Computational Economics

[Link]

GARCHNet: Value‑at‑Risk Forecasting with GARCH Models

Based on Neural Networks

Mateusz Buczynski1,2 · Marcin Chlebus2

Accepted: 4 April 2023

© The Author(s) 2023

Abstract
This paper proposes a new GARCH specification that adapts the architecture of a
long-term short memory neural network (LSTM). It is shown that classical GARCH
models generally give good results in financial modeling, where high volatility can
be observed. In particular, their high value is often praised in Value-at-Risk. How-
ever, the lack of nonlinear structure in most approaches means that conditional vari-
ance is not adequately represented in the model. On the contrary, the recent rapid
development of deep learning methods is able to describe any nonlinear relation-
ship in a clear way. We propose GARCHNet, a nonlinear approach to conditional
variance that combines LSTM neural networks with maximum likelihood estima-
tors in GARCH. The variance distributions considered in the paper are normal, t
and skewed t, but the approach allows extension to other distributions. To evaluate
our model, we conducted an empirical study on the logarithmic returns of the WIG
20 (Warsaw Stock Exchange Index), S&P 500 (Standard & Poor’s 500) and FTSE
100 (Financial Times Stock Exchange) indices over four different time periods from
2005 to 2021 with different levels of observed volatility. Our results confirm the
validity of the solution, but we provide some directions for its further development.

Keywords Value-at-risk · GARCH · Neural networks · LSTM

* Mateusz Buczynski
[Link]@[Link]
Marcin Chlebus
mchlebus@[Link]
1
Faculty of Economic Sciences, University of Warsaw, Dluga 44/50, Warsaw, Poland
2
Interdisciplinary Doctoral School, University of Warsaw, Dobra 56/66, Warsaw, Poland

13
Vol.:(0123456789)
 M. Buczynski, M. Chlebus

1 Introduction

Uncertainty in financial markets has been a main point of risk-related research

for decades (Segal et al., 2015; Vorbrink, 2014). The market standard, which was
established more than 30 years ago for use as a measure of risk, is Value-at-Risk
(VaR) (Duffie & Pan, 1997). It is the simplest way to express potential losses over
a target time horizon with a specified statistical certainty. The simplicity of VaR
does not stop numerous approaches from being proposed (Engle & Manganelli,
2004; Barone-Adesi et al., 2008; Wang et al., 2010). Such scientific abundance
dictates that various approaches to calculating VaR can be in use and still be con-
sidered “good.” Whether a model can be referred to as qualitatively good is a
point of debate among many researchers in the field (Abad et al., 2014; Nozari
et al., 2010; Ergün & Jun, 2010; Degiannakis et al., 2012; Escanciano & Olmo,
2010; Abad & Benito, 2013). Even when models have been thoroughly tested
and found to be statistically valid, insufficient attention is still paid to temporal
changes in financial time series characteristics, which can lead to overestimation
or underestimation of risk. The best example of such situation is the financial
crisis of 2008 (Degiannakis et al., 2012) or the more recent market crash caused
by COVID-19 (Omari et al., 2020). Therefore the financial industry—both regu-
lators and financial institutions—are turning to a better, probabilistic way of esti-
mating risk based on past events that is able to quickly adjust to recent shocks (So
& Philip, 2006). As of writing, the official estimate of market risk is either Value
at Risk or Expected Shortfall (ES) proposed by Basel Committee. It estimates the
expected value of a potential loss if such a loss on a given asset is less than VaR.
One of the most influential drivers of risk is variance, particularly its changing
temporal structure or tendency to cluster (Cont, 2002). There exists a broad fam-
ily of models that aim to capture such effect, the most common being General-
ized Autoregressive Conditional Heteroskedasticity (GARCH) model, proposed
by Bollerslev (1986). Since financial markets typically exhibit known stylized
facts, a more fitting approach is to use a fat tail distribution (Aloui & Mabrouk,
2010). The introduction of distributions such as t-distributions or GEDs, which
allow for modeling skewness and heavy tails, has dispelled any doubts about the
validity of GARCH models (BenSaïda, 2015; Bonato, 2012). Another extension
of these models, proposed by Francq and Zakoïan (2004), assumes that not only
the variance exhibits temporal changes, but also the mean, but in terms of finan-
cial returns, the mean is usually insignificant in the long run (Fama, 1998).
On the other hand, financial researchers are much keen on implementing
machine learning methods (Sezer et al., 2020). Deep neural networks (deep NNs)
are considered a good substitute for conventional statistical methods, not only in
the field of financial markets, but also in other areas of science (Mnih et al., 2013;
Devlin et al., 2018; Cho et al., 2014). However, for time series data, recursive
approaches such as NNs with long short-term memory (LSTM) are preferable
(Goodfellow et al., 2016). In addition, NNs offer a nonlinear estimator of the like-
lihood function (Chen & Billings, 1992). For GARCH models, the conditional
variance function usually assumes a linear or very simple nonlinear relationship

13
GARCHNet: Value‑at‑Risk Forecasting with GARCH Models Based…

between the likelihood function (and also the moments of the distribution) and
the observables (Glosten et al., 1993a; Nelson & Cao, 1992).
According to Lim et al. (2019), the best approach to using machine learning in
the time series domain is not to fully replace statistical and econometric approaches.
Rather, they propose to combine the best of both worlds, hence the idea of this paper
is to model conditional variance using NN. Several studies have already been pro-
duced on the intersection of GARCH and NN models. For example, Arnerić et al.
(2014) have proposed modeling time series using the GARCH model, but with an
extension to RNNs called Jordan NNs. Similar studies by Kristjanpoller and Minu-
tolo (2015, 2016) propose an ANN-GARCH model and their results show a 25%
reduction in mean absolute percentage error (MAPE). Research by Kim and Won
(2018) oes a step further, incorporating an LSTM layer into the neural network,
reporting a 37.2% decrease in mean absolute error (MAE). Yet another approach,
proposed by Jeong and Lee (2019) considers the RNN model to determine the
autoregressive moving average (ARMA) process, which drives not the conditional
variance, but the conditional mean. Their results reveal that this approach leads to a
reduction in MAPE of about 10%.
The aforementioned studies, however, do not specifically focus on the implemen-
tation of NNs for conditional variance alone. For example, studies by Kristjanpoller
and Minutolo (2015, 2016) use GARCH estimates of variability as inputs to the NN
model, while Kim and Won (2018) build NNs with covariates that are parameters of
artificially generated GARCH models. Our approach leans toward estimating con-
ditional moments of an assumed distribution using NNs, such as in Rothfuss et al.
(2019). The first to propose such approach were Nikolaev et al. (2011), who investi-
gated an approach with recursive NNs (RNN) to represent conditional variance and
found that incorporating nonlinear methods (RNN-GARCH) reduces model uncer-
tainty. Further on, Liu and So (2020) consider using the LSTM NN to model con-
ditional variance directly through the maximum likelihood approach of the density
function of the assumed distribution. They showed that this method can successfully
determine both the standard deviation and variance of financial returns. Another
advantage of their approach is that it can use explained artificial intelligence (XAI)
methods. However, instead of using estimation, they assumed the values of addi-
tional (in addition to the first and second moments) parameters of the distribution.
Another research by Nguyen et al. (2019) proposes a fairly similar approach, but to a
stochastic volatility (SV) model, which is related to GARCH. In their research, they
propose an SV-LSTM model that uses LSTM NN instead of using the AR(1) pro-
cess to model volatility. Their results indicate that the proposed approach can give
better out-of-sample estimates than standard SV models.
In this paper, we propose GARCHNet— a conditional specification of NN-based
GARCH models with extensive use of the LSTM layer. Our incentives are based on
the previously raised drawbacks of GARCH and the fact that the LSTM NN is able
to adequately represent any non-linear relationships found in financial time series
data. We also extend previous research in this area by proposing further distribu-
tions—we propose a GARCHNet with normal, t and skewed t distributions, and pro-
vide the necessary negative log likelihood functions for all of them, which can be
used as cost functions in NN back-propagation optimization algorithms.

13
 M. Buczynski, M. Chlebus

We also propose an empirical experiment to verify the usefulness of the GARCH-

Net model. The experiment consists of estimating Value-at-Risk forecasts one day
ahead in a window of 250 test days (approximately one trading year) and comparing
them with equivalent GARCH models. The experiment was conducted on logarith-
mic returns of the WIG 20 index (Warsaw Stock Exchange Index; Poland), S&P 500
(Standard and Poor’s 500; the USA) and FTSE 100 (Financial Time Stock Exchange;
the UK) over four different time periods (both training and test sample) from 2005
to 2021. The experiment was written and conducted in Python and pytorch (Paszke
et al., 2019).
The paper is organized as follows. In Sect. 2, we present the theoretical back-
ground of GARCHNet and the necessary background for VAR backtesting. In
Sect. 3, we describe the empirical experiment with data and model descriptions. In
Sect. 4, we present the results of the experiment, and in Sect. 5 we include conclud-
ing remarks and paths for extending our framework in future research.

2 Methodology

2.1 GARCH Models

GARCH model with no mean (pure GARCH process) can be specified as:
rt = 𝜇t + 𝜖t ,
(1)
𝜖t = 𝜎t zt ,

where rt is observed time series, 𝜇t is conditional mean of the process and 𝜎t is the
conditional standard deviation of the observed time series process. zt is an innova-
tion process and is considered to be i.i.d with unit variance, in the most straightfor-
ward approach the assumed distribution is normal: zt ∼ N(0, 1).
Many definitions of conditional variance have already been proposed in the VaR
field: standard GARCH (Bollerslev, 1986), Exponential GARCH (EGARCH) (Nel-
son, 1991), Integrated GARCH (IGARCH) (Engle & Bollerslev, 1986) or Glosten-
Jagannathan-Runkle GARCH (GJR-GARCH) (Glosten et al., 1993b). However, in
this paper, we only utilize standard GARCH(p, q) process, which defines conditional
volatility as:
∑
q
∑
p
𝜎t2 = 𝜔 + (2)
2 2
𝛽i 𝜖t−i + 𝛾i 𝜎t−i ,
i=1 i=1

where p and q are numbers of lags of conditional variance and innovation respec-
tively, 𝛽 and 𝛾 are parameter vectors to be estimated. The of stationarity of the
∑q ∑p
GARCH process is satisfied by the fact that i=1 𝛽i + i=1 𝛾i < 1.
As for optimizing this process, one possible procedure is to use quasi maximum
likelihood (QML). Given that the innovations are assumed to be independent, the
conditional log likelihood of a vector of demeaned observed time series 𝜖 of length

13
GARCHNet: Value‑at‑Risk Forecasting with GARCH Models Based…

T can be defined as the sum of all log conditional densities of particular innovations
𝜖t (see Francq & Zakoïan, 2004):

∑
T
∑
T
∑
T
𝓁(𝜃𝜃 ;𝜖𝜖 ) = 𝓁t (𝜃𝜃 ;𝜖t ) = logf (𝜖t |𝜖t−1 , … , 𝜖1 ;𝜃𝜃 ) = logf (𝜖t ;𝜃𝜃 ), (3)
t=1 t=1 t=1

where 𝜃 = (𝜔, 𝛽1 , … , 𝛽q , 𝛾1 , … , 𝛾p ) is a vector of parameters, f (𝜖t |𝜖t−1 , … , 𝜖1 ;𝜃𝜃 ) is

conditional density function of innovation 𝜖t , however, given that the innovations are
independent it will reduce to f (𝜖t ;𝜃𝜃 ).
Quasi maximum likelihood estimation of the parameters vector 𝜃 is a solution 𝜃̂
of:

𝜃̂ = arg max 𝓁(𝜃𝜃 , 𝜖 ) (4)

𝜃

1∑
T
𝓁(𝜃𝜃 , 𝜖 ) = 𝓁 (𝜃𝜃 , 𝜖t ) (5)
T t=1 t

In the case zt is normally distributed, conditional log likelihood function for one
observation is equal to:
2
1 1 𝜖t
𝓁t (𝜃𝜃 , 𝜖t ) = − log𝜎t2 − , (6)
2 2 𝜎t2

which comes down to a logarithm of a normal density function.

