Bayesian Analysis (2021) 16, Number 4, pp.

1157–1185

Bayesian Modelling of Time-Varying

Conditional Heteroscedasticity
Sayar Karmakar∗ and Arkaprava Roy†

Abstract. Conditional heteroscedastic (CH) models are routinely used to ana-

lyze financial datasets. The classical models such as ARCH-GARCH with time-
invariant coefficients are often inadequate to describe frequent changes over time
due to market variability. However, we can achieve significantly better insight by
considering the time-varying analogs of these models. In this paper, we propose a
Bayesian approach to the estimation of such models and develop a computation-
ally efficient MCMC algorithm based on Hamiltonian Monte Carlo (HMC) sam-
pling. We also established posterior contraction rates with increasing sample size
in terms of the average Hellinger metric. The performance of our method is com-
pared with frequentist estimates and estimates from the time constant analogs. To
conclude the paper we obtain time-varying parameter estimates for some popular
Forex (currency conversion rate) and stock market datasets.
Keywords: autoregressive model, B-splines, Hamiltonian Monte Carlo (HMC),
non-stationary, posterior contraction, volatility.

1 Introduction
For datasets observed over a long period, stationarity turns out to be an oversimplified
assumption that ignores systematic deviations of parameters from constancy. Thus time-
varying parameter models have been studied extensively in the literature of statistics,
economics, and related fields. For example, the financial datasets, due to external fac-
tors such as war, terrorist attacks, economic crisis, political events, etc. exhibit deviation
from time-constant stationary models. Accounting for such changes is crucial as other-
wise time-constant models can lead to incorrect policy implications as pointed out by
Bai (1997). Thus functional estimation of unknown parameter curves using time-varying
models has become an important research topic today. In this paper, we analyze popular
conditional heteroscedastic models such as AutoRegressive Conditional Heteroscedas-
ticity (ARCH) and Generalized ARCH (GARCH) from a Bayesian perspective in a
time-varying setup. Before discussing our new contributions in this paper, we provide a
brief overview of some previous works in this area.
In the regression regime, time-varying models have garnered a lot of recent atten-
tion; see, for instance, Fan and Zhang (1999), Fan and Zhang (2000), Hoover et al.
(1998), Huang et al. (2004), Lin and Ying (2001), Ramsay and Silverman (2005), Zhang
et al. (2002) among others. The models show time-heterogeneous relationship between
response and predictors. Consider the following two regression models
Model I: yi = xTi θi + ei , Model II: yi = xTi θ0 + ei , i = 1, . . . , n,

∗ Department of Statistics, University of Florida, sayarkarmakar@ufl.edu

† Department of Biostatistics, University of Florida, ark007@ufl.edu


c 2021 International Society for Bayesian Analysis https://doi.org/10.1214/21-BA1267
1158 Bayesian Modelling of Time-Varying Conditional Heteroscedasticity

where xi ∈ Rd (i = 1, . . . , n) are the covariates, T is the transpose, θ0 and θi = θ(i/n)

are the regression coefficients. Here, θ0 is a constant parameter and θ : [0, 1] → Rd is
a smooth function. Estimation of θ(·) has been considered by Hoover et al. (1998), Cai
(2007) and Zhou and Wu (2010) among others. One popular way to decide if there is
an evidence to favor time-varying models over the time-constant analogue is to perform
hypothesis testing. See, for instance, Zhang and Wu (2012), Zhang and Wu (2015),
Chow (1960), Brown et al. (1975), Nabeya and Tanaka (1988), Leybourne and McCabe
(1989), Nyblom (1989), Ploberger et al. (1989), Andrews (1993) and Lin and Teräsvirta
(1999). Zhou and Wu (2010) discussed obtaining simultaneous confidence bands (SCB)
in model I, i.e. with additive errors. However their treatment is heavily based on the
closed-form solution and it does not extend to processes defined by a more general
recursion.
For time-varying AR, MA, or ARMA processes, the results from time-varying linear
regression can be naturally extended. However, such an extension is not obvious for
conditional heteroscedastic (CH hereafter) models. These are, by the simple definition
of evolution is difficult to estimate even in the time-constant case. However, one can-
not possibly ignore its usefulness in analyzing and predicting financial datasets. These
models (even the simple time-constant ones) have remained primary tools for analyzing
and forecasting certain trends for stock market datasets since Engle (1982) introduced
the classical ARCH model and Bollerslev (1986) extended it to a more general GARCH
model. However, with the rapid dynamics of market vulnerability, the simple classical
time-constant models fail in terms of estimation or prediction due to over-compensating
the past data. Several references point out the necessity of extending these classical mod-
els to a set-up where the parameters can change across time, for example Stărică and
Granger (2005), Engle and Rangel (2005) and Fryzlewicz et al. (2008a). Consider the
simple tvARCH(1) model

Xi = σi ζi , ζi ∼ N (0, 1), σi2 = μ0 (i/n) + a1 (i/n)Xi−1

2
.

Similar models can be defined for tvGARCH(1,1) as well where σi2 has an additional
2
recursive term involving σi−1

Xi = σi ζi , ζi ∼ N (0, 1), σi2 = μ0 (i/n) + a1 (i/n)Xi−1

2 2
+ b1 (i/n)σi−1 .

When the two recursive parameters in a GARCH model sum up to 1, i.e. a1 + b1 = 1 it

is usually called an integrated GARCH (iGARCH; or bubble garch/explosive garch by
some authors) process which employing the above display can also be extended towards
a time-varying analog i.e. b1 (·) = 1 − a1 (·). A wide range of financial datasets exhibits
iGARCH phenomena.
In the parlance of time-varying parameter models in the CH setting, numerous works
discussed the CUSUM-type procedure, for instance, Kim et al. (2000) for testing change
in parameters of GARCH(1,1). Kulperger et al. (2005) studied the high moment partial
sum process based on residuals and applied it to residual CUSUM tests in GARCH
models. Interested readers can find some more change-point detection results in the
context of CH models in James Chu (1995), Chen and Gupta (1997), Lin et al. (1999),
Kokoszka et al. (2000) or Andreou and Ghysels (2006).
S. Karmakar and A. Roy 1159

A time-varying framework and a pointwise curve estimation using M-estimators for

locally stationary ARCH models were provided by Dahlhaus and Subba Rao (2006).
Since then, while several pointwise approaches were discussed in the tvARMA and
tvARCH case (cf. Dahlhaus and Polonik (2009), Dahlhaus and Subba Rao (2006),
Fryzlewicz et al. (2008a)), pointwise theoretical results for estimation in tvGARCH
processes were discussed in Rohan and Ramanathan (2013) and Rohan (2013) for
GARCH(1,1) and GARCH(p, q) models respectively. In a series of recent works Kar-
makar et al. (2021); Karmakar (2018) such models were discussed in wide generality.
However, the focus remained frequentist, and the main goal accomplished there was to
build simultaneous inference. One strong criticism for the CH type models remained
that one needs a relatively large sample size (n ∼ 2000) to achieve nominal coverage
levels. The recursive definition of the models and a subsequent kernel-based method of
estimating make it difficult to achieve satisfying results for relatively smaller sample
sizes. This motivated us to explore a Bayesian way of building and estimating these
models and use the posteriors to construct posterior estimates of the coefficient curves
θ(·).
In this paper, we develop a Bayesian estimation method for time-varying analogs of
ARCH, GARCH, and iGARCH models. We model the time-varying functional parame-
ters using cubic B-splines. In the context of general varying-coefficient modeling, spline
bases are a popular choice for its convenience and flexibility (Hastie and Tibshirani,
1993; Gu and Wahba, 1993; Cai et al., 2000; Biller and Fahrmeir, 2001; Huang et al.,
2002; Huang and Shen, 2004; Amorim et al., 2008; Fan and Zhang, 2008; Yue et al., 2014;
Franco-Villoria et al., 2019). Specific to the literature of time-varying volatility model-
ing, B-spline-based models have also been explored (Engle and Rangel, 2008; Audrino
and Bühlmann, 2009; Liu and Yang, 2016).
Our contributions in this paper are two-fold. Towards the methodological devel-
opment, note that the tvARCH, tvGARCH, and tviGARCH models require complex
shape constraints on the coefficient functions. We achieve those by imposing different
hierarchical structures on B-spline coefficients. The constraints are designed to be able
to develop an efficient sampling algorithm based on gradient-based Hamiltonian Monte
Carlo (HMC) (Neal et al., 2011; Betancourt and Girolami, 2015; Betancourt, 2017; Liv-
ingstone et al., 2019). Strong motivation towards implementing such a Bayesian method-
ology was to circumvent the requirement of a huge sample size which is almost essential
for effective estimation using the frequentist and kernel-based methods. This require-
ment on sample size has been frequently pointed out in the literature of ARCH/GARCH
models and thus this was one of our main motivations to see if a reasonable estimation
scheme can be designed in a Bayesian way.
Secondly, the existing literature on obtaining posterior concentration rates for de-
pendent data is thin, even for an extremely simple model. To the best of our knowledge,
ours is the first such attempt towards a theoretical development for these models under
Gaussian-link. Posterior contraction rates for these models with respect to the average
Hellinger metric are established. The main challenge therein is to construct exponen-
tially consistent tests for these classes of models. Using some recently developed tools
from Jeong (2019); Ning et al. (2020) we have developed such tests. We first establish
1160 Bayesian Modelling of Time-Varying Conditional Heteroscedasticity

