Bounded GARCH Processes

summary

The gist

Bounded GARCH processes propose a novel class of Bounded GARCH (B-GARCH) models designed to characterize conditional heteroskedasticity for continuous time series constrained within a compact…

In short

Bounded GARCH (B-GARCH) models are a new class of models for continuous time series constrained within a compact interval. They address boundary violations common in standard GARCH by ensuring the conditional distribution follows a Beta distribution. This framework provides a proper predictive distribution and dynamic variance equation with parameter constraints that strictly maintain the bounded support almost surely.

Key concepts

Bounded GARCH (B-GARCH)
A model designed for continuous time series restricted to a compact interval, ensuring the conditional distribution follows a Beta distribution. This approach solves boundary violation issues found in standard GARCH models when applied to bounded data, providing a proper predictive framework.
Conditional Variance Equation
The dynamic equation describing how the conditional variance ($h_t$) evolves over time. In B-GARCH, this equation includes parameters constrained such that the resulting process strictly preserves the bounded support of the entire series almost surely.
Strict Stationarity and Ergodicity Proof
Proving these fundamental properties for B-GARCH processes is novel because standard Lyapunov exponent theory does not apply directly. The proof relies on verifying contraction conditions from Doukhan & Wintenberger (2008) to establish the required strict stationarity and ergodicity.
Conditional Maximum Likelihood Estimator (CMLE)
The method used to estimate the model parameters by maximizing the conditional log-likelihood function. It involves specific terms like $ ext{log}(1 - y^2t)^{2h_t}$, and its asymptotic normality is rigorously established using information matrix calculations.

Terminology used across episodes

This episode discusses

The paper

Bounded GARCH Processes · Read on arXiv

University College Dublin

We propose a novel class of Bounded GARCH (B-GARCH) processes designed to characterise the conditional heteroskedasticity of continuous time series constrained within a compact interval. While standard GARCH specifications are tailored for R-valued data, many empirical series, such as realised correlation, unemployment rates, expenditure shares, and relative humidity, are bounded. Fitting a traditional GARCH model to such data creates a severe support mismatch, frequently generating predictions and density forecasts that breach the natural boundaries of the data (boundary violation). Furthermore, conventional approaches utilising non-linear link transformations onto the real line might induce interpretability issues and estimation biases. To address these limitations, our proposed B-GARCH framework ensures a proper predictive distribution and an equation for the dynamic conditional variance with a parameter constraint that strictly preserves the bounded support of the entire process almost surely. We establish the fundamental probabilistic properties of the process, proving strict stationarity and ergodicity, and address parameter estimation via conditional maximum likelihood. The asymptotic consistency and normality of the estimators are rigorously derived, and their finite-sample performance is validated through Monte Carlo simulations. An empirical application to the monthly realised correlation between the Vanguard Real Estate ETF (VNQ) and the iShares 20+ Year Treasury Bond ETF (TLT) demonstrates that B-GARCH models effectively capture volatility clustering while completely eliminating boundary violation issue that plagues traditional GARCH models.

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Bounded GARCH Processes".