In the case zt is t distributed, an additional parameter is necessary to be estimated
- 𝜂 - number of degrees of freedom of this distribution, with an assumption of 𝜂 > 2.
Therefore the parameter vector is 𝜃 = (𝜔, 𝛽1 , … , 𝛽q , 𝛾1 , … , 𝛾p , 𝜂), and conditional
log likelihood for one observation is:
( ) ( )
𝜂+1 (𝜂 ) 𝜂+1 𝜖2
1
𝓁t (𝜃𝜃 , 𝜖t ) = log Γ − log Γ − log(𝜋(𝜂 − 2)𝜎t2 ) − log 1 + 2 t ,
2 2 2 2 𝜎t (𝜂 − 2)
(7)
where Γ(⋅) is a gamma function and the log likelihood is a logarithm of density of t
distribution.
In the last case, we treat zt as skewed t distributed. One more parameter is intro-
duced -𝜆, responsible for the skewness of the distribution. A particular analytical
implementation of the skewed t-distribution was proposed by the Hansen (1994).
In this case, an additional assumption is that −1 < 𝜆 < 1. Parameter vector is once
again extended to 𝜃 = (𝜔, 𝛽1 , … , 𝛽q , 𝛾1 , … , 𝛾p , 𝜂, 𝜆).
� � �2 �−(𝜂+1)∕2 ⎤
⎡
bc 1 a + bx∕𝜎
𝓁t = ln ⎢ 1+ ⎥, (8)
⎢𝜎 𝜂 − 2 1 + sgn(x∕𝜎 + a∕b)𝜆 ⎥
⎣ ⎦

where

13
 M. Buczynski, M. Chlebus

� �
𝜂−2 Γ 𝜂+1
2
a = 4𝜆c , b2 = 1 + 3𝜆2 − a2 , c= √ � �, (9)
𝜂−1 𝜋(𝜂 − 2)Γ 𝜂2

All of the log likelihood functions are numerically obtainable. In addition, the spe-
cific form of the conditional variance does not affect the QML in the above form. It
is much more influenced by the assumed distribution. This opens up the possibility
of using much more complicated nonlinear forms, such as NN (Goodfellow et al.,
2016).

2.2 LSTM Neural Networks

Long Short Term Memory (LSTM) neural networks are an extension of recurrent neu-
ral networks (RNNs), proposed by Rumelhart et al. (1986). RNNs are a special type
of neural networks that introduce recursion by allowing the use of sequential, autocor-
related data. The sequence (or observed time series) is accompanied by a hidden input,
a kind of memory state that stores information provided with previous time steps. The
next input in the sequence is predicted using this recursive hidden state:
ht = g(Wx xt + Wh ht−1 + bh ), (10)
where g(⋅) is an activation function (e.g., logistic sigmoid, hyperbolic tangent or
Rectified Linear Unit (ReLU)), x = (x1 , x2 , … , xT ) is the sequence of observed time
series of length T, while h = (h1 , h2 , … , hT ) represents a random vector—hidden
state of the same length T. Wx and Wh are weight matrices (parameters) of the neural
network, corresponding to x and h respectively and bh is a bias vector. Such equa-
tion assumes that the sequence can be of infinite length or at least an arbitrarily large
number T, but due to computational obstacles (such as the problem of vanishing or
exploding gradients (Pascanu et al., 2012)) the sequence length T is practically lim-
ited to only a few timesteps.
The problem mentioned above is practically solved by introduction of LSTM
(Hochreiter & Schmidhuber, 1997). LSTMs expand the idea of hidden states by intro-
ducing gating mechanisms, which tell whether to preserve or ignore the input from the
hidden state. Given that, LSTMs can “remember” or “forget” particular timesteps if
necessary, building the long-term dependency parameter matrix. In detail, there are
three gates: forget, input and output.
The following equations calculated iteratively build up LSTM network:
it =g(Wix xt + Wih ht−1 + Wic ct−1 + bi ) (11)

ft =g(Wfx xt + Wfh ht−1 + Wfc ct−1 + bf ), (12)

ct =ft ⊙ ct−1 + it ⊙ tanh(Wcx xt + Wch ht−1 + bc ), (13)

13
GARCHNet: Value‑at‑Risk Forecasting with GARCH Models Based…

ot =g(Wox xt + Woh ht−1 + Woc ct + bo ), (14)

ht =ot ⊙ h(ct ), (15)

yt =Wyh ht + by , (16)
where W terms denote weight matrices (e.g.: Wix is a matrix of weights from the
input gate to the input x), the b terms denote bias vectors (e.g. bi is the input gate
bias vector), g(⋅) and h(⋅) denote sigmoid and hyperbolic tangent activation functions
respectively here, i, f and o denote input, forget and output gates respectively, ct is
another hidden state vector, specifically named cell activation vector (responsible for
activating specific gates). The output of the neural network can be any distribution
p(yy|xx;𝜃𝜃 ), however most often some particular moment of this distribution is esti-
mated directly—in our case we would like it to be conditional variance.

2.3 GARCHNet

Our idea of a GARCH process specification is to use a neural network as an approxi-

mation of the true conditional variance specification. To optimize the NN, the likeli-
hood functions described in Sect. 2.1 come to our aid. They are used as cost func-
tions—in negative log likelihoods form. In GARCHNet, the GARCH specification is
as follows:
ln = Wln ln−1 ln−1 + bln (17)
and
l1 = Wly yt + bl1 , (18)
where yt is calculated as in Eq. 16 and n determines the number of following fully
connected layers. Given that, conditional variance is the function of the last n fully
connected layer:

𝜎t2 = g(WVl ln + bV ), (19)

where g(⋅) is a function with non-negative output (e.g. softplus), while WVl is a
matrix of weights from the last hidden layer to the output layer and bV denotes its
bias. The input of such an LSTM neural network is p of the last observed realiza-
tions of the time series (selected earlier). Its output will be an estimate of the condi-
tional variance. Because of the specific mechanism that drives the forgetting mecha-
nism of LSTM layers, we do not have to worry that the sequence that is fed into
the model may be too long. NNs are typically optimized using a backpropagation
algorithm (Goodfellow et al., 2016), which includes calculating gradients for each
neuron in the layer and then iteratively applying changes in weights based on the
value of the cost function.

13
 M. Buczynski, M. Chlebus

However, in the density functions of t and skewed t distributions, there are two addi-
tional parameters that are necessary to be estimated or assumed. In our scenario these
parameters are estimated with the same NN as a function of time. Therefore degrees of
freedom 𝜂 and skewness 𝜆 are estimated as:
𝜂 = g(WEl ln + bE ) + 2, (20)

𝜆 = h(WSl ln + bS ), (21)
where g(⋅) is a function with non-negative output (e.g. softplus), h(⋅) is a function
with output in the range (−1, 1), while W are matrices of weights from the last hid-
den layer to the particular output layer (degrees of freedom 𝜂 and skewness 𝜆 respec-
tively) and b vectors denote their biases. Please note that we are adding two units
to the output of degrees of freedom 𝜂 to meet the assumption that 𝜂 > 2. A comple-
mentary approach would imply changes in the log likelihood function.
This means that in the most advanced scenario, for skewed t distribution, there are
three last hidden layers (one for conditional variance 𝜎t2, one for degrees of freedom 𝜂
and one for skewness 𝜆), each resulting in one different output neuron.
Originally, conditional variance’s parameters (𝜔, 𝛽 and 𝛾 ) should be non-negative
(Bollerslev, 1986), which together with non-negativity of random variables (𝜎t2 and z2t )
suffices for the conditional variance to be non-negative as well. In the case of neural
network, such assumption could lead to the worsening of the accuracy of estimated
solution (Chorowski & Zurada, 2014). Instead of using such limitation, we have pro-
posed to use softplus function (or any other that outputs non-negative values and is eas-
ily differentiable). Softplus function is defined as:
Softplus(x) = log(1 + exp(x)). (22)
In the case of skewness we have proposed to use hyperbolic tangent function so that
the output meets the assumption that −1 < 𝜆 < 1. Hyperbolic tangent function is
defined as:
exp(x) − exp(−x)
tanh(x) = . (23)
exp(x) + exp(−x)

2.4 Value‑at‑Risk

Value-at-Risk (VaR) defines the worst possible loss with a given probability 𝛼, assum-
ing normal market conditions for a specific time period t (Philippe, 2006). In other
words, VaR is a quantile of the distribution of the observed financial time series. In our
case, these are log returns of the price quotations of the respective stock index.
P(rt < VaR𝛼 (t)|Ωt−1 ) = 𝛼, (24)
where rt is the realization of the observed financial time series and Ωt−1 is an infor-
mation set given at the time t − 1.