posterior contraction rates with respect to average log-affinity and then the same rate
is transferred to the average Hellinger metric. The frequentist literature on inference
about time-varying needs very stringent moment assumption and local stationarity as-
sumptions which are often difficult to verify. Moreover, for econometric datasets, the
existence of even the fourth moment is often questionable. Thus this paper offers some
alternative way to estimate coefficients under lesser assumptions.
The rest of the paper is organized as follows. Section 2 describes the proposed
Bayesian model in detail. Section 3 discusses an efficient computational scheme for
the proposed method. We calculate posterior contraction rate in Section 4. In Section 5
we study the performance of our proposed method in the light of. Section 6 deals with
real data application of the proposed methods for the three separate models and con-
cludes with a brief interpretation of the results. We wrap the paper up with discussions,
some concluding remarks, and possible future directions in Section 7. The supplemen-
tary materials Karmakar and Roy (2021) contain theoretical proofs and some additional
results.

2 Modeling
We elaborate on the models and our Bayesian framework for time-varying analogs of
three specific cases that are popularly used to analyze econometric datasets.

2.1 tvARCH Model

Let {Xi } satisfy the following time-varying ARCH(p) model for Xi given Fi−1 = {Xj :
j ≤ (i − 1)},

Xi |Fi−1 ∼ N(0, σi2 ), (2.1)


p
σi2 = μ(i/n) + 2
ak (i/n)Xi−k , (2.2)
k=1

where the parameter functions μ(·), ai (·) satisfy


P = {μ, ak : μ(x) ≥ 0, 0 ≤ ak (x) ≤ 1, sup ak (x) < 1}. (2.3)
x
k

In a Bayesian regime we put priors on μ(·) and ai (·). To respect the shape-constraints
as imposed by P we reformulate the problem. With Bj as the B-spline basis functions,
let

K1
μ(x) = exp(βj )Bj (x),
j=1


K2
ak (x) = θkj Mk Bj (x), 0 ≤ θkj ≤ 1,
j=1
S. Karmakar and A. Roy 1161

exp(δi )
Mi = p , i = 1, . . . , p,
k=0 exp(δk )
δl ∼N (0, c1 ), for 0 ≤ l ≤ p,
βj ∼N (0, c2 ) for 1 ≤ j ≤ K1 ,
θkj ∼U (0, 1) for 1 ≤ k ≤ p, 1 ≤ j ≤ K2 .

P is P-supported.The
The prior induced by above construction
P
verification is very
straightforward. In above construction, j=0 Mj = 1. Thus j=1 Mj ≤ 1. Since 0 ≤
P P P
θkj ≤ 1, supx ai (x) ≤ Mi . Thus supx i=1 ai (x) ≤ i=1 Mi ≤ 1. We have j=1 Mj ≤ 1
if and only if δ0 = −∞, which has probability zero. On the other hand, we also have
μ(·) ≥ 0 as we have exp(βj ) ≥ 0. Thus, the induced priors, described above are well
supported in P.

2.2 tvGARCH Model

Let {Xi } satisfy the following time-varying GARCH(p, q) model for Xi given Fi−1 =
{Xj : j ≤ (i − 1)},

Xi |Fi−1 ∼ N(0, σi2 ),


p 
q
σi2 = μ(i/n) + 2
ak (i/n)Xi−k + 2
bj (i/n)σi−j . (2.4)
k=1 j=1

Additionally we impose the following constraints on parameter space for the time-
varying parameters,
 
P1 = {μ, ai : μ(x) ≥ 0, 0 ≤ ai (x), 0 ≤ bj (x), sup ak (x) + bj (x) < 1}. (2.5)
x
k j

The condition on the AR parameters imposed by (2.5) is somewhat popular in time-

varying AR literature. See Dahlhaus and Subba Rao (2006); Fryzlewicz et al. (2008b);
Karmakar et al. (2021) for example. Different from these references, we additionally do
not assume existence of any unobserved local-stationary process that are close to the
observed process.
To proceed with Bayesian computation, we again put priors on the unknown func-
tions μ(·), ai (·) and bj (·)’s such that they are supported in P1 . Again the restrictions
imposed by (2.5) are respected. The complete description of prior is


K1
μ(x) = exp(βj )Bj (x),
j=1


K2
ak (x) = θkj Mk Bj (x), 0 ≤ θkj ≤ 1, 1 ≤ k ≤ p,
j=1
1162 Bayesian Modelling of Time-Varying Conditional Heteroscedasticity


K3
bk (x) = ηkj Mk+p Bj (x), 0 ≤ ηij ≤ 1, 1 ≤ k ≤ q,
j=1
exp(δi )
Mi =  p , i = 1, . . . , p + q,
k=0 exp(δk )
δl ∼N (0, c1 ), for 0 ≤ l ≤ p + q,
βj ∼N (0, c2 ) for 1 ≤ j ≤ K1 ,
θkj ∼U (0, 1) for 1 ≤ k ≤ p, 1 ≤ j ≤ K2 ,
ηkj ∼U (0, 1) for 1 ≤ k ≤ q, 1 ≤ j ≤ K3 .
Here Bj ’s are the B-spline basis functions. The parameters δj ’s are unbounded. The
verification of support condition 2.5 for the proposed prior is similar.

2.3 tviGARCH Model

Although the GARCH(1,1) remains one of the most popular models to analyze econo-
metric datasets, empirical evidenceshows that these datasets regularly raise suspicion
to the parameter space restriction i ai + j bj < 1. Note that we used a time-varying
analog of this restriction for the tvGARCH modeling
 in Section
 2.2. This often creates
a very stringent condition as the validity of i ai (t) + j bj (t) < 1 is questionable.
The special case for a time-constant GARCH model where this restriction fails is called
an iGARCH model in the literature. We consider the following time-varying analog of
iGARCH.
Xi |Fi−1 ∼ N(0, σi2 ),

p 
q
σi2 = μ(i/n) + 2
ak (i/n)Xi−k + 2
bj (i/n)σi−j . (2.6)
k=1 j=1

We impose the following constraints on parameter space for the time-varying pa-
rameters,
 
P = {μ, ai : μ(x) ≥ 0, 0 ≤ ak (x) ≤ 1, ak (x) + bj (x) = 1}. (2.7)
k j

The prior functions that allow us to reformulate the problem keeping it consistent with
(2.7) is described below:

K1
μ(x) = exp(βj )Bj (x),
j=1


K2
ak (x) = θkj Mk Bj (x), 0 ≤ θkj ≤ 1, 1 ≤ k ≤ p,
j=1


K3
bi (x) = ηkj Mk+p Bj (x), 0 ≤ ηij ≤ 1, 1 ≤ i ≤ (q − 1),
j=1
S. Karmakar and A. Roy 1163

p 
q−1 
bq (x) =1 − ak (x) + bj (x) ,
k=1 j=1
exp(δi )
Mi = p+q−1 , i = 1, . . . , p + q − 1,
k=0 exp(δk )
δl ∼N (0, c1 ), for 0 ≤ l ≤ p + q − 1,
βj ∼N (0, c2 ) for 1 ≤ j ≤ K1 ,
θkj ∼U (0, 1) for 1 ≤ k ≤ p, 1 ≤ j ≤ K2 ,
ηkj ∼U (0, 1) for 1 ≤ k ≤ (q − 1), 1 ≤ j ≤ K3 .