Jane: Here's the gist of the paper. Bounded GARCH processes propose a novel class of Bounded GARCH (B-GARCH) models designed to characterize conditional heteroskedasticity for continuous time series constrained within a compact interval, which is crucial because standard GARCH specifications often suffer from boundary violation when applied to bounded data The proposed framework ensures a proper predictive distribution and an equation for the dynamic conditional variance with a parameter constraint that strictly preserves the bounded support of the entire process almost surely Model Formulation and Constraints The B-GARCH(p, q) processes are defined such that a stochastic sequence is bounded if its conditional distribution follows a Beta distribution where the variance evolves according to a stochastic dynamic driven by lagged squared observations and previous conditional variances Specifically, the process satisfies E(YtFt-1) = 0, while the conditional variance ht follows ht = α0 + Xp i=1 αiY 2t−i + Xq j=1 βjht−j < 1 The choice of the beta distribution is motivated by key theoretical and empirical considerations, serving as the natural counterpart for continuous data restricted to a compact interval A common feature shared by all these models is that they are designed to model the conditional mean, thereby modelling the conditional variance only indirectly Theoretical Foundations and Properties The framework establishes fundamental probabilistic properties, including strict stationarity and ergodicity for the B-GARCH processes The proof of strict stationarity and ergodicity for our processes cannot be established using standard theory (top Lyapunov exponent theory) from classical GARCH models, so our proof represents a novel contribution to the literature The process preserves the support Yt ∈ (−1, 1), allowing us to model bounded series without the need for non-linear transformations Under stationarity, Var(Yt) = α0 / (α0 + Xp i=1 αi + Xq j=1 βj) Parameter Estimation and Asymptotics Parameter estimation is achieved via conditional maximum likelihood estimation, and the asymptotic consistency and normality of the estimators are rigorously derived The conditional log-likelihood function involves terms like log(1 − y 2t) 2ht + log Γ 1/2ht− log Γ 1/ht− 1/2 The conditional maximum likelihood estimator (CMLE) θb is derived, and its asymptotic normality is established as √n(θb − θ) d−→ N(0, Σ −1) with Σ defined by the information matrix I(θ) Monte Carlo simulations validate these theoretical results and demonstrate robust finite-sample uncertainty quantification Empirical Application and Performance An empirical application to the monthly realised correlation between the Vanguard Real Estate ETF (VNQ) and the iShares 20+ Year Treasury Bond ETF (TLT) demonstrates that B-GARCH models effectively capture volatility clustering while completely eliminating boundary violation issue that plague traditional GARCH models The proposed methodology provides a general statistical framework for continuous processes on compact domains across the physical, economic, and social sciences The AR(1)-B-GARCH(p, q) processes can be generalized to accommodate asymmetry by introducing a dynamic conditional mean µt rather than assuming a zero-mean process specification The B-GARCH class of processes successfully resolves the support mismatch and subsequent boundary violations that plague classical GARCH specifications when applied to bounded data The paper concludes by proposing an extension of the B-GARCH models by allowing for a monotonic continuous link function v(ht) into the conditional variance formulation, which provides greater flexibility and facilitates the incorporation of exogenous covariates Future research avenues include investigating the kurtosis properties of the BGARCH(1,0) process under specific conditions and applying this framework to other bounded time series such as expenditure fractions The paper also notes that while the benchmark Gaussian AR-GARCH(1,1) model exhibits similar mean persistence, its conditional variance structure is unconstrained with respect to the support of yt The proposed bounded GARCH models maintain support preservation across the entire sample while accurately capturing dynamic heteroskedasticity Appendix Details The appendix provides technical proofs for key propositions, such as Proposition 2.3 which details a set of inequalities that bound the kurtosis of the B-GARCH(1, 0) process The proof for Theorem 3.2 relies on verifying contraction conditions of Doukhan & Wintenberger (2008) to establish strict stationarity and ergodicity The proof for Theorem 3.3 shows that the supremum of the difference between the conditional log-likelihood function and its initialisation is bounded, leading to consistency The proof for Theorem 2.5 uses stochastic order results to show that Z' is stochastically smaller than Z'' if h' ≤ h'' The proof for Condition (C3) involves showing that the required derivatives are bounded above by a finite constant, ensuring the existence and uniqueness of the maximum likelihood estimators The proof for Theorem 3.2 establishes that Σ is positive definite, confirming the asymptotic normality result The final proof for Theorem 3.3 uses the Markov inequality to obtain the desired result concerning convergence The appendix also includes calculations for covariances and variances of Ut(θ) based on equations (20) and (21) The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the joint estimation of saturation rate in Model 3 The appendix concludes with detailed proofs for the necessary conditions (C1), (C2), and (C3) required by Lemma 3.1 The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ The appendix also includes calculations for the covariance matrix Σ, showing it is positive definite The appendix provides detailed derivations for the bounds on the log-likelihood differences, showing convergence to zero as n→∞ <ref:2611.

Paper summary: Tom: So we're looking at this paper called "Bounded GARCH Processes." Basically, they've got this new way to model volatility when your data isn't just any real number, but it’s stuck inside a certain range.