13
GARCHNet: Value‑at‑Risk Forecasting with GARCH Models Based…

When parametric models are employed, such as GARCH, VaR is calculated as

an 𝛼 quantile of the assumed innovation distribution—F −1 (𝛼) (F is inverse cumu-
lative distribution function of the assumed innovation distribution), weighted by
the estimate of the conditional standard deviation 𝜎t , plus an estimate of condi-
tional mean 𝜇t (Angelidis et al., 2004):

VaR𝛼 = 𝜇t + 𝜎t F −1 (𝛼). (25)

2.4.1 Quality of VaR Forecasts

The primary tool for assessing the quality of the VaR forecast is the number of
cases in which the VaR forecast was lower (in absolute terms) than the realiza-
tion of the observed time series—excess count or proportion of failures (Chlebus,
2017):

1∑
N
𝛼̂ = I , (26)
N t=1 VaR𝛼 >rt
∑N
where N is the number of testing instances and I
t=1 VaR𝛼 >rt
is the
number of exceedances = n.
Statistically, this number comes from a binomial distribution (assuming the
exceptions are IID). The Basel Committee strictly regulates what values consti-
tute a “safe zone” or require a look at the model. Specifically, the name of such a
test is the Traffic Light Test (Costanzino & Curran, 2018). In the case of VaR at
2.5% significance level and 250 testing instances the ’safe’ (green) zone ends with
10 exceptions (95% cumulative probability) and yellow (warning zone) ends with
16 exceptions (99.99% cumulative probability).
The unconditional coverage (UC) test by Kupiec (1995) builds up on the idea
that the overall number of exceptions should follow the binomial distribution. To
test that a likelihood ratio test is proposed:
( )
(1 − 𝛼)N−n 𝛼 n
LRUC = −2ln . (27)
̂ N−n 𝛼̂ n
(1 − 𝛼)

There is also a conditional coverage (CC) test by Christoffersen (1998). In addition

to the unconditional coverage, Christoffersen test measures the likelihood of unusu-
ally frequent VaR exceptions—an effect of exceptions clustering.
The conditional coverage test consists of both unconditional coverage and
independence tests: LRcc = LRuc + LRind . The independence test LRind verifies
whether the exceptions follow a first-order Markov chain.
Even more restrictive is dynamic quantile (DQ) test by Engle and Manganelli
(2004). They define another random variable Hitt = It − 𝛼 . The null hypothesis
of this test is that the expected value of the Hitt explained with the information
available at t − 1 is zero. To test that, they implement a linear regression model:

13
 M. Buczynski, M. Chlebus

∑
K
Hitt = 𝛿 + 𝛽k Xt−k + 𝜖t , (28)
k=1

where matrix X might include both lags of Hit, r or VaR. DQ test statistic is then:

Hit� X(X � X)−1 X � Hit

DQ = (29)
𝛼(1 − 𝛼)
Another interesting dimension of model comparison are loss functions (LF). Their
value determines the loss if the model fails. There are two parties who are usually
interested in these values—the regulator and the companies themselves. Both weigh
certain business aspects differently. From the regulator’s point of view, the most
important aspect is the value lost on the occurrence of a VaR exception, while from
the company’s point of view, the opportunity cost of holding excess reserves.
To compare the models we have chosen following loss functions, from the pro-
posed by Abad et al. (2015):

• Lopez quadratic LF (LLF):

{
1 + (VaRt − rt )2 if rt < VaRt ,
LLFt =
0 otherwise; (30)

• Caporin regulator’s LF (CRLF):

{
|1 − |rt ∕VaRt || if rt < VaRt ,
CLFt =
0 otherwise; (31)

• Caporin firm’s LF (CFLF):

CFLFt = |1 − |rt ∕VaRt || for all rt ; (32)

• Abad, Benito, Lopez’s LF (ABLLF):
{
(VaRt − rt )2 if rt < VaRt ,
ABLLFt =
𝛽(rt − VaRt ) otherwise, (33)

where 𝛽 is a parameter that represents a cost of capital, originally an interest

rate.
Gneiting (2011) also suggests to specify a scoring function for single-valued point
forecasts, such as the ones that we generate in this paper. Specifically, for 𝛼-quantile
forecasts he proposes to use a generalized piecewise linear (GPL) scoring function,
in a form of:
{ 1
(1(xt ≥ yt ) − 𝛼) |b| (xtb − ybt ) if b ∈ ℝ ⧵ {0},
S𝛼,b,t = x (34)
(1(xt ≥ yt ) − 𝛼)log yt if b = 0,
t

where x is a vector of predictions and y is a vector of realized rates of return. In

case b = 1, we end up with asymmetric piecewise linear scoring function, which we

13
GARCHNet: Value‑at‑Risk Forecasting with GARCH Models Based…

propose to use in this paper. The overall result for a model is a sum for all the test
cases.

3 Data and Model Specifications

3.1 Data

GARCHNet’s performance was measured empirically by backtesting on the log

returns of price quotations of WIG20 (Warsaw Stock Exchange; Poland), S&P 500
(New York Stock Exchange, the USA) and FTSE 100 (London Stock Exchange, the
UK). Therefore our observed time series is:
rt = logpt − logpt−1 . (35)
Such data is openly available, e.g.: from Stooq (2021). As a reference, we also esti-
mated the corresponding GARCH models on the same data samples.
We subjectively selected four different time periods consisting of 1250 obser-
vations each (1000 was our training window length and we covered 250 test sam-
ples). The start dates of specific periods are as follows: (i) 2005-01-01 (testing on
year 2009), (ii) 2007-01-01 (testing on year 2011), (iii) 2013-01-01 (testing on year
2017), (iv) 2016-01-01 (testing on year 2020). In our opinion these periods provide
a full view of possible volatility levels between training window and predicted sam-
ple- training and testing samples both show low volatility (sample starting in 2013
respectively) or the volatility is different for training and testing samples (low vola-
tility training samples starting in 2005 and 2016); and high volatility training sample
starting in 2007). Such a spectrum allows us to test the model in varying market
conditions.

3.2 Models

We compared the proposed GARCHNet specifications with the corresponding

standard GARCH models. To do this, we also had to select the number of observa-
tions p, which is the sequence for the LSTM model. Because of the similarity to the
original meaning of p in the GARCH model (the number of lagged conditional vari-
ances in its model), we controlled both of these parameters with p. The test included
p ∈ 5, 10, 20, 100. In addition, we have reported the results for GARCH models
with p ∈ 1, 2 assuming that a comparison with our results should also provide infor-
mation on how the new models behave in relation to standard approach. We set the 𝛼
significance level for VaR estimates at 2.5%. We also set the random seed equal to 1.
We used a rolling-window estimation approach (Zanin & Marra, 2012). For each
forecast sample, we prepared a new model with new randomly initialized weights
and trained it using the last 1000 observations. We have also tested a hypothesis that
frequent updates of the model might not necessarily improve its quality, while only
increasing the time overhead in training. We tested a framework where the model
was fully reset (random weights fully initialized) less often than with each timestep

13
 M. Buczynski, M. Chlebus

forecast. The model might be refitted with fresh data, between resets to include new
information. In the most extreme approach we assumed that it is only trained fully
once (on the first 1000 observations) and then we have increased the frequency of
training up to 500 updates (update very other training sample). Such approach has
been proven faulty in results comparison, due to large jumps in volatility estimates.
The results are not reported here.
For the neural network determining the conditional variance, we used a rather
small architecture (see Fig. 1): one LSTM layer with 100 neurons (fed by a sequence
of length p), followed by three (n = 3) fully connected layers with 64, 32 and 1
neuron(s), respectively. For t and skew t distributions, there were two (and three,
respectively) output layers corresponding to the number of parameters being opti-
mized. Parameter optimization was performed using the Adam optimizer with a
learning rate of 3e-4 and a batch size of 512. Due to the rolling-window method,
it was difficult to choose an automatic threshold for the number of epochs to avoid
overfitting, so each model was trained for 300 epochs.

4 Results

The results of the experiment are satisfactory. Figure 2 shows the relationship
between GARCH and GARCHnet predictions with innovations with a t distribu-
tion. It can be seen that the GARCHNet predictions do not deviate from the rate of
return, even more—for some intervals GARCHNet confirms the presence of vola-
tility shocks much faster. It can also be noted that the GARCHNet model tends to

Fig. 1  The diagram with the architecture of GARCHNet model

13
GARCHNet: Value‑at‑Risk Forecasting with GARCH Models Based…

Fig. 2̄  Exemplary comparison for the GARCH and GARCHNet with t distribution for p = 20 for WIG
20. Note: logarithmic rate of return (in blue); GT (in orange)—GARCH with t distributed innovations;
GNT (in green)—GARCHNet with t distributed innovations. (Color figure online)

estimate a higher VaR than GARCH, except for the most recent period, where the
relationship is reversed.
In the Tables 1, 2 and 3 we have presented the results of the statistical tests and
the number of exceptions for each index tested. The results are mostly the same for
all indexes, but the biggest difference is seen for the WIG 20. There is no GARCH
model that outperforms its GARCHNet counterpart across all periods tested and for
all p sequence lengths tested in terms of number of exceptions. However, GARCH-
Net with a t-distribution appears to have the largest excess. In the case of the WIG
20, for only two cases was the number of exceptions higher than for GARCH with a
t distribution, for the S&P 500 it was five cases and for the FTSE 100 nine cases. It
should be noted that most of these exceedances occurred in the last two periods. For
the other models, the overhead is much smaller and sometimes negative. However,
we believe that the predictive power of such a model could be improved with a bet-
ter neural architecture.
GARCHNet with a skewed t-distribution is worse by a small margin, which is not
consistent with its GARCH counterpart—GARCH with a skewed t-distribution is
the best model compared to other members of its family. This may indicate that the
distribution parameter estimation approach is inefficient for Adam’s optimizer, or
that the approach we used should be reconsidered. For example, the parameter esti-
mation should be changed to an estimation for the entire training sample, rather than
based on a fairly small sample (of length p) of time series in the prediction phase.
GARCHNet with a normal distribution tends to be worse than its standard counter-
part, but in the last two periods this relationship is much smaller. We assume that
this is due to the worse predictive power of the GARCH model in turbulent periods.

13
 Table 1  Results of GARCH and GARCHNet models regarding p values of considered statistical tests for WIG20
p Period I (2009) Period II (2011) Period III (2017) Period IV (2020)
Model H UC CC DQ GPL H UC CC DQ GPL H UC CC DQ GPL H UC CC DQ GPL