3 Posterior Computation and Implementation

3.1 tvARCH Structure
The complete likelihood L of the proposed Bayesian method is given by

 
p
n
 
L ∝ exp − {μ(i/n) + 2
ak (i/n)Xi−k + Xi2 log μ(i/n)
i=p k=1


p 
K1 
p
+ 2
ai (i/n)Xi−i } − βj2 /(2c2 ) − δl2 /(2c1 ) 10≤θkj ≤1 ,
i=1 j=1 l=0

K 1 K 2 exp(δj )
where μ(x) = j=1 exp(βj )Bj (x), ak (x) = j=1 θkj Mk Bj (x) and Mj =  p exp(δ k)
.
k=0
We develop efficient Markov Chain Monte Carlo (MCMC) algorithm to sample the pa-
rameter β, θ and δ from the above likelihood. The computation of derivatives allows
us to develop an efficient gradient-based MCMC algorithm to sample these parame-
ters. We calculate the gradients of negative log-likelihood (− log L) with respect to the
parameters β, θ and δ. The gradients are given below,
 
d log L Bj (i/n)Xi2
− = exp(βj ) 1 −  2 ) + βj /c2 , (3.1)
βj i
(μ(i/n) + j aj (i/n)Xi−j
 
d log L Bj (i/n)Xi2
− = Mk 1 −  2 ) , (3.2)
θkj i
(μ(i/n) + j aj (i/n)Xi−j
d log L  
− = δj /c1 + (Mj 1{j=k} − Mj Mk ) θkj Bj (x)
δj i
k
  Bj (i/n)Xi2
× 1−  2 ) , (3.3)
i
(μ(i/n) + j aj (i/n)Xi−j

where 1{j=k} stands for the indicator function which takes the value 1 when j = k.
1164 Bayesian Modelling of Time-Varying Conditional Heteroscedasticity

3.2 tvGARCH / tviGARCH Structure

The complete likelihood L2 of the proposed Bayesian method of (2.4) is given by
 
p 
q
n
 
L2 ∝ exp − {μ(i/n) + ai (i/n)Xt−i + bi (i/n)λt−i + Xt log μ(i/n)
t=p i=1 i=1


p 
q 
K1 
p
+ ai (i/n)Xt−i + bi (i/n)λt−i } − βj2 /(2c2 ) − δl2 /(2c1 )
i=1 i=1 j=1 l=0

− (d1 + 1) log λ0 − d1 /λ0 10≤θij ,ηij ≤1 .

We calculate the gradients of negative log-likelihood (− log L2 ) with respect to the

parameters β, θ, η and δ. The gradients are given below,
  2
d log L2 Bj (i/n)Xi−j
− = exp(βj ) 1−  2
 2 ) +βj /c2 ,
βj t
(μ(i/n) + j aj (i/n)Xi−j ) + k bk (i/n)σi−k
  2
d log L2 Bj (i/n)Xi−j
− = Ml 1 −  2
 2 ) ,
θlj t
(μ(i/n) + j aj (i/n)Xi−j ) + k bk (i/n)σi−k
  2
d log L2 Bj (i/n)σi−j
− = Mp+k 1 −  2
 2 ,
ηkj t
(μ(i/n) + j a j (i/n)X i−j ) + k bk (i/n)σi−k )
d log L2 
− = δj /c1 + (Mj 1{j=k} − Mj Mk )×
δj
k
   2
Bj (i/n)Xi−j
θij Bj (x) 1 −  2 )+
 2 1{j≤p} +
t
(μ(i/n) + j aj (i/n)Xi−j k bk (i/n)σi−k )
i≤p

   Bj (i/n)σt2
ηkj Bj (x) 1 −  2 )+
 2 1{j>p} .
t
(μ(i/n) + j aj (i/n)Xi−j k bk (i/n)σi−k )
1≤k≤q

While fitting tvGARCH(p, q), we assume for any t < 0, Xt2 = 0, σt2 = 0. Thus, we need
to additionally estimate the parameter σ02 . The derivative of the likelihood concerning
σ02 is calculated numerically using the jacobian function from R package pracma. For
the tviGARCH, the derivatives are similar so we avoid computing them for the sake of
brevity.
Based on these gradient functions, we develop gradient-based Hamiltonian Monte
Carlo (HMC) sampling. Note that, parameter spaces of θkj ’s have bounded support.
We circumvent this by mapping any Metropolis candidate falling outside the parameter
space back to the nearest boundary. HMC has two parameters, required to be spec-
ified. These are the leap-frog step and the step-size parameter. It is difficult to tune
both of them simultaneously. We choose to tune the step size parameter to maintain
an acceptance range between 0.6 to 0.8. After every 100 iterations, the step-length is
adjusted (increased or reduced) accordingly if it falls outside. Neal et al. (2011) showed
S. Karmakar and A. Roy 1165

that a higher leapfrog step is better for estimation accuracy at the expense of greater
computation. To maintain a balance between accuracy and computational complexity,
we keep it fixed at 30 and obtain good results.

4 Large-Sample Properties
We now focus on obtaining posterior contraction rates for our proposed Bayesian models.
The posterior consistency is studied in the asymptotic regime of increasing number of
time points n. We study the posterior consistency with respect to the average Hellinger
distance on the coefficient functions which is
  
1 1
d21,n = d2H (κ1 , κ2 ) = ( f 1 − f 2 )2 ,
n n
n
where f1 = i=1 Pκ1 (Xi |Xi−1 ) and f2 denotes the corresponding likelihoods.
(n)
Definition: For a sequence n if Πn (d(f, f0 )|X (n) ≥ Mn n |X (n) ) → 0 in Fκ0 -probability
for every sequence Mn → ∞, then the sequence n is called the posterior contraction
rate.
All the proofs are postponed to the supplementary materials. The proof is based
on the general contraction rate result for independent non-i.i.d. observations (Ghosal
and Van der Vaart, 2017) and some results on B-splines based finite random series. The
exponentially consistent tests are constructed leveraging on the famous Neyman-Pearson
Lemma as in Ning et al. (2020). Thus the first step is to calculate posterior contraction
 1/2 1/2
rate with respect to average log-affinity rn2 (f1 , f2 ) = − n1 log f1 f2 . Then we show
that rn2 (f1 , f2 )  2n implies n1 d2H (f1 , f2 )  2n . We also consider following simplified
priors for αj and τi to get better control over tail probabilities,

αj ∼ Gamma(g1 , g1 ), τi ∼ U (0, 1). (4.1)

4.1 tvARCH Model

Let κ = (μ, a1 ) stand for the complete set of parameters. For sake of generality of the
method, we put a prior on K1 and K2 with probability mass function given by,

Π(Ki = k) = bi1 exp[−bi2 k(log k)bi3 ], (4.2)

for i = 1, 2. These priors have not been considered while fitting the model as it would re-
quire computationally expensive reversible jump MCMC strategy. The contraction rate
will depend on the smoothness of true coefficient functions μ and a and the parameters
b13 and b23 from the prior distributions of K1 and K2 . Let κ0 = (μ0 , a01 ) be the truth
of κ.
Assumptions (A): There exists constants MX > 1, 0 < Mμ < MX such that,

(A.1) The coefficient functions satisfy supx μ0 (x) < Mμ and supx a01 (x) < 1 − Mμ /MX .
1166 Bayesian Modelling of Time-Varying Conditional Heteroscedasticity

(A.2) inf x min(μ0 (x), a01 (x)) > ρ for some small ρ > 0.

(A.3) E(X02 ) < MX .

Assumptions (A.1)–(A.3) ensure


Mμ
Eκ0 (Xi2 ) = Eκ0 (Eκ0 (Xi2 |Xi−1 )) < Mμ + 1 − MX < MX
MX

by recursion.
Theorem 1. Under assumptions (A.1)–(A.3), let the true functions μ0 (·) and a10 (·) be
Hölder smooth functions with regularity level ι1 and ι2 respectively, then the posterior
contraction rate with respect to the distance d21,n is
 
max n−ι1 /(2ι1 +1) (log n)ι1 /(2ι1 +1)+(1−b13 )/2 , n−ι2 /(2ι2 +1) (log n)ι2 /(2ι2 +1)+(1−b23 )/2 ,

where bij are specified in (4.2).