Jane: Right, so the main problem they’re tackling is that standard GARCH models are built for numbers that can go anywhere on the real line. When you use them on data like realized correlation, which is bounded between negative one and one, you run into this big issue called boundary violation where their predictions shoot outside that range.

Lu: Exactly. The authors propose a B-GARCH framework specifically designed to keep the conditional variance equation constrained so that the entire process stays within its defined boundaries almost surely, which is a pretty strong claim for continuous time series <ref:2611.08158#pg3>.

Meng: But keeping it bounded while still modeling volatility clustering sounds tricky. How do you make the dynamic variance equation work without violating those boundaries constantly?

Lalam: They define the process such that the conditional variance follows a stochastic dynamic driven by lagged squared observations and previous conditional variances, and they constrain that equation with a parameter constraint to strictly preserve the bounded support of the entire process almost surely <ref:2611.08158#pg3>.

Tom: That's interesting because it avoids those messy non-linear transformations that people usually use when they try to fit GARCH to bounded data, which often introduce estimation biases.

Jane: And they do this by basing the conditional distribution on a Beta distribution, which makes sense because that's the natural counterpart for continuous data restricted to an interval like negative one and positive one <ref:2611.08158#pg3>.

Lu: Beyond just fitting it, they established the rigorous theoretical foundations needed for inference, showing strict stationarity and ergodicity for these B-GARCH processes, which is a novel contribution because standard Lyapunov exponent theory doesn't quite cover them <ref:2611.08158#pg3>.

Meng: And then they have the parameter estimation part. They use conditional maximum likelihood estimation to get an estimator, and they proved that this estimator is consistent and normal asymptotically, which means it should be reliable for real-world data <ref:2611.08158#pg3>.

Paper summary: Lalam: The conditional log-likelihood function they derive involves terms like the logarithm of one minus y squared times the conditional variance plus logarithms involving the Gamma function scaled by half the conditional variance and its inverse, which is what drives their estimation process <ref:2611.08158#pg3>.

Tom: So, empirically they tested this on something real. I remember reading that they applied it to the monthly realized correlation between Vanguard Real Estate ETF and Treasury Bond ETF <ref:2611.08158#pg3>.

Jane: That’s right, and what’s important there is that the B-GARCH models effectively capture volatility clustering while completely eliminating that boundary violation issue that plagues traditional GARCH models <ref:2611.08158#pg3>.

Lu: They also noted something about generalization; they said the AR one B-GARCH p, q processes can be extended to accommodate asymmetry by introducing a dynamic conditional mean instead of assuming zero-mean <ref:2611.08158#pg3>.

Meng: From an engineering standpoint, I'm curious about how this plays out at scale. What does the computational cost look like when you have to track these conditional constraints and estimate parameters for a process that’s strictly bounded?

Lalam: The asymptotic normality of the conditional maximum likelihood estimator is established as root n theta b minus theta converging to a normal distribution with zero mean and a covariance matrix defined by sigma minus one, which gives us the uncertainty quantification we need <ref:2611.08158#pg3>.

Tom: So, for someone just listening to the show who isn't deep in econometrics, what does this actually mean for their data? What's the big picture here?

Jane: It means that if you’re dealing with any continuous series that naturally has limits—like correlation or expenditure shares—this framework gives you a proper statistical tool instead of forcing your data into a square box where it just breaks <ref:2611.08158#pg2>.

Lu: It moves beyond just fitting the equation; they’ve done the heavy lifting on proving that these processes are strictly stationary and ergodic, which is a big step for theoretical reliability in this area <ref:2611.08158#pg3>.

Meng: I worry about those proofs being too complex to implement efficiently in a production environment. Are these constraints computationally cheap enough for daily model updates?

Lalam: The appendix shows the derivations for the necessary conditions, like condition C3, which ensures the existence and uniqueness of maximum likelihood estimators by bounding required derivatives with a finite constant <ref:2611.08158#pg3>.

Paper summary: Tom: Okay, so we've seen that it handles bounded data correctly and has solid estimation theory. But what are they missing? What’s the catch?