13
1 GN 7.00 0.76 0.10 0.38 0.32 5.00 0.61 0.01 0.79 0.23 1.00 0.01 0.62 0.03 0.14 33.00 0.00 0.00 0.00 0.55
GT 10.00 0.16 0.00 0.06 0.37 6.00 0.93 0.04 0.30 0.26 2.00 0.05 0.36 0.13 0.14 29.00 0.00 0.00 0.00 0.52
GS 9.00 0.29 0.00 0.06 0.41 6.00 0.93 0.05 0.30 0.27 2.00 0.05 0.36 0.13 0.14 27.00 0.00 0.00 0.00 0.53
2 GN 9.00 0.29 0.00 0.06 0.35 9.00 0.29 0.00 0.35 0.28 1.00 0.01 0.62 0.03 0.13 27.00 0.00 0.00 0.00 0.55
GT 11.00 0.08 0.00 0.05 0.37 8.00 0.49 0.00 0.39 0.29 6.00 0.93 0.00 0.30 0.16 23.00 0.00 0.00 0.00 0.50
GS 8.00 0.49 0.00 0.05 0.37 8.00 0.49 0.00 0.39 0.30 7.00 0.76 0.00 0.03 0.17 21.00 0.00 0.00 0.00 0.49
5 GN 10.00 0.16 0.00 0.26 0.39 9.00 0.29 0.25 0.35 0.27 1.00 0.01 0.62 0.03 0.13 18.00 0.00 0.00 0.00 0.48
GT 10.00 0.16 0.00 0.06 0.39 9.00 0.29 0.26 0.35 0.27 1.00 0.01 0.61 0.03 0.13 16.00 0.00 0.00 0.00 0.44
GS 9.00 0.29 0.00 0.35 0.39 10.00 0.16 0.00 0.06 0.27 1.00 0.01 0.61 0.03 0.13 15.00 0.00 0.00 0.01 0.42
GNN 13.00 0.02 0.00 0.02 0.35 4.00 0.33 0.09 0.08 0.26 0.00 – – – 0.13 18.00 0.00 0.00 0.00 0.48
GNT 5.00 0.61 0.24 0.18 0.36 1.00 0.01 0.60 0.03 0.30 0.00 – – – 0.17 15.00 0.00 0.00 0.00 0.46
GNS 6.00 0.93 0.99 0.86 0.33 1.00 0.01 0.62 0.03 0.29 0.00 – – – 0.16 14.00 0.01 0.00 0.00 0.46
10 GN 7.00 0.76 0.00 0.78 0.36 11.00 0.08 0.00 0.05 0.26 1.00 0.01 0.62 0.03 0.13 13.00 0.02 0.00 0.00 0.53
GT 7.00 0.76 0.00 0.78 0.37 11.00 0.08 0.00 0.05 0.26 1.00 0.01 0.62 0.03 0.13 13.00 0.02 0.00 0.00 0.49
GS 6.00 0.93 0.13 0.86 0.37 10.00 0.16 0.00 0.06 0.27 1.00 0.01 0.62 0.03 0.13 13.00 0.02 0.00 0.00 0.50
GNN 9.00 0.29 0.11 0.35 0.33 9.00 0.29 0.31 0.35 0.29 2.00 0.05 0.69 0.13 0.13 18.00 0.00 0.00 0.00 0.46
GNT 5.00 0.61 0.26 0.18 0.33 4.00 0.33 0.92 0.59 0.26 0.00 – – – 0.17 17.00 0.00 0.00 0.00 0.43
GNS 6.00 0.93 0.99 0.86 0.32 5.00 0.61 0.26 0.18 0.27 0.00 – – – 0.16 17.00 0.00 0.00 0.00 0.46
20 GN 8.00 0.49 0.17 0.39 0.37 8.00 0.49 0.19 0.39 0.27 1.00 0.01 0.62 0.03 0.13 13.00 0.02 0.00 0.00 0.51
GT 7.00 0.76 0.18 0.38 0.37 8.00 0.49 0.24 0.39 0.28 1.00 0.01 0.61 0.03 0.13 15.00 0.00 0.00 0.00 0.48
GS 8.00 0.49 0.00 0.39 0.39 8.00 0.49 0.24 0.39 0.28 1.00 0.01 0.61 0.03 0.13 15.00 0.00 0.00 0.00 0.47
GNN 10.00 0.16 0.01 0.06 0.34 9.00 0.29 0.33 0.35 0.26 3.00 0.15 0.63 0.34 0.14 13.00 0.02 0.00 0.05 0.43
GNT 6.00 0.93 0.32 0.30 0.35 5.00 0.61 0.97 0.79 0.26 0.00 – – – 0.17 14.00 0.01 0.00 0.01 0.42
GNS 0.28 2.00 0.05 0.15 16.00 0.00 0.00 0.00 0.51
M. Buczynski, M. Chlebus

3.00 0.15 0.93 0.34 0.32 4.00 0.33 0.09 0.08 0.75 0.13
 Table 1  (continued)
p Period I (2009) Period II (2011) Period III (2017) Period IV (2020)
Model H UC CC DQ GPL H UC CC DQ GPL H UC CC DQ GPL H UC CC DQ GPL

100 GN 4.00 0.33 0.98 0.59 0.36 17.00 0.00 0.00 0.00 0.32 4.00 0.33 0.08 0.59 0.12 14.00 0.01 0.00 0.00 0.54
GT 3.00 0.15 0.94 0.34 0.36 17.00 0.00 0.00 0.00 0.32 2.00 0.05 0.82 0.13 0.12 14.00 0.01 0.00 0.00 0.54
GS 2.00 0.05 0.80 0.13 0.36 15.00 0.00 0.00 0.00 0.33 3.00 0.15 0.01 0.34 0.13 14.00 0.01 0.00 0.00 0.54
GNN 11.00 0.08 0.08 0.17 0.34 5.00 0.61 0.25 0.18 0.28 1.00 0.01 0.62 0.03 0.12 15.00 0.00 0.00 0.00 0.43
GNT 4.00 0.33 0.04 0.08 0.36 1.00 0.01 0.61 0.03 0.35 0.00 – – – 0.16 14.00 0.01 0.00 0.01 0.41
GNS 6.00 0.93 0.98 0.86 0.32 5.00 0.61 0.28 0.79 0.30 3.00 0.15 0.77 0.34 0.15 15.00 0.00 0.00 0.00 0.46

GN—GARCH with normally distributed innovations (d.i.); GT—GARCH with t d.i.; GST—GARCH with skewed t d.i.; GNN—GARCHNet with normally d.i.; GNT—
GARCHNet with t d.i.; GNS—GARCHNet with skewed t d.i.; H—number of hits within test period; UC—unconditional coverage test; CC—conditional coverage test;
DQ—dynamic quantile test; GPL—generalized piecewise linear scoring function
All statistical tests that were failed to be rejected at the 5% significance level are in bold. Best Hit and GPL value are also in bold
GARCHNet: Value‑at‑Risk Forecasting with GARCH Models Based…

13
 Table 2̄  Results of GARCH and GARCHNet models regarding p values of considered statistical tests for S&P 500
p Period I (2009) Period II (2011) Period III (2017) Period IV (2020)
Model H UC CC DQ GPL H UC CC DQ GPL H UC CC DQ GPL H UC CC DQ GPL

13
1 GN 24.00 0.00 0.00 0.00 0.34 7.00 0.76 0.00 0.78 0.27 3.00 0.15 0.00 0.34 0.08 21.00 0.00 0.00 0.00 0.32
GT 23.00 0.00 0.00 0.00 0.34 7.00 0.76 0.00 0.78 0.26 3.00 0.15 0.01 0.34 0.08 24.00 0.00 0.00 0.00 0.34
GS 21.00 0.00 0.00 0.00 0.32 5.00 0.61 0.00 0.79 0.23 1.00 0.01 0.34 0.03 0.08 25.00 0.00 0.00 0.00 0.34
2 GN 22.00 0.00 0.00 0.00 0.31 8.00 0.49 0.00 0.60 0.26 2.00 0.05 0.02 0.13 0.07 20.00 0.00 0.00 0.00 0.31
GT 20.00 0.00 0.00 0.00 0.31 8.00 0.49 0.01 0.60 0.25 1.00 0.01 0.24 0.03 0.08 17.00 0.00 0.00 0.00 0.28
GS 19.00 0.00 0.00 0.00 0.32 7.00 0.76 0.00 0.78 0.25 1.00 0.01 0.50 0.03 0.09 17.00 0.00 0.00 0.00 0.31
5 GN 18.00 0.00 0.00 0.00 0.34 12.00 0.04 0.00 0.10 0.25 3.00 0.15 0.14 0.34 0.08 13.00 0.02 0.00 0.03 0.29
GT 18.00 0.00 0.00 0.00 0.34 10.00 0.16 0.00 0.26 0.25 3.00 0.15 0.38 0.34 0.08 13.00 0.02 0.00 0.05 0.27
GS 17.00 0.00 0.00 0.00 0.32 9.00 0.29 0.00 0.35 0.24 1.00 0.01 0.59 0.03 0.09 5.00 0.61 0.99 0.79 0.21
GNN 10.00 0.16 0.02 0.24 0.29 9.00 0.29 0.35 0.35 0.24 4.00 0.33 0.93 0.59 0.09 21.00 0.00 0.00 0.00 0.45
GNT 1.00 0.01 0.58 0.03 0.36 2.00 0.05 0.00 0.13 0.29 2.00 0.05 0.77 0.13 0.11 7.00 0.76 0.10 0.38 0.40
GNS 6.00 0.93 0.11 0.30 0.30 2.00 0.05 0.82 0.13 0.25 3.00 0.15 0.84 0.34 0.11 14.00 0.01 0.00 0.01 0.46
10 GN 14.00 0.01 0.01 0.01 0.32 9.00 0.29 0.00 0.41 0.25 3.00 0.15 0.57 0.34 0.08 12.00 0.04 0.00 0.06 0.28
GT 14.00 0.01 0.01 0.01 0.32 11.00 0.08 0.00 0.13 0.25 3.00 0.15 0.59 0.34 0.08 3.00 0.15 0.94 0.34 0.22
GS 13.00 0.02 0.02 0.03 0.31 7.00 0.76 0.00 0.78 0.24 1.00 0.01 0.60 0.03 0.09 3.00 0.15 0.94 0.34 0.22
GNN 9.00 0.29 0.15 0.41 0.29 13.00 0.02 0.00 0.02 0.25 5.00 0.61 1.00 0.79 0.09 17.00 0.00 0.00 0.00 0.46
GNT 1.00 0.01 0.57 0.03 0.31 3.00 0.15 0.01 0.34 0.27 3.00 0.15 0.92 0.34 0.10 7.00 0.76 0.10 0.38 0.40
GNS 6.00 0.93 0.13 0.86 0.29 3.00 0.15 0.01 0.34 0.26 5.00 0.61 0.69 0.79 0.11 15.00 0.00 0.00 0.01 0.48
20 GN 11.00 0.08 0.20 0.13 0.30 11.00 0.08 0.00 0.17 0.27 6.00 0.93 0.87 0.86 0.09 6.00 0.93 0.42 0.86 0.24
GT 12.00 0.04 0.07 0.06 0.30 10.00 0.16 0.00 0.26 0.27 8.00 0.49 0.83 0.60 0.09 3.00 0.15 0.94 0.34 0.26
GS 10.00 0.16 0.17 0.24 0.30 7.00 0.76 0.00 0.78 0.25 6.00 0.93 0.81 0.86 0.09 3.00 0.15 0.94 0.34 0.24
GNN 11.00 0.08 0.08 0.13 0.29 13.00 0.02 0.00 0.02 0.26 5.00 0.61 1.00 0.79 0.09 15.00 0.00 0.00 0.00 0.42
GNT 3.00 0.15 0.89 0.34 0.33 2.00 0.05 0.00 0.13 0.28 1.00 0.01 0.61 0.03 0.11 9.00 0.29 0.24 0.35 0.39
GNS 4.00 0.29 8.00 0.28 4.00 0.10 14.00 0.01 0.00 0.00 0.51
M. Buczynski, M. Chlebus

0.33 0.97 0.59 0.49 0.06 0.39 0.33 0.96 0.59

 Table 2̄  (continued)
p Period I (2009) Period II (2011) Period III (2017) Period IV (2020)
Model H UC CC DQ GPL H UC CC DQ GPL H UC CC DQ GPL H UC CC DQ GPL