4.2 tvGARCH Model

Let κ = (μ, a1 , b1 ) stand for the complete set of parameters. For sake of generality of
the method, we put a prior on K1 , K2 and K3 with probability mass function given by,

Π(Ki = k) = bi1 exp[−bi2 k(log k)bi3 ], (4.3)

for i = 1, 2. These priors have not been considered while fitting the model as it would re-
quire computationally expensive reversible jump MCMC strategy. The contraction rate
will depend on the smoothness of true coefficient functions μ and a and the parameters
b13 and b23 from the prior distributions of K1 and K2 . Let κ0 = (μ0 , a01 ) be the truth
of κ.
Assumptions (B): There exists constants MX > 1, 0 < Mμ < MX such that,

(B.1) The coefficient functions satisfy supx μ0 (x) < Mμ and supx (a01 (x) + b01 (x)) <
1 − Mμ /MX .

(B.2) inf x min(μ0 (x), a01 (x), b01 (x)) > ρ for some small ρ > 0.

(B.3) E(X02 ) < MX , σ00

2
< MX .

Assumptions (B.1) and (B.3) ensure


Mμ
Eκ0 (Xi2 ) = Eκ0 (Eκ0 (Xi2 |Xi−1 )) < Mμ + 1 − MX < MX
MX

by recursion. Similarly we have E(σi2 ) = Eκ0 (Eκ0 (Xi2 |Fi )) = Eκ0 (Xi2 ) < MX .
S. Karmakar and A. Roy 1167

Theorem 2. Under assumptions (B.1)–(B.3), let the true functions μ0 (·), a10 (·) and
b10 (·) be Hölder smooth functions with regularity level ι1 . ι2 and ι3 respectively, then the
posterior contraction rate with respect to the distance d21,n is

max n−ι1 /(2ι1 +1) (log n)ι1 /(2ι1 +1)+(1−b13 )/2 , n−ι2 /(2ι2 +1) (log n)ι2 /(2ι2 +1)+(1−b23 )/2 ,

n−ι3 /(2ι3 +1) (log n)ι2 /(2ι3 +1)+(1−b33 )/2 ,

where bij are specified in (4.3).

4.3 tviGARCH Model

Let κ = (μ, a1 ) stand for the complete set of parameters. For sake of generality of the
method, we put a prior on K1 and K2 with probability mass function given by,

Π(Ki = k) = bi1 exp[−bi2 k(log k)bi3 ], (4.4)

for i = 1, 2. These priors have not been considered while fitting the model as it would re-
quire computationally expensive reversible jump MCMC strategy. The contraction rate
will depend on the smoothness of true coefficient functions μ and a and the parameters
b13 and b23 from the prior distributions of K1 and K2 . Let κ0 = (μ0 , a01 ) be the truth
of κ.

(C.1) The coefficient functions satisfy supx μ0 (x) < Mμ < ∞ for some Mμ .
(C.2) inf x (μ0 (x)) > ρ, inf x a01 (x) > ρ, supx a0,1 (x) < 1 − ρ for some ρ > 0.
Theorem 3. Under assumptions (C.1)–(C.2), let the true functions μ0 (·) and a10 (·) be
Hölder smooth functions with regularity level ι1 and ι2 respectively, then the posterior
contraction rate with respect to the distance d21,n is
 
−ι1 /(2ι1 +1) ι1 /(2ι1 +1)+(1−b13 )/2 −ι2 /(2ι2 +1) ι2 /(2ι2 +1)+(1−b23 )/2
max n (log n) ,n (log n) ,

where bij are specified in (4.4).

5 Simulation
We run simulations to study the performance of our proposed Bayesian method in cap-
turing the true coefficient functions under different true models. The hyperparameters
c1 and c2 of the normal prior are all set 100, which makes the prior weakly informa-
tive. We consider 4, 5 and 6 equidistant knots for the B-splines when n = 200, 500 and
1000 respectively. We collect 10000 MCMC samples and consider the last 5000 as post
burn-in samples for inferences. We shall compare the estimated functions with the true
functions in terms of the posterior estimates of functions along with its 95% pointwise
1168 Bayesian Modelling of Time-Varying Conditional Heteroscedasticity

credible bands. The credible bands are calculated from the MCMC samples at each
point t = 1/T, 2/T, . . . , 1. We take the posterior mean as the posterior estimate of the
unknown functions.
Since, to the best of our knowledge, there is no other Bayesian model for these
time-varying conditional heteroscedastic models, we compare our Bayesian estimates
with corresponding frequentist time-varying estimates. For computing the time-varying
estimates of these models, we use the kernel-based method from Karmakar et al. (2021).
The M-estimator of the parameter vector θ(t) are obtained using the conditional quasi
log-likelihood. For instance, in the tvARCH(1) case, say θ(t) = (μ(t), a1 (t))

n 
t − i/n
θ̂bn (t) = argmin K (Xi |Fi−1 , θ) t ∈ [0, 1],
θ∈Θ i=2
bn

where (·) denotes the Gaussian log-likelihood. Note that these methods are fast but
usually need a cross-validated choice of bandwidth bn . We use K(x) = 3/4(1−x2 )I(|x| ≤
1) and an appropriately chosen bandwidth as suggested by the authors therein. Since
our discussion also involves iGARCH formulation, we wrote a separate kernel-based
frequentist estimation for iGARCH models analogously. Apart from these two time-
varying estimates, we also obtain a time-constant fit on the same data to help initiate a
discussion on whether there was a necessity of introducing coefficients varying with time.
For this tseries and rugarch R packages are used respectively for ARCH/GARCH and
iGARCH fits.
To compare these estimates, we evaluate the average mean square errors (AMSE) for
the three estimates. Note that in an usual linear regression
 of response y on predictor
X scenario, the fitted MSE is often defined as n1 (yi − ŷi )2 . Since here, Xi |Fi−1 ∼
N (0, σi2 ), we use the following definition of AMSE
1 2
AMSE = (Xi − σ̂i2 )2 ,
n i

where the σ̂i2 is computed with the fitted parameter values as per the model under
consideration. For example, for a tvGARCH(1,1) model we have

σ̂i2 = μ̂(i/n) + â(i/n)Xi−1

2 2
+ b̂(i/n)σ̂i−1 ,

where μ̂(·), â(·) and b̂(·) are the estimated curves from the posterior. Replacing the
response yi by Xi2 is natural as often autocorrelations of Xi2 are checked to gauge
presence of CH effect. Moreover, one of the early methods to deal with CH models was
to view Xi2 approximated by an TVAR(1) process. See Bose and Mukherjee (2003) and
references therein. Similar estimators as our proposed AMSE to evaluate the fitting
accuracy has been used in the literature previously. See Starica (2003); Fryzlewicz et al.
(2008a); Rohan and Ramanathan (2013); Karmakar et al. (2021) for example.
In the next three subsections, we provide the results for the three models, namely,
tvARCH, tvGARCH, and tviGARCH. Our conclusions from these results are illustrated
at the end of the section.
S. Karmakar and A. Roy 1169

Figure 1: tvARCH(1): True coefficient functions (red), estimated curve (black) along
with the 95% pointwise credible bands (green) are shown for T = 200, 500, 1000 from
top to bottom.

5.1 tvARCH Case

We start by considering the following tvARCH(1) model from 2.2. Three different
choices for n are considered, n = 200, 500 and 1000. The true functions are,

 
μ0 (x) =10 exp − (x − 0.5)2 /0.1 ,
a10 (x) =0.4(x − 0.15)2 + 0.1.

We compare the estimated functions with the truth for sample size 1000 in Figures 1.
Table 1 illustrates the performance of our method with respect to other competing
methods.
1170 Bayesian Modelling of Time-Varying Conditional Heteroscedasticity

ARCH(1) Frequentist tvARCH(1) Bayesian tvARCH(1)

n = 200 96.42 90.34 85.22
n = 500 128.07 122.53 118.45
n = 1000 138.06 130.33 127.06
Table 1: AMSE comparison for different sample sizes across different methods when the
true model is tvARCH with p = 1.

5.2 tvGARCH Case

Next we explore the following GARCH(1,1) model (cf. 2.4)for different choices of n. The
true functions are,

μ0 (x) =1 − 0.8 sin(πx/2),

a10 (x) =0.5 − (x − 0.3)2 ,
b10 (x) =0.4 − 0.5(x − 0.4)2 .

Note that, estimation of GARCH, due to the additional bi (·) parameter curves is a
significantly more challenging problem and often requires a much higher sample size to
have a reasonable estimation. We show by the means of the following pictures in Figure 2
that the estimation looks reasonable even for smaller sample sizes. The AMSE score
comparisons are shown in Tables 2. The performance of our method is also contrasted
with other competing methods.

