sets up a new way of thinking about stability for these kinds of nonlinear vector models that aren't just simple linear ones <ref:2307.06190#pg1>. We need to think about what happens when everyone tries to run this model at once, especially if those threshold nonlinearities kick in simultaneously >
Iris: And they do deliver consistent maximum likelihood estimators for the model parameters, which is important because then we know the parameters we're estimating are actually what they claim to be <ref:2307.06190#pg2>. The identification part seems solid enough for now >
Ruth: Right, so this framework offers flexibility for modeling these constrained systems, but we still need to be careful about the assumptions when applying those stability tests <ref:2307.06190#pg3>. It’s a new way of thinking about stability for these kinds of nonlinear vector models that aren't just simple linear ones <ref:2307.06190#pg1>. We need to think about what happens when everyone tries to run this model at once >
James A. Duffy, Sophocles Mavroeidis, Sam Wycherley
University of Oxford · Stanford University
econ.EM, math.ST, stat.TH
Submitted: 2023-07-12
Updated: 2026-11-16
Comments: ii + 36 pp., 4 figures. Split from arXiv:2211.09604v1. Author accepted manuscript
Code: https://github.com/samwycherley/thresholdr
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 100/100
The gist: This paper studies a class of multivariate threshold autoregressive models known as censored and kinked structural vector autoregressions (CKSVAR)…
Key concepts
- CKSVAR
- A multivariate model that combines a standard Vector Autoregression (VAR) with a threshold nonlinearity. One variable's coefficients change based on whether it is above or below a specific threshold, allowing the model to handle series facing occasional binding constraints.
- Stationarity Conditions
- The paper establishes new sufficient conditions for stationarity that go beyond simple VAR polynomial root analysis. It relies on advanced Markov process theory, treating the CKSVAR as a regime-switching model to ensure the generated time series are stable and ergodic.
- Joint Spectral Radius (JSR)
- A stability criterion used to determine if a nonlinear system is stable. While exact computation is hard, this paper uses it and its constrained version (CJSR) as an improved, numerically computable estimate for the stability degree of the CKSVAR system.
Terminology
Summary
This paper studies a class of multivariate threshold autoregressive models known as censored and kinked structural vector autoregressions (CKSVAR), which are notably able to accommodate series that are subject to occasionally binding constraints because the CKSVAR model is able to accommodate series that are subject to occasionally binding constraints. p. 0
Model Definition and Canonical Form
The CKSVAR model involves a VAR(k) model in p variables where one series, yt, enters with coefficients that differ according to whether it is above or below a time-invariant threshold b, while the other p−1 series, collected in xt, enter linearly and the canonical CKSVAR is defined by a specific structure of matrices. The model can be rewritten compactly as ϕ+(L)y+t + ϕ-(L)y-t + Φx(L)xt = c + ut. p. 3 For the purposes of this paper, the authors may take b = 0 without loss of generality. p. 4
Stationarity and Ergodicity Conditions
The paper develops a set of sufficient conditions for the processes generated by a CKSVAR to be stationary, ergodic, and weakly dependent. The stationarity of the model defies simple analytical characterization available in linear VARs. p. 2 The authors rely on existing results from the theory of ergodic Markov processes, casting the CKSVAR as an instance of the ‘regime-switching’ or VTAR models considered in that literature to establish conditions sufficient for the series generated by a CKSVAR to be stationary, geometrically ergodic, and β-mixing. p. 2
Stability Criteria and Joint Spectral Radius
The authors develop a hierarchy of sufficient conditions for stability, only the most elementary of which is related to the joint spectral radius (JSR). The JSR has previously been applied to nonlinear time-series models, and though their criteria refer to quantities that cannot be computed exactly, they can be approximated numerically to a high degree of precision. p. 3 The constrained joint spectral radius (CJSR) may yield an improved estimate of the stability degree of the system compared to the JSR. p. 0
Asymptotic Inference and Computational Details
The sufficient conditions also deliver consistency and asymptotic normality of maximum likelihood estimators of the model parameters. The analysis involves reparametrizing the CKSVAR model so as to clearly separate identified parameters from those that are unidentified, leading to a parameter space where the MLEs are consistent and asymptotically normal. p. 3 Computational details involve using semidefinite programming (SDP) and sum of squares (SOS) programming to approximate the CJSR, which is computationally inexpensive but can be too conservative. p. 1
Conclusion
The CKSVAR provides a flexible yet tractable framework in which to structurally model vector time series that are subject to an occasionally binding constraint, and more general threshold nonlinearities. The usual criteria for stationarity and ergodicity, in terms of the roots of the autoregressive polynomial(s), no longer apply, being neither necessary nor sufficient for the CKSVAR to generate a stationary time series. p. 3 We must instead employ other criteria, such as the joint spectral radius or the related quantities developed above, to establish stationarity via control over the deterministic subsystem of the CKSVAR. p. 0
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Accurate stationary checks for multivariate threshold autoregressive models can be performed using criteria based on global asymptotic stability, such as
the joint spectral radius (JSR), has previously been applied to nonlinear time-series models.
This allows AI systems to reject model configurations that generate explosive trajectories before extensive simulation is required. -
The AI system can establish geometric ergodicity for CKSVAR processes by verifying the stability of the associated deterministic system, as stated in Theorem 3.1:
Suppose DGP∗ and ERR hold, and that the deterministic system (3.2) associated to the model (3.1) is stable.
This provides a rigorous mathematical foundation for inferring long-term behavior of complex nonlinear series. -
The system can perform asymptotic inference on structural parameters by utilizing the martingale central limit theorem results derived from the stationary distribution, as shown in Theorem 3.2:
Suppose DGP holds, the deterministic system (3.3) is stable, ut ∼i.i.d. N[0, Ip], E∥z0∥m0 4, and θ0 ∈ int Θ.
This leads to a limiting distribution for estimators:n 1/2 (ˆθn − θ0) ⇝ N[0, I(θ0)−1].
-
The AI can identify parameters in the presence of an unknown orthogonal matrix by utilizing the reparametrization derived from QR decomposition, specifically that
the parameters Ψ:= (µ, ϕ+i, ψ−i, Ψx i k i=0 are exactly identified.
This provides a framework for parameter estimation even when structural identification is not fully known. -
The system can compute an approximate upper bound for stability criteria using semidefinite programming (SDP) or sum of squares (SOS) programming to estimate the joint spectral radius, as mentioned in Section 4:
significant progress has been made in calculating approximate upper bounds for both of these quantities, with an (in principle) arbitrarily high degree of accuracy.
Sources
- Cointegration with Occasionally Binding Constraints
- Common Trends and Long-Run Identification in Nonlinear Structural VARs
- Identification in (Endogenously) Nonlinear SVARs Is Easier Than You Think
Related papers
- Backtesting High-Dimensional Expected Shortfall
- Higher-Order Neyman Orthogonality in Moment-Condition Models
- Online Bayesian Inference for a Dirichlet Process Mixture of Binary Logit Models
- Debiased Conditional Independence Tests for Causal Effects
- Sharp Testable Implications of Encouragement Designs
- Discrete Choice with Generalized Social Interactions