100 GN 7.00 0.76 0.18 0.78 0.32 14.00 0.01 0.00 0.01 0.31 9.00 0.29 0.24 0.41 0.09 10.00 0.16 0.00 0.26 0.26
GT 7.00 0.76 0.19 0.78 0.32 14.00 0.01 0.00 0.01 0.31 8.00 0.49 0.31 0.60 0.09 13.00 0.02 0.00 0.05 0.28
GS 7.00 0.76 0.21 0.78 0.33 13.00 0.02 0.00 0.03 0.30 8.00 0.49 0.36 0.60 0.09 19.00 0.00 0.00 0.00 0.29
GNN 13.00 0.02 0.01 0.03 0.28 12.00 0.04 0.00 0.10 0.25 5.00 0.61 1.00 0.79 0.09 23.00 0.00 0.00 0.00 0.50
GNT 0.00 – – – 0.38 2.00 0.05 0.00 0.13 0.30 1.00 0.01 0.62 0.03 0.11 8.00 0.49 0.16 0.39 0.43
GNS 6.00 0.93 0.54 0.86 0.31 4.00 0.33 0.07 0.59 0.29 3.00 0.15 0.90 0.34 0.10 17.00 0.00 0.00 0.00 0.51

GN—GARCH with normally distributed innovations (d.i.); GT—GARCH with t d.i.; GST—GARCH with skewed t d.i.; GNN—GARCHNet with normally d.i.; GNT—
GARCHNet with t d.i.; GNS—GARCHNet with skewed t d.i.; H—number of hits within test period; UC—unconditional coverage test; CC—conditional coverage test;
DQ—dynamic quantile test; GPL—generalized piecewise linear scoring function
All statistical tests that were failed to be rejected at the 5% significance level are in bold. Best Hit and GPL value are also in bold
GARCHNet: Value‑at‑Risk Forecasting with GARCH Models Based…

13
 Table 3  Results of GARCH and GARCHNet models regarding p values of considered statistical tests for FTSE 100
p Period I (2009) Period II (2011) Period III (2017) Period IV (2020)
Model H UC CC DQ GPL H UC CC DQ GPL H UC CC DQ GPL H UC CC DQ GPL

13
1 GN 11.00 0.07 0.02 0.15 0.22 3.00 0.15 0.36 0.34 0.17 4.00 0.34 0.45 0.59 0.09 22.00 0.00 0.00 0.00 0.25
GT 11.00 0.07 0.03 0.15 0.23 3.00 0.15 0.39 0.34 0.17 2.00 0.05 0.78 0.14 0.09 14.00 0.01 0.00 0.02 0.23
GS 7.00 0.71 0.10 0.76 0.22 3.00 0.15 0.35 0.34 0.18 2.00 0.05 0.18 0.14 0.08 11.00 0.08 0.06 0.17 0.22
2 GN 12.00 0.03 0.00 0.03 0.21 4.00 0.33 0.82 0.59 0.18 3.00 0.15 0.23 0.34 0.09 21.00 0.00 0.00 0.00 0.26
GT 11.00 0.07 0.00 0.04 0.21 3.00 0.15 0.63 0.34 0.18 3.00 0.15 0.50 0.34 0.09 10.00 0.16 0.00 0.24 0.23
GS 6.00 0.98 0.03 0.86 0.18 3.00 0.15 0.60 0.34 0.19 3.00 0.15 0.56 0.34 0.09 8.00 0.49 0.00 0.60 0.24
5 GN 17.00 0.00 0.00 0.00 0.25 5.00 0.61 0.22 0.18 0.19 3.00 0.15 0.20 0.34 0.08 16.00 0.00 0.00 0.00 0.22
GT 17.00 0.00 0.00 0.00 0.25 5.00 0.61 0.22 0.18 0.19 3.00 0.15 0.68 0.34 0.09 12.00 0.04 0.01 0.10 0.22
GS 10.00 0.14 0.10 0.24 0.24 3.00 0.15 0.73 0.34 0.19 3.00 0.15 0.61 0.34 0.09 7.00 0.76 0.07 0.38 0.23
GNN 13.00 0.02 0.04 0.05 0.27 9.00 0.29 0.01 0.06 0.24 7.00 0.76 0.15 0.78 0.11 21.00 0.00 0.00 0.00 0.43
GNT 0.00 – – – 0.30 2.00 0.05 0.79 0.13 0.23 3.00 0.15 0.68 0.34 0.12 14.00 0.01 0.00 0.01 0.40
GNS 6.00 0.93 0.93 0.86 0.24 3.00 0.15 0.94 0.34 0.24 3.00 0.15 0.82 0.34 0.12 13.00 0.02 0.00 0.05 0.41
10 GN 17.00 0.00 0.00 0.00 0.24 6.00 0.93 0.98 0.86 0.18 3.00 0.15 0.76 0.34 0.09 7.00 0.76 0.55 0.78 0.21
GT 16.00 0.00 0.00 0.00 0.24 4.00 0.33 0.91 0.59 0.18 4.00 0.34 0.78 0.59 0.09 4.00 0.33 0.09 0.59 0.21
GS 7.00 0.71 0.49 0.76 0.22 3.00 0.15 0.73 0.34 0.19 4.00 0.34 0.90 0.59 0.09 2.00 0.05 0.77 0.13 0.22
GNN 10.00 0.16 0.44 0.24 0.25 10.00 0.16 0.01 0.06 0.22 6.00 0.93 0.22 0.86 0.11 17.00 0.00 0.00 0.00 0.38
GNT 2.00 0.05 0.81 0.13 0.29 4.00 0.33 0.93 0.59 0.22 4.00 0.33 0.94 0.59 0.11 13.00 0.02 0.00 0.02 0.35
GNS 5.00 0.61 0.91 0.79 0.26 3.00 0.15 0.92 0.34 0.23 5.00 0.61 0.36 0.79 0.13 12.00 0.04 0.00 0.03 0.34
20 GN 8.00 0.45 0.77 0.57 0.22 5.00 0.61 0.20 0.18 0.18 3.00 0.15 0.66 0.34 0.08 7.00 0.76 0.96 0.78 0.22
GT 8.00 0.45 0.88 0.57 0.21 5.00 0.61 0.20 0.18 0.18 4.00 0.34 0.95 0.59 0.09 6.00 0.93 0.99 0.86 0.22
GS 4.00 0.36 0.99 0.62 0.22 3.00 0.15 0.01 0.02 0.18 4.00 0.34 0.98 0.59 0.09 4.00 0.33 0.98 0.59 0.23
GNN 12.00 0.04 0.04 0.06 0.26 10.00 0.16 0.00 0.01 0.21 8.00 0.49 0.07 0.60 0.11 19.00 0.00 0.00 0.00 0.39
GNT 7.00 0.76 0.89 0.78 0.27 4.00 0.33 0.99 0.59 0.21 2.00 0.05 0.78 0.13 0.11 12.00 0.04 0.01 0.10 0.35
GNS 9.00 0.26 4.00 0.22 4.00 0.13 15.00 0.00 0.00 0.01 0.38
M. Buczynski, M. Chlebus

0.29 0.07 0.41 0.33 0.98 0.59 0.33 0.60 0.59

 Table 3  (continued)
p Period I (2009) Period II (2011) Period III (2017) Period IV (2020)
Model H UC CC DQ GPL H UC CC DQ GPL H UC CC DQ GPL H UC CC DQ GPL

100 GN 4.00 0.36 0.99 0.62 0.25 11.00 0.08 0.02 0.05 0.21 15.00 0.00 0.00 0.01 0.10 7.00 0.76 0.14 0.78 0.22
GT 4.00 0.36 0.99 0.62 0.25 11.00 0.08 0.02 0.05 0.21 14.00 0.01 0.00 0.02 0.10 8.00 0.49 0.12 0.60 0.23
GS 2.00 0.05 0.84 0.15 0.25 8.00 0.49 0.60 0.39 0.21 12.00 0.04 0.01 0.10 0.10 4.00 0.33 0.97 0.59 0.22
GNN 15.00 0.00 0.00 0.00 0.28 7.00 0.76 0.56 0.38 0.21 8.00 0.49 0.03 0.60 0.10 24.00 0.00 0.00 0.00 0.43
GNT 5.00 0.61 0.96 0.79 0.28 3.00 0.15 0.86 0.34 0.24 2.00 0.05 0.81 0.13 0.12 13.00 0.02 0.00 0.00 0.40
GNS 6.00 0.93 0.89 0.86 0.26 2.00 0.05 0.81 0.13 0.23 3.00 0.15 0.75 0.34 0.13 12.00 0.04 0.00 0.01 0.41

GN—GARCH with normally distributed innovations (d.i.); GT—GARCH with t d.i.; GST—GARCH with skewed t d.i.; GNN—GARCHNet with normally d.i.; GNT—
GARCHNet with t d.i.; GNS—GARCHNet with skewed t d.i.; H—number of hits within test period; UC—unconditional coverage test; CC—conditional coverage test;
DQ—dynamic quantile test; GPL—generalized piecewise linear scoring function
All statistical tests that were failed to be rejected at the 5% significance level are in bold. Best Hit and GPL value are also in bold
GARCHNet: Value‑at‑Risk Forecasting with GARCH Models Based…

13
 M. Buczynski, M. Chlebus

We have also prepared results for GARCH models with p ∈ 1, 2. We note that for
each GARCHNet model there is a p that will provide results better than the stand-
ard GARCH approach. This is most apparent for periods with a large discrepancy
between the level of variability in the training and test samples. GARCHnet models
have typically from 5 to 1 exceedances fewer than GARCH models with p ∈ 1, 2.
Let us focus on the first two periods: starting in 2005 and 2007. Both of these
periods show a high number of failures to reject the null hypothesis of the tests con-
sidered regardless of the model tested, but we can see that the GARCHNet mod-
els have better results there (the largest differences for S&P 500). The number of
exceedances for GARCHNet do not show any outstanding features, but we note
that the DQ test was not rejected much more often than for the standard GARCH
approach. The non-linear structure of the proposed conditional variance may not be
fully explained by the linear structure of the DQ test and the similar, linearly struc-
tured standard GARCH models might be outperforming NN here. In terms of statis-
tical tests, the GARCHNet approach appears to provide models of generally higher
quality. It can also be noted that the GPL statistic indicating the best predictions fell
in favour of the GARCHNet models as being better in 14 out of 24 cases. The worst
forecasts were made for the FTSE 100 index, where only 1 GARCHNet forecast was
better. As for the GARCH(1,1) or GARCH(2,2) benchmark, it had the worst GPL in
2009, while the best in 2011.
Let us now turn to the samples starting in 2013 and 2016. In these two cases, we
see a clearly higher number of rejections of the null hypotheses—both due to under-
estimation and overestimation of risk. In these two periods, however, the results of
the GARCHNet models are in line with those of the standard GARCHs, with a slight
tendency to underestimate risk (in 2016 it was mainly the GARCH models that were
not rejected for the null hypothesis). On average, GARCHNet models have very sim-
ilar number of exceptions. Both model families were not able to respond correctly to
the COVID-19 financial market crashes, hence the high number of exceptions in the
last analyzed period. It should be noted that the COVID 19 period should be seen as
a stress-test for these models, and given the very similar performance of GARCH-
Net we would like to emphasise that it performs well under all conditions.
We note that the sequence length has a non-linear effect on the quality of the
model. This depends primarily on the variability of the sample used for prediction,
mainly for the standard GARCH model—see the outstanding exceptions for p = 100
in the sample starting in 2007 and the much numbers values for the other p values.
This effect weakens in the case of GARCHNet, but we still notice large discrepan-
cies for different values of p. Regarding the proposed length of p, we would suggest
20, which represents four trading weeks—one trading month and therefore gave the
most remarkable results.
In Tables 4, 5 and 6 we have presented the cost function values. We note that due
to the lower number of exceptions of the GARCHNet models, the cost function val-
ues of the regulator are lower than for its counterparts in most of the analyzed cases,
moreover—the worst GARCHNet approaches are often better than the best GARCH
in turbulent periods, while the GARCH models are better in calm periods. This is a
very desirable feature of a VaR model, as in the case of an exception the potential
loss is not as severe. However, from the company’s point of view, the GARCHNet