GARCH(1,1) Frequentist tvGARCH(1,1) Bayesian tvGARCH(1,1)

n = 200 33.99 31.84 29.43
n = 500 45.46 34.77 33.33
n = 1000 42.60 37.09 36.55
Table 2: AMSE comparison for different sample sizes across different methods when the
true model is tvGARCH(1,1).

5.3 tviGARCH Case

Finally we consider the tviGARCH(1,1) model (cf. 2.6) a special case of GARCH. Note
that due to the constraint a1 (·) + b1 (·) = 1 we only consider the mean function and
AR(1) function for plotting. For this case, our true functions are as follows
 
μ0 (x) = exp − (x − 0.5)2 /0.1 ,
a10 (x) =0.4(x − 1)2 + 0.1.

The frequentist computation for tviGARCH method is carried out based on a kernel-
based estimation scheme along the same line as Karmakar et al. (2021). The estimated
plots along with the 95% credible intervals are shown in Figure 3 for three sample sizes
n = 200, 500, 1000 and the AMSE scores in Table 3.
S. Karmakar and A. Roy 1171

Figure 2: tvGARCH(1,1): True coefficient functions (red), estimated curve (black) along
with the 95% pointwise credible bands (green) are shown for T = 200, 500, 1000 from
top to bottom.

To summarize, our estimated functions are close to true functions for all the cases.
We also find that the credible bands are getting tighter with increasing sample size. Thus
estimates are improving in precision as sample size increases as shown in Figures 1 to 3.
AMSEs of our Bayesian estimates are at least better for all the cases as in Tables 1 to 3.
For tviGARCH, AMSE* is considered due to the huge and somewhat incomparable
values of AMSE due to non-existent variance.

6 Real Data Application

Towards applying our methods on real-life datasets we stick to econometric data for
varying time horizons. These datasets show considerable time-variation justifying our
models to be suitable for understanding how the parameter functions have evolved.
1172 Bayesian Modelling of Time-Varying Conditional Heteroscedasticity

Figure 3: tviGARCH(1,1): True coefficient functions (red), estimated curve (black) along
with the 95% pointwise credible bands (green) are shown for T = 200, 500, 1000 from
top to bottom.

Typically we model the log-return data of the daily closing price of these data to avoid
the unit-root scenario. The log-return is defined as follows and is close to the relative
return 
Pi − Pi−1 Pi − Pi−1
Yi = log Pi − log Pi−1 = log 1 + ≈ ,
Pi−1 Pi−1
where Pi is the closing price on the ith day. Conditional heteroscedastic models are
popularly used for model building, analysis and forecasting. Here we extend such models
to a more sophisticated and general scenario by allowing the coefficient functions to vary.
In this section, we analyze two datasets: USD to JPY conversion and NASDAQ, a
popular US stock market data. We analyze the NASDAQ data through tvGARCH(1,1)
and tviGARCH(1,1) models and USDJPY conversion rate data through tvARCH(1)
models. We just fit one lag for these models as multiple lag fits are similar and larger
lags seem to be insignificant. This result is consistent with the findings in Karmakar
S. Karmakar and A. Roy 1173

iGARCH(1,1) Frequentist tviGARCH(1,1) Bayesian tviGARCH(1,1)

200 8.20 23.86 8.14
500 9.06 18.72 9.06
1000 10.59 25.92 10.59
Table 3: AMSE* comparison for different sample sizes across different methods when the
true model is tviGARCH with p = 1, q = 1. AMSE* stands for mean of the log(AMSE).

et al. (2021), Fryzlewicz et al. (2008b) etc. Moreover, as Fryzlewicz et al. (2008b) claims,
stock indices and Forex rates are more suited to GARCH and ARCH type of models re-
spectively for their superior predictive performance. Each of these datasets was collected
up to 31 July 2020. We exhibit our results for the last 200, 500 and 1000 days which
capture the last 6 months, around 1.5 years, and around 3 years of data respectively.
All these datasets were collected from www.investing.com. Note that these datasets
are usually available for weekdays barring holidays and typically there are about 260
data points every single year.

6.1 USDJPY Data: tvARCH(1) Model

We obtain the following Figure 4 that shows our estimation for fitting a tvARCH(1)
model on the USD to JPY conversion data for the last 200, 500 and 1000 days ending
on 31 July 2020. The AMSE is also computed and contrasted with other competing
methods in Table 4. Figure 4 depicts the estimated functions with 95% credible bands
for different sample sizes. One can see the bands become much shorter for larger sample
sizes. The mean coefficient function μ(·) is generally time-varying for all three cases as
one cannot fit a horizontal line through the 95% posterior bands. There seems to be a
rise in the mean value around 100 days ago from July 31, 2020, which is right around
the time the COVID-19 pandemic hit the world. With the analysis of n = 1000 days,
we see that the volatility is quite high around October 2016 which coincides with the
presidential election time of 2016. The AR(1) coefficient does not show the huge time-
varying property. We also tabulate the AMSE for the three sample sizes in Table 4 and
one can see for smaller sample sizes such as n = 200, the proposed Bayesian tvARCH
achieves a significantly lower score but when the sample size grows then the performance
becomes similar.
ARCH(1) Frequentist tvARCH(1) Bayesian tvARCH(1)
n = 200 1.4572 1.2259 1.1712
n = 500 0.6281 0.5313 0.5218
n = 1000 0.4265 0.3773 0.3785
Table 4: AMSE comparison: tvARCH(1) model – USDJPY data.

6.2 NASDAQ Data: tvGARCH(1,1) Model

As has become standard in analyzing stock market datasets using GARCH models, we
use time-varying GARCH for small orders. We obtain the following Figure 5 for fitting
1174 Bayesian Modelling of Time-Varying Conditional Heteroscedasticity

Figure 4: USDJPY data (tvARCH(1) model) Estimated curve (black) along with the
95% pointwise credible bands (green) are shown for T = 200, 500, 1000 from top to
bottom.

a tvGARCH(1,1) model on the NASDAQ data for the last 200, 500, 1000 days ending
on 31 July 2020. One can see the a1 (·) values are generally low and the b1 (·) values
are higher which is consistent with how these outcomes turn out for time-constant esti-
mates for econometric datasets. One can also see the role sample size plays in curating
these time-varying estimates. For n = 200, the b1 (·) achieves high value of 0.6 around
mid-March 2020 but for higher sample sizes it shows values as high as 0.8. One can
also note the striking similarity for the analysis of the last 500 and 1000 days which is
S. Karmakar and A. Roy 1175

fairly consistent with the idea that estimation is more stable for such CH type models
with a higher sample size. Nonetheless, the estimates for n = 200 seem quite smooth as
well which can be seen as a benefit of our methodology. Table 5 provides a comparison
of AMSE scores across the three methods for three sample sizes. The Bayesian tv-
GARCH(1,1) performs relatively better than other methods and estimated curves have
smaller credible bands with a growing sample size. The behavior of the mean function
also shows higher volatility around the pandemic.

GARCH(1,1) Frequentist tvGARCH(1,1) Bayesian tvGARCH(1,1)

n = 200 203.5917 203.5917 202.6192
n = 500 104.7443 90.5395 90.3126
n = 1000 46.16759 46.9225 45.5618
Table 5: AMSE comparison: tvGARCH(1,1) model – NASDAQ data.

6.3 NASDAQ Data: tviGARCH(1,1) Model

In Figure 5 the sum of estimated coefficient functions a(·) + b(·) is close to 1 for a
significant time-horizon. This motivates us to also fit tviGARCH(1,1) to analyze the
same NASDAQ data. The estimated functions are presented in Figure 6 for the last
n = 200, 500 and 1000 days. Table 6 compares the AMSE scores for the same three
methods as before with varying sample sizes. The estimated mean and AR(1) functions
of Figure 6 change a little from the estimated functions of tvGARCH(1,1) fit in Figure 5.
Moreover, the effect of the three sample sizes is clear here with n = 1000 showing very
precise bands and can reveal an interesting time-varying pattern.
In terms of AMSE, one can see in Table 6 that the frequentist methods did worse than
even the time-constant versions. The time-constant estimates were computed using the
rugarch package in R. The Bayesian tviGARCH method provides significantly better
AMSE uniform overall sample sizes. Here the mean function also shows higher volatility
around the time when the pandemic struck us. Volatility due to the presidential election
in 2016 can also be observed here.

iGARCH(1,1) Frequentist tviGARCH(1,1) Bayesian tviGARCH(1,1)

n = 200 217.4988 278.4635 206.6886
n = 500 96.5001 132.544 90.4456
n = 1000 54.1171 260.4696 46.3704
Table 6: AMSE comparison: tviGARCH(1,1) model – Nasdaq data.