Jane: Well, one thing they pointed out is that while their model respects the support of the data, it still doesn't perfectly replicate everything a standard model does; for instance, their benchmark Gaussian AR-GARCH one by one with one is unconstrained with respect to the support of y t <ref:2611.08158#pg3>.

Lu: And they did suggest future work focusing on the kurtosis properties of the B-GARCH one by zero under certain conditions, which suggests there’s still room to explore how this framework handles extreme events <ref:2611.08158#pg3>.

Meng: That sounds like a direction I could look into for stress testing. If we can bound the kurtosis of the simpler version, that helps with risk management when deploying these kinds of models to production <ref:2611.08158#pg3>.

Lalam: They also mentioned extending this by allowing a monotonic continuous link function into the conditional variance formulation, which adds flexibility and lets you incorporate external factors like covariates <ref:2611.08158#pg3>.

Tom: So to wrap up, the B-GARCH Processes paper is a rigorous statistical tool that solves the boundary violation problem for continuous data by constraining the conditional variance equation so it stays within the data's natural limits <ref:2611.08158#pg1>.

Jane: It means we can apply this to things like realized correlation or expenditure shares without having to guess how to transform the data first, giving us a proper predictive distribution right out of the gate <ref:2611.08158#pg3>.

Lu: The implication is that for any bounded time series, we have a formal way to model its conditional heteroskedasticity that respects the physical or structural boundaries of the data set <ref:2611.08158#pg2>.

Meng: For me, it means if I’m building a system that tracks something bounded, I don't have to spend all my time wrestling with boundary violations in my loss function; this structure handles it for the variance dynamics <ref:2611.08158#pg3>.

Lalam: It’s pretty confident in its results because they derived the asymptotic normality of the conditional maximum likelihood estimator using rigorous stochastic order results, which shows convergence to zero for log-likelihood differences as n goes to infinity <ref:2611.08158#pg3>.

Conclusion: Tom: So what does that title actually mean in plain English?

Jane: It means they’ve taken the standard GARCH model, which usually lets things go off to infinity, and they’ve put a leash on it so it respects the boundaries of the data you are actually looking at.

Lu: The core idea is ensuring that no matter what happens in time, your conditional variance equation stays within those bounds almost surely. That’s a big constraint for continuous time series modeling.

Meng: But how does keeping everything constrained work without making the math impossibly complicated? What's the actual mechanism behind that constraint?

Lalam: They define the process such that the conditional distribution follows a Beta distribution, and they build a parameter constraint into the variance equation so it strictly preserves that bounded support. It’s about forcing stability within those limits.

Tom: So it’s not just fitting data; it’s building a statistical structure that inherently understands physical or economic limits on the data itself. That sounds like something useful for things like financial correlations.

Jane: Exactly. The authors show this framework works empirically on things like the monthly correlation between real estate and bonds, and it successfully captures volatility clustering without those messy boundary violations you see in older GARCH models.

Lu: They also proved that these processes are strictly stationary and ergodic, which is a significant theoretical win because standard theory doesn't cover them. That means you can trust the long-term behavior of the model.

Meng: If it’s theoretically sound, does that translate to something practical for me when I’m deploying models in production? Does this mean faster training or just more reliable results?

Lalam: The estimation part is rigorous too; they derive the asymptotic normality of the maximum likelihood estimator, showing how sure we can be about our parameter estimates as we collect more data.

Tom: It sounds like they’ve got the mechanics down—the math for consistency and normality is solid. But what’s the next step for this class of models?

Jane: They suggest an extension where you can link a continuous function into that variance formula, which lets you bring in outside information or covariates, adding more flexibility to the structure.

Lu: Future work involves looking at how these processes handle things like kurtosis under specific conditions, exploring the tails of the distribution even within these bounded constraints.

Meng: I’m interested in that—if we can bound the kurtosis for a simpler version, that gives us better risk management tools when we actually put these models into our systems.

Lalam: That’s where it gets interesting; connecting this structural constraint to managing extreme events is a key research direction moving forward.

More episodes

← Home