13
 Table 4  Results of GARCH and GARCHNet models regarding the values of cost functions for WIG 20
p Period I (2009) Period II (2011) Period III (2017) Period IV (2020)
Model LLF CRLF CFLF ABLLF LLF CRLF CFLF ABLLF LLF CRLF CFLF ABLLF LLF CRLF CFLF ABLLF

1 GN 7.00 1.34 149.35 0.11 5.00 0.76 171.17 0.08 1.00 0.06 165.27 0.05 33.02 28.33 186.14 0.07
GT 10.00 2.37 152.50 0.11 6.00 0.92 175.43 0.09 2.00 0.94 167.80 0.05 29.01 21.64 182.63 0.07
GS 9.00 1.98 164.05 0.13 6.00 1.11 175.40 0.09 2.00 0.98 168.68 0.05 27.01 27.54 187.65 0.07
2 GN 9.00 2.40 151.26 0.11 9.00 2.46 171.50 0.08 1.00 0.00 164.26 0.05 27.02 23.99 183.62 0.07
GT 11.00 3.14 150.53 0.11 8.00 2.45 173.58 0.08 6.00 6.76 171.11 0.05 23.01 14.50 168.64 0.08
GS 8.00 1.66 158.44 0.12 8.00 2.66 173.46 0.08 7.00 8.19 173.16 0.05 21.01 16.12 173.97 0.08
5 GN 10.00 4.36 150.33 0.11 9.00 2.37 170.55 0.08 1.00 0.08 161.26 0.05 18.01 11.81 163.84 0.08
GT 10.00 4.30 150.16 0.10 9.00 2.36 170.80 0.08 1.00 0.11 161.31 0.05 16.01 8.10 157.90 0.09
GS 9.00 3.29 154.67 0.11 10.00 2.46 170.61 0.08 1.00 0.09 161.64 0.05 15.01 7.71 155.69 0.08
GNN 13.00 3.07 145.98 0.10 4.00 2.20 171.00 0.08 0.00 0.00 163.41 0.05 18.01 10.33 155.58 0.08
GNT 5.00 1.27 161.73 0.12 1.00 0.30 192.77 0.11 0.00 0.00 180.42 0.07 15.01 7.37 164.25 0.09
GNS 6.00 0.99 156.43 0.12 1.00 0.89 185.09 0.10 0.00 0.00 173.43 0.06 14.01 8.43 162.00 0.09
GARCHNet: Value‑at‑Risk Forecasting with GARCH Models Based…

10 GN 7.00 2.80 150.07 0.11 11.00 2.02 168.25 0.08 1.00 0.06 160.39 0.05 13.02 13.43 170.30 0.10
GT 7.00 2.89 151.39 0.11 11.00 2.08 168.80 0.08 1.00 0.10 160.93 0.05 13.01 10.68 168.74 0.10
GS 6.00 2.22 157.48 0.12 10.00 2.32 169.18 0.08 1.00 0.09 161.92 0.05 13.01 11.11 168.57 0.10
GNN 9.00 1.83 148.53 0.11 9.00 3.53 165.76 0.08 2.00 0.17 162.25 0.05 18.01 9.65 156.53 0.08
GNT 5.00 0.89 158.37 0.12 4.00 0.85 176.82 0.09 0.00 0.00 179.22 0.07 17.01 6.29 161.69 0.09
GNS 6.00 0.60 158.83 0.12 5.00 0.68 179.52 0.10 0.00 0.00 173.25 0.06 17.01 7.83 160.13 0.09
20 GN 8.00 3.26 151.99 0.11 8.00 2.61 166.98 0.08 1.00 0.14 155.89 0.05 13.01 12.52 168.87 0.10
GT 7.00 3.13 153.64 0.11 8.00 2.88 167.53 0.08 1.00 0.17 156.87 0.05 15.01 9.31 166.77 0.10
GS 8.00 2.66 159.02 0.12 8.00 2.91 167.65 0.08 1.00 0.16 158.53 0.05 15.01 9.55 165.46 0.10
GNN 10.00 2.41 148.58 0.10 9.00 2.44 160.81 0.07 3.00 0.58 160.02 0.05 13.01 7.57 159.94 0.09
GNT 6.00 1.49 158.92 0.12 5.00 1.23 174.39 0.09 0.00 0.00 177.80 0.07 14.01 5.92 167.01 0.10
GNS 3.00 0.80 158.26 0.12 4.00 0.86 180.93 0.10 2.00 0.08 170.60 0.06 16.01 9.59 166.54 0.10

13
 Table 4  (continued)
p Period I (2009) Period II (2011) Period III (2017) Period IV (2020)
Model LLF CRLF CFLF ABLLF LLF CRLF CFLF ABLLF LLF CRLF CFLF ABLLF LLF CRLF CFLF ABLLF

13
100 GN 4.00 1.18 159.74 0.12 17.00 7.76 155.69 0.07 4.00 0.24 147.64 0.05 14.02 14.14 172.01 0.11
GT 3.00 1.16 160.12 0.13 17.00 7.74 156.00 0.07 2.00 0.34 148.40 0.05 14.02 14.14 172.50 0.11
GS 2.00 0.46 167.10 0.14 15.00 8.18 155.54 0.07 3.00 0.35 151.60 0.05 14.02 14.75 170.15 0.10
GNN 11.00 2.49 148.48 0.10 5.00 2.09 174.93 0.09 1.00 0.11 157.15 0.05 15.01 7.64 158.51 0.09
GNT 4.00 0.93 165.29 0.13 1.00 0.26 201.54 0.14 0.00 0.00 173.89 0.06 14.01 5.17 163.89 0.09
GNS 6.00 1.10 154.74 0.11 5.00 0.86 188.20 0.11 3.00 0.30 168.87 0.06 15.01 8.11 161.97 0.09

GN—GARCH with normally distributed innovations (d.i.); GT—GARCH with t d.i.; GST—GARCH with skewed t d.i.; GNN—GARCHNet with normally d.i.; GNT—
GARCHNet with t d.i.; GNS—GARCHNet with skewed t d.i.
Minimum cost values for each period and p pairs are in bold
M. Buczynski, M. Chlebus
 Table 5  Results of GARCH and GARCHNet models regarding the values of cost functions for S&P 500
p Period I (2009) Period II (2011) Period III (2017) Period IV (2020)
Model LLF CRLF CFLF ABLLF LLF CRLF CFLF ABLLF LLF CRLF CFLF ABLLF LLF CRLF CFLF ABLLF

1 GN 24.00 9.42 140.49 0.06 7.00 2.65 170.49 0.08 3.00 0.24 185.72 0.03 21.00 22.97 131.91 0.05
GT 23.00 8.79 141.04 0.06 7.00 2.39 170.40 0.08 3.00 0.24 186.62 0.03 24.00 66.08 193.32 0.06
GS 21.00 7.53 141.55 0.06 5.00 1.32 168.64 0.07 1.00 0.01 193.19 0.03 25.00 60.19 189.41 0.06
2 GN 22.00 7.70 141.03 0.06 8.00 3.26 167.41 0.07 2.00 0.03 185.52 0.03 20.00 19.94 124.71 0.05
GT 20.00 7.37 140.72 0.06 8.00 3.10 165.56 0.07 1.00 0.07 186.39 0.03 17.00 21.70 151.62 0.06
GS 19.00 8.27 144.61 0.06 7.00 2.56 168.71 0.07 1.00 0.13 193.45 0.03 17.00 36.73 166.88 0.06
5 GN 18.00 7.60 150.63 0.07 12.00 3.64 163.74 0.07 3.00 0.49 183.38 0.03 13.00 15.05 131.98 0.06
GT 18.00 7.46 151.61 0.07 10.00 3.27 164.50 0.07 3.00 0.60 184.56 0.03 13.00 23.86 172.72 0.07
GS 17.00 5.89 152.17 0.07 9.00 2.44 167.89 0.07 1.00 0.26 192.56 0.03 5.00 1.44 132.39 0.07
GNN 10.00 3.51 160.34 0.09 9.00 2.17 164.09 0.07 4.00 1.32 184.83 0.03 21.01 14.04 151.87 0.08
GNT 1.00 0.30 193.62 0.14 2.00 0.22 195.86 0.11 2.00 0.48 200.83 0.04 7.00 3.06 180.58 0.11
GNS 6.00 2.89 170.12 0.10 2.00 0.57 181.74 0.09 3.00 1.16 197.31 0.04 14.01 9.60 160.44 0.09
GARCHNet: Value‑at‑Risk Forecasting with GARCH Models Based…

10 GN 14.00 5.09 155.29 0.08 9.00 3.70 162.83 0.07 3.00 0.77 179.15 0.03 12.00 8.50 137.29 0.07
GT 14.00 5.10 156.32 0.08 11.00 3.56 163.27 0.07 3.00 0.75 181.30 0.03 3.00 0.94 147.45 0.08
GS 13.00 4.29 157.63 0.08 7.00 2.59 167.02 0.07 1.00 0.31 190.61 0.03 3.00 1.13 144.74 0.08
GNN 9.00 3.52 158.75 0.09 13.00 3.49 163.41 0.07 5.00 2.14 181.57 0.03 17.01 15.61 151.13 0.08
GNT 1.00 0.14 185.61 0.12 3.00 1.09 182.48 0.09 3.00 1.13 195.19 0.03 7.00 3.79 176.25 0.11
GNS 6.00 1.57 172.76 0.10 3.00 1.18 181.36 0.09 5.00 2.15 194.80 0.04 15.01 10.22 164.94 0.10
20 GN 11.00 2.87 162.14 0.09 11.00 5.24 162.34 0.07 6.00 3.49 170.27 0.02 6.00 1.30 152.11 0.08
GT 12.00 2.80 162.82 0.09 10.00 4.85 162.53 0.07 8.00 3.26 170.96 0.02 3.00 1.27 159.82 0.09
GS 10.00 2.24 166.33 0.10 7.00 3.02 165.73 0.07 6.00 1.79 179.00 0.03 3.00 0.63 156.57 0.09
GNN 11.00 2.85 157.96 0.09 13.00 4.01 163.35 0.07 5.00 1.76 183.39 0.03 15.01 13.23 152.31 0.08
GNT 3.00 1.10 184.23 0.12 2.00 1.02 185.34 0.10 1.00 0.32 201.21 0.04 9.00 4.58 168.50 0.10
GNS 4.00 1.22 172.16 0.10 8.00 2.48 182.00 0.09 4.00 1.10 196.60 0.04 14.01 13.76 174.42 0.10