6.4 Model Comparison

For the analysis of NASDAQ data, we have used two different models and thus it
is pertinent to answer how should one choose between a competing class of models.
We provide some measures in this subsection to decide between these two competing
models. We start by comparing the performance of tvGARCH and tviGARCH models
1176 Bayesian Modelling of Time-Varying Conditional Heteroscedasticity

Figure 5: NASDAQ data (tvGARCH(1,1) model) Estimated curve (black) along with
the 95% pointwise credible bands (green) are shown for T = 200, 500, 1000 from top to
bottom.

in terms of Bayes factor (Kass and Raftery, 1995). Our calculation of the Bayes factor is
based on the posterior samples using the harmonic mean identity of Neton and Raftery
(1994). Let us denote B200 , B500 and B1000 as the Bayes factors for three sample sizes
where
P (D(i) |tvGARCH)
Bi = ,
P (D(i) )|tviGARCH)
S. Karmakar and A. Roy 1177

Figure 6: NASDAQ data (tviGARCH(1,1) model) Estimated curve (black) along with
the 95% pointwise credible bands (green) are shown for T = 200, 500, 1000 from top to
bottom.

for sample size i and the corresponding dataset D(i) . The values we obtain are
2 log(B200 ) = 8.16, 2 log(B500 ) = 19.08 and 2 log(B1000 ) = 24.14. According to guide-
lines from section 3.2 of Kass and Raftery (1995), there is ‘positive’ evidence in favor of
tvGARCH for sample sizes 200 and 500. However, the same evidence becomes ‘strong’
for sample size 1000.
1178 Bayesian Modelling of Time-Varying Conditional Heteroscedasticity

We also try to address out-of-sample predictive performance comparison here. Note

that for a time-varying GARCH or time-varying iGARCH model this is generally a
difficult task due to the assumed non-stationarity of the model. Thus, we take following
approach to calculate out of sample joint predictive log-likelihoods for model compar-
ison. Let us assume we have the data D(n) with n data points. To evaluate the joint
predictive log-likelihood for the last m(< n) most recent data points, we fit the models
in (2.4) and (2.6) with the first n − m data. Note that the assumed time horizons for
these two models are n. Based on the estimated B-spline coefficients, and other param-
eters from each model, we can compute the joint predictive log-likelihood of the last m
data points as

1 
n
1 
L(n)
m = −Xi2 /σ̂i2 − log(σ̂i ) − log(2π) ,
m i=n−m+1 2

where σ̂i2 = μ̂(i/n) + â1 (i/n)Xi−1

2 2
+ b̂1 (i/n)σ̂i−1 . For the tviGARCH model, we have
b̂1 (·) = 1 − â1 (·).
Using this predictive log-likelihood we decide to evaluate the two fits from tvGARCH
and tviGARCH in the following manner. For each of the sample sizes, we run it on three
separate regimes of the data, the full data, and two halves of the data. In all these 8
settings, (three sample sizes, three possible regions of the data, but the latest half of
the 1000-sized data is the same as the full data for sample size 500) we compute 10,
20, and 50 steps ahead forecast. We tabulate these results in Table 7. One can see that
generally speaking, there is somewhat conclusive evidence towards the iGARCH model
for a smaller sample size. This supports our motivation why we additionally provide a
tviGARCH(1,1) modeling on the same dataset.
Based on our model comparison exercises, we have an interesting phenomenon where
for in-sample model fit, tvGARCH is better but in terms of out-of-sample prediction,
tviGARCH outperforms tvGARCH in almost all the cases. Note that, tvGARCH has
one additional free parameter and thus is expected to fit the data better but since the
estimated a1 (·) and b1 (·) coefficients are close to one satisfying the iGARCH formulation,
the out-of-sample performance for tviGARCH may have exceeded that for tvGARCH.
As per the suggestion from a reviewer, we also add a one-step-ahead point forecasting
exercise between these models. Here the computation method remains the same as out-
lined in the predictive log-likelihood computation however we only restrict ourselves to
m = 1 to make the discussion concise. For this part of the exercise we choose to compute
posterior mean of (Xn2 − σ̂ 2 )2 where to ensure out-of-sample prediction σ̂ 2 is estimated
solely based on X1 , . . . , Xn−1 . As one-step-ahead forecasts can be prohibitively mislead-
ing given it depends so much on one single location, we decide to take an average of
over 15 random starting points over the entire time spectrum of 10 years resulting in
2518 points. For each of the sample sizes, we tabulate the performance in the following
Table 8. Note that, here we are only comparing the two Bayesian time-varying models
to see which one fits our data better. The advantage of predicting the future coefficients
using B-spline is not available in the kernel-based frequentist method and thus is not
included here in the discussion.
S. Karmakar and A. Roy 1179

Steps Full First Half Second (Latest) Half

n (m) GARCH iGARCH GARCH iGARCH GARCH iGARCH
200 10 −2.1 × 108 −1.982 −3.532 −3.154 −8.9 × 106 −2.251
20 −2.689 −1.881 −14.889 −2.475 −90281 −2.278
50 −3.640 −2.487 −86.81 −2.877 −5.839 −4.221
500 10 −2.842 −2.068 −10.260 −1.898 −1161 −2.079
20 −2.341 −1.848 −3897 −1.407 −343.49 −1.893
50 −2.147 −2.112 −44.381 −3.061 −932.7 −2.371
1000 10 −1.856 −1.936 −2.499 −2.790 −2.842 −2.068
20 −1.911 −1.789 −1.999 −1.903 −2.341 −1.848
50 −2.266 −2.214 −1.893 −1.536 −2.147 −2.112
Table 7: Joint log-likelihood for 10, 20, 50 steps ahead: Comparing tv-
GARCH(1,1)/tviGARCH(1,1) model – NASDAQ data. Better model is in bold.

Bayesian tvGARCH(1,1) Bayesian tviGARCH(1,1)

n = 200 0.9914 0.3477
n = 500 1.5208 1.4102
n = 1000 1.6378 1.7033
Table 8: One-step-ahead out-of-sample forecast for NASDAQ data: Comparing tv-
GARCH(1,1)/tviGARCH(1,1) model.

One can see, we again observe the same advantage of tviGARCH modeling over the
tvGARCH one for smaller sample sizes. This is an interesting find of this paper in the
context of the Bayesian model fitting to these datasets.

7 Discussion and Conclusion

In this paper, we consider a Bayesian estimation framework for time-varying condi-
tional heteroscedastic models. Our prior specifications are amenable to Hamiltonian
Monte Carlo for efficient computation. One of the key motivations towards going to a
Bayesian regime was to achieve reasonable estimation even for a small sample size. Our
simulation coverage shows good performance for all three models tvARCH, tvGARCH,
tviGARCH for both small and large sample sizes. Importantly, in all three of the cases,
we were able to establish posterior contraction rates. These calculations are, to the best
of our knowledge, the first such work in even simple dependent models let alone the
complicated recursions that these conditional heteroscedastic models demand. More-
over, the assumptions on the true functions and the number of moments needed were
very minimal. An interesting future theoretical work would be to calculate posterior
contraction rate with respect to empirical 2 -distance which is a more desirable metric
for function estimation. While analyzing real data, we see that the parameter curves
vary significantly for the intercept terms, but not that much for AR or CH parameters.
The associated codes to fit the three models are available at https://github.com/
royarkaprava/tvARCH-tvGARCH-tviGARCH.
1180 Bayesian Modelling of Time-Varying Conditional Heteroscedasticity

As future work, it will be interesting to explore multivariate time-varying volatil-

ity modeling (Tse and Tsui, 2002; Kwan et al., 2005) through a Bayesian framework
similar to ours. Another interesting time-heterogeneity that we plan to explore through
the glass of a Bayesian framework is regime-switching CH models where instead of the
smooth time-varying functions the parameters change abruptly. We have a brief discus-
sion in Section 6.4 about how to choose between competing models. Those discussions
can easily be extended to choose a proper number of lags or to choose between different
types of ARCH/GARCH models. We believe this would provide an interesting paral-
lel to the usual penalized likelihood-based methods for model selection in time-series.
Finally note that, even though we do provide some insights onto future prediction for
these datasets for real data applications, that was not the main focus in this paper.
Forecasting for the time-varying model is extremely tricky since it requires ‘estimation’
of the future parameter values. Although in-filled asymptotics can help in this regard,
still the literature so far is very sparse in this direction for both Bayesian and frequentist
regimes. We plan to explore this extensively in near future.

Supplementary Material
Proof of Theorems (DOI: 10.1214/21-BA1267SUPP; .pdf). The supplementary material
includes the proof of Theorems 1, 2 and 3 and a general discussion of the main strategy
behind them. We also include the traceplots for the MCMC chain from our simulations.

References
Amorim, L. D., Cai, J., Zeng, D., and Barreto, M. L. (2008). Regression splines in the
time-dependent coefficient rates model for recurrent event data. Statistics in Medicine,
27(28):5890–5906. MR2597750. doi: https://doi.org/10.1002/sim.3400. 1159
Andreou, E. and Ghysels, E. (2006). Monitoring disruptions in financial markets. Jour-
nal of Econometrics, 135(1-2):77–124. MR2328397. doi: https://doi.org/10.1016/
j.jeconom.2005.07.023. 1158
Andrews, D. W. K. (1993). Tests for parameter instability and structural change with
unknown change point. Econometrica, 61(4):821–856. MR1231678. doi: https://
doi.org/10.2307/2951764. 1158
Audrino, F. and Bühlmann, P. (2009). Splines for financial volatility. Journal of
the Royal Statistical Society: Series B (Statistical Methodology), 71(3):655–670.
MR2749912. doi: https://doi.org/10.1111/j.1467-9868.2009.00696.x. 1159
Bai, J. (1997). Estimation of a change point in multiple regression models. The Review
of Economics and Statistics, 79(4):551–563. 1157
Betancourt, M. (2017). A conceptual introduction to Hamiltonian Monte Carlo.
arXiv preprint arXiv:1701.02434. MR1699395. doi: https://doi.org/10.1017/
CBO9780511470813.003. 1159
S. Karmakar and A. Roy 1181

Betancourt, M. and Girolami, M. (2015). Hamiltonian Monte Carlo for hierarchi-

cal models. Current Trends in Bayesian Methodology with Applications, 79(30):2–4.
MR3644666. 1159
Biller, C. and Fahrmeir, L. (2001). Bayesian varying-coefficient models using adaptive
regression splines. Statistical Modelling, 1(3):195–211. 1159
Bollerslev, T. (1986). Generalized autoregressive conditional heteroskedasticity. Jour-
nal of Econometrics, 31(3):307–327. MR0853051. doi: https://doi.org/10.1016/
0304-4076(86)90063-1. 1158
Bose, A. and Mukherjee, K. (2003). Estimating the arch parameters by solving
linear equations. Journal of Time Series Analysis, 24(2):127–136. MR1965808.
doi: https://doi.org/10.1111/1467-9892.00296. 1168
Brown, R. L., Durbin, J., and Evans, J. M. (1975). Techniques for testing the constancy
of regression relationships over time. Journal of the Royal Statistical Society: Series
B (Statistical Methodology), 37:149–192. With discussion by D. R. Cox, P. R. Fisk,
Maurice Kendall, M. B. Priestley, Peter C. Young, G. Phillips, T. W. Anderson, A.
F. M. Smith, M. R. B. Clarke, A. C. Harvey, Agnes M. Herzberg, M. C. Hutchison,
Mohsin S. Khan, J. A. Nelder, Richard E. Quant, T. Subba Rao, H. Tong and W. G.
Gilchrist and with reply by J. Durbin and J. M. Evans. MR0378310. 1158
Cai, Z. (2007). Trending time-varying coefficient time series models with serially cor-
related errors. Journal of Econometrics, 136(1):163–188. MR2328589. doi: https://
doi.org/10.1016/j.jeconom.2005.08.004. 1158
Cai, Z., Fan, J., and Yao, Q. (2000). Functional-coefficient regression models for non-
linear time series. Journal of the American Statistical Association, 95(451):941–956.
MR1804449. doi: https://doi.org/10.2307/2669476. 1159
Chen, J. and Gupta, A. K. (1997). Testing and locating variance changepoints with ap-
plication to stock prices. Journal of the American Statistical Association, 92(438):739–
747. MR1467863. doi: https://doi.org/10.2307/2965722. 1158
Chow, G. C. (1960). Tests of equality between sets of coefficients in two linear re-
gressions. Econometrica, 28:591–605. MR0141193. doi: https://doi.org/10.2307/
1910133. 1158
Dahlhaus, R. and Polonik, W. (2009). Empirical spectral processes for locally station-
ary time series. Bernoulli, 15(1):1–39. MR2546797. doi: https://doi.org/10.3150/
08-BEJ137. 1159
Dahlhaus, R. and Subba Rao, S. (2006). Statistical inference for time-varying ARCH
processes. Annals of Statistics, 34(3):1075–1114. MR2278352. doi: https://doi.org/
10.1214/009053606000000227. 1159, 1161
Engle, R. F. (1982). Autoregressive conditional heteroscedasticity with estimates of the
variance of United Kingdom inflation. Econometrica, 50(4):987–1007. MR0666121.
doi: https://doi.org/10.2307/1912773. 1158
1182 Bayesian Modelling of Time-Varying Conditional Heteroscedasticity

Engle, R. F. and Rangel, J. G. (2005). The spline garch model for unconditional volatility
and its global macroeconomic causes. 1158
Engle, R. F. and Rangel, J. G. (2008). The spline-GARCH model for low-frequency
volatility and its global macroeconomic causes. The Review of Financial Studies,
21(3):1187–1222. 1159
Fan, J. and Zhang, W. (1999). Statistical estimation in varying coefficient models.
Annals of Statistics, 27(5):1491–1518. MR1742497. doi: https://doi.org/10.1214/
aos/1017939139. 1157
Fan, J. and Zhang, W. (2000). Simultaneous confidence bands and hypothesis test-
ing in varying-coefficient models. Scandinavian Journal of Statistics, 27(4):715–731.
MR1804172. doi: https://doi.org/10.1111/1467-9469.00218. 1157
Fan, J. and Zhang, W. (2008). Statistical methods with varying coefficient models.
Statistics and its Interface, 1(1):179. MR2425354. doi: https://doi.org/10.4310/
SII.2008.v1.n1.a15. 1159
Franco-Villoria, M., Ventrucci, M., Rue, H., et al. (2019). A unified view on
Bayesian varying coefficient models. Electronic Journal of Statistics, 13(2):5334–5359.
MR4047589. doi: https://doi.org/10.1214/19-EJS1653. 1159
Fryzlewicz, P., Sapatinas, T., and Subba Rao, S. (2008a). Normalized least-squares
estimation in time-varying ARCH models. Annals of Statistics, 36(2):742–786.
MR2396814. doi: https://doi.org/10.1214/07-AOS510. 1158, 1159, 1168
Fryzlewicz, P., Sapatinas, T., and Subba Rao, S. (2008b). Normalized least-squares
estimation in time-varying ARCH models. Annals of Statistics, 36(2):742–786.
MR2396814. doi: https://doi.org/10.1214/07-AOS510. 1161, 1173
Ghosal, S. and Van der Vaart, A. (2017). Fundamentals of nonparametric Bayesian
inference, volume 44. Cambridge University Press. MR3587782. doi: https://doi.
org/10.1017/9781139029834. 1165
Gu, C. and Wahba, G. (1993). Smoothing spline anova with component-wise Bayesian
“confidence intervals”. Journal of Computational and Graphical Statistics, 2(1):97–
117. MR1272389. doi: https://doi.org/10.2307/1390957. 1159
Hastie, T. and Tibshirani, R. (1993). Varying-coefficient models. Journal of the Royal
Statistical Society: Series B (Methodological), 55(4):757–779. MR1229881. 1159
Hoover, D. R., Rice, J. A., Wu, C. O., and Yang, L.-P. (1998). Nonparamet-
ric smoothing estimates of time-varying coefficient models with longitudinal data.
Biometrika, 85(4):809–822. MR1666699. doi: https://doi.org/10.1093/biomet/
85.4.809. 1157, 1158
Huang, J. Z. and Shen, H. (2004). Functional coefficient regression models for non-
linear time series: a polynomial spline approach. Scandinavian Journal of Statistics,
31(4):515–534. MR2101537. doi: https://doi.org/10.1111/j.1467-9469.2004.
00404.x. 1159
S. Karmakar and A. Roy 1183