13
 Table 5  (continued)
p Period I (2009) Period II (2011) Period III (2017) Period IV (2020)
Model LLF CRLF CFLF ABLLF LLF CRLF CFLF ABLLF LLF CRLF CFLF ABLLF LLF CRLF CFLF ABLLF

13
100 GN 7.00 3.20 170.06 0.11 14.00 10.82 159.78 0.06 9.00 4.91 164.51 0.02 10.00 4.70 145.32 0.08
GT 7.00 2.99 170.84 0.11 14.00 10.83 160.00 0.06 8.00 4.74 165.36 0.02 13.00 2.50 142.64 0.08
GS 7.00 2.85 174.40 0.11 13.00 9.37 163.95 0.07 8.00 2.88 172.42 0.03 19.00 4.26 140.20 0.08
GNN 13.00 4.46 150.00 0.08 12.00 2.78 166.65 0.07 5.00 1.69 181.49 0.03 23.01 20.16 149.77 0.08
GNT 0.00 0.00 195.64 0.15 2.00 0.17 196.28 0.12 1.00 0.19 203.37 0.04 8.00 4.74 177.72 0.12
GNS 6.00 1.98 170.62 0.10 4.00 1.12 189.40 0.10 3.00 0.73 197.44 0.04 17.01 13.49 171.97 0.09

GN—GARCH with normally distributed innovations (d.i.); GT—GARCH with t d.i.; GST—GARCH with skewed t d.i.; GNN—GARCHNet with normally d.i.; GNT—
GARCHNet with t d.i.; GNS—GARCHNet with skewed t d.i.
Minimum cost values for each period and p pairs are in bold
M. Buczynski, M. Chlebus
 Table 6  Results of GARCH and GARCHNet models regarding the values of cost functions for FTSE 100
p Period I (2009) Period II (2011) Period III (2017) Period IV (2020)
Model LLF CRLF CFLF ABLLF LLF CRLF CFLF ABLLF LLF CRLF CFLF ABLLF LLF CRLF CFLF ABLLF

1 GN 11.00 3.02 139.09 0.06 3.00 0.24 160.73 0.06 4.00 1.10 167.47 0.03 22.00 8.83 139.87 0.05
GT 11.00 3.38 138.56 0.06 3.00 0.22 161.97 0.07 2.00 1.09 169.64 0.03 14.00 5.66 143.37 0.06
GS 7.00 1.96 144.41 0.06 3.00 0.12 166.15 0.07 2.00 0.41 170.83 0.03 11.00 3.58 146.93 0.07
2 GN 12.00 2.27 136.50 0.06 4.00 0.56 157.79 0.06 3.00 1.02 166.50 0.03 21.00 9.51 139.66 0.06
GT 11.00 2.32 135.57 0.06 3.00 0.53 159.78 0.07 3.00 0.92 169.16 0.03 10.00 5.17 141.45 0.06
GS 6.00 0.54 143.02 0.06 3.00 0.41 164.11 0.07 3.00 0.74 171.06 0.03 8.00 4.20 148.95 0.07
5 GN 17.00 4.25 137.41 0.06 5.00 1.03 155.91 0.06 3.00 0.88 165.83 0.03 16.00 4.84 136.58 0.06
GT 17.00 4.13 137.42 0.06 5.00 0.91 157.17 0.06 3.00 1.06 168.82 0.03 12.00 3.24 142.86 0.07
GS 10.00 3.18 141.72 0.06 3.00 0.65 162.31 0.07 3.00 0.70 170.32 0.03 7.00 2.33 151.39 0.08
GNN 13.00 4.45 159.11 0.07 9.00 3.32 163.92 0.07 7.00 2.40 175.37 0.03 21.01 11.16 153.83 0.08
GNT 0.00 0.00 189.75 0.12 2.00 0.25 182.15 0.09 3.00 0.72 190.99 0.04 14.01 6.25 169.67 0.10
GNS 6.00 0.75 170.62 0.09 3.00 0.67 182.84 0.09 3.00 0.90 191.40 0.04 13.01 8.67 162.36 0.08
GARCHNet: Value‑at‑Risk Forecasting with GARCH Models Based…

10 GN 17.00 3.40 138.21 0.06 6.00 0.91 155.45 0.06 3.00 1.11 166.79 0.03 7.00 2.08 144.59 0.07
GT 16.00 3.44 138.39 0.06 4.00 0.86 156.53 0.06 4.00 0.82 167.43 0.03 4.00 1.16 152.55 0.08
GS 7.00 1.98 143.67 0.07 3.00 0.61 161.91 0.07 4.00 0.74 169.15 0.03 2.00 0.72 161.21 0.08
GNN 10.00 2.86 156.25 0.08 10.00 2.97 161.26 0.07 6.00 2.74 172.87 0.03 17.01 7.85 153.85 0.08
GNT 2.00 0.19 186.80 0.11 4.00 1.42 173.38 0.08 4.00 0.90 182.29 0.04 13.01 4.07 165.68 0.09
GNS 5.00 1.71 170.08 0.09 3.00 0.45 179.45 0.09 5.00 1.70 187.47 0.04 12.01 4.59 159.39 0.08
20 GN 8.00 1.52 148.35 0.07 5.00 0.94 154.08 0.06 3.00 1.06 162.87 0.03 7.00 1.35 151.41 0.07
GT 8.00 1.12 148.35 0.07 5.00 0.91 155.09 0.06 4.00 1.36 164.39 0.03 6.00 0.87 157.10 0.08
GS 4.00 0.96 155.56 0.08 3.00 0.58 161.43 0.07 4.00 1.32 166.62 0.03 4.00 0.48 163.38 0.09
GNN 12.00 3.19 154.93 0.08 10.00 2.26 157.55 0.06 8.00 2.86 171.49 0.03 19.01 8.22 154.22 0.08
GNT 7.00 2.16 167.65 0.09 4.00 1.33 166.77 0.07 2.00 1.09 182.02 0.04 12.00 4.02 164.26 0.09
GNS 9.00 1.87 167.87 0.09 4.00 0.31 178.34 0.08 4.00 1.32 189.69 0.04 15.01 6.28 159.83 0.08

13
 Table 6  (continued)
p Period I (2009) Period II (2011) Period III (2017) Period IV (2020)
Model LLF CRLF CFLF ABLLF LLF CRLF CFLF ABLLF LLF CRLF CFLF ABLLF LLF CRLF CFLF ABLLF

13
100 GN 4.00 0.52 161.52 0.09 11.00 2.96 147.84 0.06 15.00 5.23 157.82 0.03 7.00 1.30 143.68 0.07
GT 4.00 0.52 161.53 0.09 11.00 2.94 147.99 0.06 14.00 5.25 158.63 0.03 8.00 1.68 143.66 0.07
GS 2.00 0.22 168.56 0.10 8.00 1.95 154.29 0.06 12.00 4.79 161.21 0.03 4.00 0.69 150.57 0.08
GNN 15.00 5.41 151.06 0.07 7.00 1.56 161.76 0.07 8.00 4.19 166.02 0.03 24.01 11.96 156.11 0.08
GNT 5.00 1.29 175.83 0.10 3.00 1.35 180.93 0.08 2.00 0.48 190.10 0.04 13.01 6.26 166.85 0.10
GNS 6.00 1.15 167.94 0.09 2.00 0.15 183.66 0.09 3.00 1.13 191.08 0.04 12.01 7.89 159.91 0.09

GN—GARCH with normally distributed innovations (d.i.); GT—GARCH with t d.i.; GST—GARCH with skewed t d.i.; GNN—GARCHNet with normally d.i.; GNT—
GARCHNet with t d.i.; GNS—GARCHNet with skewed t d.i.
Minimum cost values for each period and p pairs are in bold
M. Buczynski, M. Chlebus
GARCHNet: Value‑at‑Risk Forecasting with GARCH Models Based…

models do not look so good. In most cases, the values of the company’s cost func-
tion are the worst—only in a few cases was the value of the cost function for the
GARCHNet model lower. This is rather undesirable behavior due to the use of a
non-linear approach. The GARCHNet model with a normal distribution appears to
have the lowest ABLLF cost function value among the GARCHNet models and can
usually compete with the same cost function calculated for its GARCH family coun-
terpart. In summary, based on the cost function results, we assume that GARCHNet
at this stage is a relatively conservative model. The results converge across the index
tested—with noticeable differences, but these are due to the distribution of the data
rather than the model specification.

5 Conclusion

In this paper, we have proposed a new approach to the specification of conditional

variance in GARCH models—the GARCHNet model, which incorporates a simple
neural network with a long short-term memory. The idea behind the GARCHNet
model is that the neural network can easily approximate non-linear relationships,
and these are by far the most common in financial market volatility. Furthermore,
the simplicity of the GARCH maximum likelihood estimation allows the original
log likelihood functions to be used as cost functions in the GARCHNet neural net-
work model. We proposed three different GARCHNet models, each with a different
assumed distribution of innovations: normal, t and skewed t. The neural network
that is used as the conditional variance specification is rather small. It contains an
LSTM layer as input, followed by three fully connected layers. When the assumed
distribution requires parameters other than mean and variance, these are optimized
by the same neural network.
The GARCHNet models were compared with the original GARCH models in an
empirical study. VAR estimates were created using a rolling window method—we
trained the model using 1000 observations and estimated a forecast, then moved one
time step forward and estimated another forecast. Such a procedure was repeated
250 times. Logarithmic returns of the WIG20 index (Warsaw Stock Exchange,
Poland), S&P 500 (New York Stock Exchange, the USA) and FTSE 100 (London
Stock Exchange, the UK) were used as data.
Our results show that GARCHNet is an outstanding model that can explain con-
ditional variance at least at the same level as traditional approaches. Value-at-risk
forecasts are rather conservative, but fewer exceptions are observed for this reason.
GARCHNet would be more often chosen by regulators than by company manage-
ment itself due to its relatively higher opportunity cost. We also note a rather large
advantage of this model—by obtaining much more data (see p values greater than
10) the model can generate predictions of the same or better quality than GARCH
models that consider smaller data samples.
We can see several options to enhance the model already:

13
 M. Buczynski, M. Chlebus

1. The best lenght of sequence p

p is one of the most influential parameters of the model, as it determines the
amount of information that one forecast contains, but we did not notice any trends
that would determine its impact on the quality of the model.
2. Stopping criterion
Given that the validation sample is absent in the case of the one-day ahead fore-
cast (time step to time step), the available options for objectively determining
the end of the model training phase are exhausted. We believe that, in the case of
VaR, a stopping criterion based on statistical tests would be accurate.
3. Neural network architecture and hyper-parameter tuning
The neural network proposed here is rather small. We think that increasing the
number of parameters (and thus the depth of the NN) would positively affect
the quality of the model. Furthermore, we have not included any tuning of the
hyperparameters—most of them have been assumed rather than tested.
4. Another approach to estimation of the distribution’s parameters
The distribution parameters are estimated here by a separate layer that depends on
the previous layers. The deteriorated performance of GARCHNet with t distribu-
tion and skewed t distribution can only be the result of the approach taken. Two
other options that can be considered are either separate neural networks for the
estimation of additional parameters optimized in a single procedure (parameters
dependent on the forecast sample); or the inclusion of these parameters as sepa-
rate weights for optimization (parameters dependent on the training sample).
5. Possible extension of this approach to time-series
Given that GARCH models are not only used in VaR modelling, we are primarily
interested in the performance of GARCHNet in traditional time series forecasting.