Huang, J. Z., Wu, C. O., and Zhou, L. (2002). Varying-coefficient models

and basis function approximations for the analysis of repeated measurements.
Biometrika, 89(1):111–128. MR1888349. doi: https://doi.org/10.1093/biomet/
89.1.111. 1159
Huang, J. Z., Wu, C. O., and Zhou, L. (2004). Polynomial spline estimation and inference
for varying coefficient models with longitudinal data. Statistica Sinica, 14(3):763–788.
MR2087972. 1157
James Chu, C.-S. (1995). Detecting parameter shift in garch models. Econometric Re-
views, 14(2):241–266. 1158
Jeong, S. (2019). Frequentist properties of Bayesian procedures for high-dimensional
sparse regression. MR4094193. 1159
Karmakar, S. (2018). Asymptotic theory for simultaneous inference under dependence.
Technical report, University of Chicago. 1159
Karmakar, S. and Roy, A. (2021). “Supplementary Material of “Bayesian Modelling
of Time-Varying Conditional Heteroscedasticity”.” Bayesian Analysis. doi: https://
doi.org/10.1214/21-BA1267SUPP. 1160
Karmakar, S., Richter, S., and Wu, W. B. (2021). Simultaneous inference for
time-varying models. Journal of Econometrics, doi: https://doi.org/10.1016/j.
jeconom.2021.03.002. 1159, 1161, 1168, 1170, 1172
Kass, R. E. and Raftery, A. E. (1995). Bayes factors. Journal of the American Sta-
tistical Association, 90(430):773–795. MR3363402. doi: https://doi.org/10.1080/
01621459.1995.10476572. 1176, 1177
Kim, S., Cho, S., and Lee, S. (2000). On the cusum test for parameter changes in garch
(1, 1) models. Communications in Statistics – Theory and Methods, 29(2):445–462.
MR1749743. doi: https://doi.org/10.1080/03610920008832494. 1158
Kokoszka, P., Leipus, R., et al. (2000). Change-point estimation in ARCH mod-
els. Bernoulli, 6(3):513–539. MR1762558. doi: https://doi.org/10.2307/3318673.
1158
Kulperger, R., Yu, H., et al. (2005). High moment partial sum processes of residu-
als in garch models and their applications. Annals of Statistics, 33(5):2395–2422.
MR2211090. doi: https://doi.org/10.1214/009053605000000534. 1158
Kwan, C., Li, W., and Ng, K. (2005). A multivariate threshold garch model with time-
varying correlations. Econometric Reviews. 1180
Leybourne, S. J. and McCabe, B. P. M. (1989). On the distribution of some test statistics
for coefficient constancy. Biometrika, 76(1):169–177. MR0991435. doi: https://doi.
org/10.1093/biomet/76.1.169. 1158
Lin, C.-F. J. and Teräsvirta, T. (1999). Testing parameter constancy in linear models
against stochastic stationary parameters. Journal of Econometrics, 90(2):193–213.
MR1703341. doi: https://doi.org/10.1016/S0304-4076(98)00041-4. 1158
1184 Bayesian Modelling of Time-Varying Conditional Heteroscedasticity

Lin, D. Y. and Ying, Z. (2001). Semiparametric and nonparametric regression analysis

of longitudinal data. Journal of the American Statistical Association, 96(453):103–
126. With comments and a rejoinder by the authors. MR1952726. doi: https://doi.
org/10.1198/016214501750333018. 1157
Lin, S.-J., Yang, J., et al. (1999). Testing shifts in financial models with conditional
heteroskedasticity: an empirical distribution function approach. School of Finance and
Economics, University of Technology, Sydney. 1158
Liu, R. and Yang, L. (2016). Spline estimation of a semiparametric garch model.
Econometric Theory, 32(4):1023. MR3530459. doi: https://doi.org/10.1017/
S0266466615000055. 1159
Livingstone, S., Betancourt, M., Byrne, S., and Girolami, M. (2019). On the geometric
ergodicity of Hamiltonian Monte Carlo. Bernoulli, 25(4A):3109–3138. MR4003576.
doi: https://doi.org/10.3150/18-BEJ1083. 1159
Nabeya, S. and Tanaka, K. (1988). Asymptotic theory of a test for the constancy
of regression coefficients against the random walk alternative. Annals of Statis-
tics, 16(1):218–235. MR0924867. doi: https://doi.org/10.1214/aos/1176350701.
1158
Neal, R. M. et al. (2011). Mcmc using hamiltonian dynamics. Handbook of Markov
Chain Monte Carlo, 2(11):2. MR2858447. 1159, 1164
Neton, M. and Raftery, A. (1994). Approximate Bayesian inference by the weighted
likelihood bootstrap (with discussion). Journal of the Royal Statistical Society: Series
B (Statistical Methodology), pages 1–48. MR1257793. 1176
Ning, B., Jeong, S., Ghosal, S., et al. (2020). Bayesian linear regression for mul-
tivariate responses under group sparsity. Bernoulli, 26(3):2353–2382. MR4091112.
doi: https://doi.org/10.3150/20-BEJ1198. 1159, 1165
Nyblom, J. (1989). Testing for the constancy of parameters over time. Journal of the
American Statistical Association, 84(405):223–230. MR0999682. 1158
Ploberger, W., Krämer, W., and Kontrus, K. (1989). A new test for structural stability
in the linear regression model. Journal of Econometrics, 40(2):307–318. MR0994952.
doi: https://doi.org/10.1016/0304-4076(89)90087-0. 1158
Ramsay, J. O. and Silverman, B. W. (2005). Functional data analysis. Springer Series
in Statistics. Springer, New York, second edition. MR2168993. 1157
Rohan, N. (2013). A time varying garch (p, q) model and related statistical infer-
ence. Statistics & Probability Letters, 83(9):1983–1990. MR3079033. doi: https://
doi.org/10.1016/j.spl.2013.04.030. 1159
Rohan, N. and Ramanathan, T. V. (2013). Nonparametric estimation of a time-varying
GARCH model. Journal of Nonparametric Statistics, 25(1):33–52. MR3039969.
doi: https://doi.org/10.1080/10485252.2012.728600. 1159, 1168
Starica, C. (2003). Is GARCH (1, 1) as good a model as the accolades of the Nobel
prize would imply? Available at SSRN 637322. 1168
S. Karmakar and A. Roy 1185

Stărică, C. and Granger, C. (2005). Nonstationarities in stock returns. The Review of

Economics and Statistics, 87(3):503–522. 1158
Tse, Y. K. and Tsui, A. K. C. (2002). A multivariate generalized autoregressive con-
ditional heteroscedasticity model with time-varying correlations. Journal of Business
& Economic Statistics, 20(3):351–362. MR1939906. doi: https://doi.org/10.1198/
073500102288618496. 1180
Yue, Y. R., Simpson, D., Lindgren, F., Rue, H., et al. (2014). Bayesian adaptive smooth-
ing splines using stochastic differential equations. Bayesian Analysis, 9(2):397–424.
MR3217001. doi: https://doi.org/10.1214/13-BA866. 1159
Zhang, T. and Wu, W. B. (2012). Inference of time-varying regression models. An-
nals of Statistics, 40(3):1376–1402. MR3015029. doi: https://doi.org/10.1214/
12-AOS1010. 1158
Zhang, T. and Wu, W. B. (2015). Time-varying nonlinear regression models: non-
parametric estimation and model selection. Annals of Statistics, 43(2):741–768.
MR3319142. doi: https://doi.org/10.1214/14-AOS1299. 1158
Zhang, W., Lee, S.-Y., and Song, X. (2002). Local polynomial fitting in semivary-
ing coefficient model. Journal of Multivariate Analysis, 82(1):166–188. MR1918619.
doi: https://doi.org/10.1006/jmva.2001.2012. 1157
Zhou, Z. and Wu, W. B. (2010). Simultaneous inference of linear models with time vary-
ing coefficients. Journal of the Royal Statistical Society. Series B. (Statistical Method-
ology), 72(4):513–531. MR2758526. doi: https://doi.org/10.1111/j.1467-9868.
2010.00743.x. 1158

Acknowledgments
We would like to thank the editor, the associate editor, and two anonymous referees for their
constructive suggestions that improved the quality of the manuscript.