Authors’ Contributions Both authors contributed equally to the research.

Funding Not applicable.

Data Availability The data that support the findings of this study are available from the corresponding
author upon request.

Code Availability The codes that were used in this study are available from the corresponding author
upon request.

Declarations
Conflict of interest The authors declare that they have no conflict of interest.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License,
which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as
you give appropriate credit to the original author(s) and the source, provide a link to the Creative Com-
mons licence, and indicate if changes were made. The images or other third party material in this article
are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the
material. If material is not included in the article’s Creative Commons licence and your intended use is
not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission

13
GARCHNet: Value‑at‑Risk Forecasting with GARCH Models Based…

directly from the copyright holder. To view a copy of this licence, visit [Link]
ses/​by/4.​0/.

References
Abad, P., & Benito, S. (2013). A detailed comparison of value at risk estimates. Mathematics and Com-
puters in Simulation, 94, 258–276.
Abad, P., Benito, S., & López, C. (2014). A comprehensive review of value at risk methodologies. The
Spanish Review of Financial Economics, 12(1), 15–32.
Abad, P., Muela, S., & Lopez, C. (2015). The role of the loss function in value-at-risk comparisons. Jour-
nal of Risk Model Validation, 9, 1–19.
Aloui, C., & Mabrouk, S. (2010). Value-at-risk estimations of energy commodities via long-memory,
asymmetry and fat-tailed GARCH models. Energy Policy, 38(5), 2326–2339.
Angelidis, T., Benos, A., & Degiannakis, S. (2004). The use of GARCH models in VaR estimation. Sta-
tistical Methodology, 1(1), 105–128.
Arnerić, J., Šestanović, T., & Aljinović, Z. (2014). GARCH based artificial neural networks in forecasting
conditional variance of stock returns. Croatian Operational Research Review, 5, 329–343.
Barone-Adesi, G., Engle, R. F., & Mancini, L. (2008). A GARCH option pricing model with filtered his-
torical simulation. The Review of Financial Studies, 21(3), 1223–1258.
BenSaïda, A. (2015). The frequency of regime switching in financial market volatility. Journal of Empiri-
cal Finance, 32, 63–79.
Bollerslev, T. (1986). Generalized autoregressive conditional heteroskedasticity. Journal of Economet-
rics, 31(3), 307–327.
Bonato, M. (2012). Modeling fat tails in stock returns: A multivariate stable-GARCH approach. Compu-
tational Statistics, 27(3), 499–521.
Chen, S., & Billings, S. A. (1992). Neural networks for nonlinear dynamic system modelling and identifi-
cation. International Journal of Control, 56(2), 319–346.
Chlebus, M. (2017). Ews-garch: New regime switching approach to forecast value-at-risk. Central Euro-
pean Economic Journal, 3(50), 1–25.
Cho, K., van Merriënboer, B., Gulcehre, C., Bougares, F., Schwenk, H., & Bengio, Y. (2014). Learning
phrase representations using RNN encoder-decoder for statistical machine translation.
Chorowski, J., & Zurada, J. M. (2014). Learning understandable neural networks with nonnegative
weight constraints. IEEE Transactions on Neural Networks and Learning Systems, 26(1), 62–69.
Christoffersen, P. F. (1998). Evaluating interval forecasts. International Economic Review, 39, 841–862.
Cont, R. (2002). Empirical properties of asset returns: Stylized facts and statistical issues. Quantitative
Finance, 1, 223–236.
Costanzino, N., & Curran, M. (2018). A simple traffic light approach to backtesting expected shortfall.
Risks, 6(1), 2.
Degiannakis, S., Floros, C., & Livada, A. (2012). Evaluating value-at-risk models before and after the
financial crisis of 2008: International evidence. Managerial Finance, 38, 436–452.
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018). BERT: Pre-training of deep bidirectional
transformers for language understanding.
Duffie, D., & Pan, J. (1997). An overview of value at risk. Journal of Derivatives, 4(3), 7–49.
Engle, R. F., & Bollerslev, T. (1986). Modelling the persistence of conditional variances. Econometric
Reviews, 5(1), 1–50.
Engle, R. F., & Manganelli, S. (2004). Caviar: Conditional autoregressive value at risk by regression
quantiles. Journal of Business & Economic Statistics, 22(4), 367–381.
Ergün, A. T., & Jun, J. (2010). Time-varying higher-order conditional moments and forecasting intraday
VaR and expected shortfall. The Quarterly Review of Economics and Finance, 50(3), 264–272.
Escanciano, J. C., & Olmo, J. (2010). Backtesting parametric value-at-risk with estimation risk. Journal
of Business and Economic Statistics, 28(1), 36–51.
Fama, E. F. (1998). Market efficiency, long-term returns, and behavioral finance1the comments of brad
barber, david hirshleifer, s.p. kothari, owen lamont, mark mitchell, hersh shefrin, robert shiller, rex
sinquefield, richard thaler, theo vermaelen, robert vishny, ivo welch, and a referee have been helpful.
kenneth french and jay ritter get special thanks. 1. Journal of Financial Economics, 49(3), 283–306.

13
 M. Buczynski, M. Chlebus

Francq, C., & Zakoïan, J.-M. (2004). Maximum likelihood estimation of pure GARCH and ARMA-
GARCH processes. Bernoulli, 10(4), 605–637.
Glosten, L. R., Jagannathan, R., & Runkle, D. E. (1993a). On the relation between the expected value
and the volatility of the nominal excess return on stocks. The Journal of Finance, 48(5), 1779–1801.
Glosten, L. R., Jagannathan, R., & Runkle, D. E. (1993b). On the relation between the expected value
and the volatility of the nominal excess return on stocks. The Journal of Finance, 48(5), 1779–1801.
Gneiting, T. (2011). Making and evaluating point forecasts. Journal of the American Statistical Associa-
tion, 106(494), 746–762.
Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning. The MIT Press.
Hansen, B. E. (1994). Autoregressive conditional density estimation. International Economic Review,
35(3), 705–730.
Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9(8),
1735–1780.
Jeong, Y., & Lee, S. (2019). Recurrent neural network-adapted nonlinear ARMA-GARCH model with
application to s &p 500 index data. Journal of the Korean Data and Information Science Society,
30(5), 1187–1195.
Kim, H. Y., & Won, C. H. (2018). Forecasting the volatility of stock price index: A hybrid model inte-
grating LSTM with multiple GARCH-type models. Expert Systems with Applications, 103, 25–37.
Kristjanpoller, W., & Minutolo, M. C. (2015). Gold price volatility: A forecasting approach using the arti-
ficial neural network-GARCH model. Expert Systems with Applications, 42(20), 7245–7251.
Kristjanpoller, W., & Minutolo, M. C. (2016). Forecasting volatility of oil price using an artificial neural
network-GARCH model. Expert Systems with Applications, 65, 233–241.
Kupiec, P. (1995). Techniques for verifying the accuracy of risk measurement models. The Journal of
Derivatives, 3(2), 73–84.
Lim, B., Arik, S., Loeff, N., & Pfister, T. (2019). Temporal fusion transformers for interpretable multi-
horizon time series forecasting.
Liu, W., & So, M. (2020). A GARCH model with artificial neural networks. Information, 11, 489.
Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., & Riedmiller, M. (2013).
Playing atari with deep reinforcement learning.
Nelson, D., & Cao, C. (1992). Inequality constraints in the univariate GARCH model. Journal of Busi-
ness & Economic Statistics, 10, 229–35.
Nelson, D. B. (1991). Conditional heteroskedasticity in asset returns: A new approach. Econometrica,
59(2), 347–370.
Nguyen, N., Tran, M.-N., Gunawan, D., & Kohn, R. (2019). A long short-term memory stochastic volatil-
ity model.
Nikolaev, N., Tino, P., & Smirnov, E. (2011). Time-dependent series variance estimation via recurrent
neural networks. In T. Honkela, W. Duch, M. Girolami, & S. Kaski (Eds.), Artificial neural networks
and machine learning—ICANN 2011 (pp. 176–184). Springer.
Nozari, M., Raei, S., Jahangiry, P., & Bahramgiri, M. (2010). A comparison of heavy-tailed estimates and
filtered historical simulation: Evidence from emerging markets. International Review of Business
Research Papers, 6, 347–359.
Omari, C., Mundia, S., & Ngina, I. (2020). Forecasting value-at-risk of financial markets under the global
pandemic of covid-19 using conditional extreme value theory. Journal of Mathematical Finance,
10(4), 28.
Pascanu, R., Mikolov, T., & Bengio, Y. (2012). On the difficulty of training recurrent neural networks. In
30th International Conference on Machine Learning, ICML 2013.
Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N.,
Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S.,
Steiner, B., Fang, L., Bai, J., & Chintala, S. (2019). Pytorch: An imperative style, high-performance
deep learning library, pp. 8024–8035.
Philippe, J. (2006). Value at risk (3rd ed.). McGraw-Hill.
Rothfuss, J., Ferreira, F., Walther, S., & Ulrich, M. (2019). Conditional density estimation with neural
networks: Best practices and benchmarks.
Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning representations by back-propagating
errors. Nature, 323(6088), 533–536.
Segal, G., Shaliastovich, I., & Yaron, A. (2015). Good and bad uncertainty: Macroeconomic and financial
market implications. Journal of Financial Economics, 117(2), 369–397.

13
GARCHNet: Value‑at‑Risk Forecasting with GARCH Models Based…

Sezer, O. B., Gudelek, M. U., & Ozbayoglu, A. M. (2020). Financial time series forecasting with deep
learning : A systematic literature review: 2005–2019. Applied Soft Computing, 90, 106181.
So, M. K., & Philip, L. (2006). Empirical analysis of GARCH models in value at risk estimation. Journal
of International Financial Markets, Institutions and Money, 16(2), 180–197.
Stooq. (2021). Historical data: Wig20 (wig20). Data retrieved from Stooq. [Link]
wig20​ &i=d
Vorbrink, J. (2014). Financial markets with volatility uncertainty. Journal of Mathematical Economics,
53, 64–78.
Wang, Z.-R., Chen, X.-H., Jin, Y.-B., & Zhou, Y.-J. (2010). Estimating risk of foreign exchange portfolio:
Using VaR and CVaR based on GARCH–EVT-copula model. Physica A: Statistical Mechanics and
its Applications, 389(21), 4918–4928.
Zanin, L., & Marra, G. (2012). Rolling regression versus time-varying coefficient modelling: An empiri-
cal investigation of the Okun’s Law in some euro area countries. Bulletin of Economic Research,
64(1), 91–108.

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps
and institutional affiliations.

13