International Review of Financial Analysis 91 (2024) 102941

Contents lists available at ScienceDirect

International Review of Financial Analysis

journal homepage: www.elsevier.com/locate/irfa

GARCH-M model with an asymmetric risk premium: Distinguishing between

‘good’ and ‘bad’ volatility periods
Juri Trifonov ∗, Bogdan Potanin
HSE University, Pokrovsky Blvd 11, 109028, Moscow, Russia

ARTICLE INFO ABSTRACT

JEL classification: We proposed a new method (GARCH-M-GJR-LEV) that captures the asymmetry in the variance and return
C18 equations. The development of the model is encouraged by the stylized fact that investors demand a higher
C22 risk premium during ‘‘bad’’ volatility periods rather than ‘‘good’’ ones. To study the properties of the obtained
C58
estimators, we conducted simulated data analysis, considering a data-generating process characterized by
Keywords: asymmetric responses of risk premium to volatility changes. As a result, we have found statistical evidence
GARCH in favor of a significant advantage of the proposed method compared to existing alternatives. Further, the
Leverage effect
proposed model was applied to study the S&P 500 market index. We have found evidence of an asymmetric
Risk premium
relationship between the risk premium and volatility changes during most periods under consideration. Due
Conditional volatility
to this, the GARCH-M-GJR-LEV model usually outperformed the alternative GARCH family models according
to the information criteria.

1. Introduction conditional variance into the mean equation, introducing the contribu-
tion of the risk premium to the dynamics of returns. This model has
Volatility is considered one of the essential factors in financial become widely used by researchers. Nevertheless, Bollerslev (2022) in-
markets. The indicator reflects assets’ risk level and is usually mea- dicates that inconsistent results concerning the sign of the risk premium
sured by the standard deviation of returns. Thus, volatility modeling is and its robustness have been found in several papers: Bollerslev et al.
indispensable in risk management policy, option pricing models, and (2006), Hong and Linton (2020) and Rossi and Timmermann (2015).
other fields of finance (Miralles-Marcelo et al., 2013). The GARCH
According to Bollerslev (2022), recent studies illustrate contradictory
family models, first introduced by Bollerslev (1986) and Engle (1982),
outcomes since the contribution sign of the risk premium differs among
are supposed to be the most common method of modeling conditional
the periods and assets. Moreover, he suggests that the asymmetric
volatility of financial assets.
The introduction of GARCH encouraged the development of a new relation between volatility and return may cause such inconsistency.
literature stratum dedicated to creating various GARCH-type models. The classical GARCH-M model does not allow to capture such asym-
These modifications of classical GARCH were stimulated by the need to metry. Thus, the estimator of the risk premium contribution may be
adapt the model to the financial theory and stylized facts in financial inconsistent.
markets. Although the concept of asymmetric relation is novel in the frame-
According to the main portfolio theories of Markowitz (1952) and work of the GARCH-M model, it has been extensively studied in the
Sharpe (1964), asset returns and volatility are the key factors that context of classical GARCH models. It refers to the asymmetric re-
describe pricing. Specifically, returns and volatility have a positive sponses of conditional variance to shocks in returns. There is a widely
correlation. Thus, investors demand higher returns when the asset is known stylized fact in financial markets that volatility reacts more
more volatile. This idea has become the foundation for introducing the sharply to negative shocks in returns than positive ones (Black, 1976;
central financial economics concept. The concept states that every risky Zhang, 2006). Because classical GARCH is symmetric, it does not
asset incorporates a risk premium component in its return. Because
allow to account for such effects (Nelson, 1991). This problem has
classical GARCH models do not allow to account for a risk premium
led to the development of leveraged GARCH models that account
existence in the returns equation, this led to the development of a new
for the asymmetry effect of random shocks in the variance equa-
GARCH-type model. The GARCH-in-Mean (GARCH-M) model has been
tion. Most popular models of this type, including EGARCH (Nelson,
introduced by Engle et al. (1987). Its key feature was incorporating the

∗ Corresponding author.
E-mail addresses: [email protected] (J. Trifonov), [email protected] (B. Potanin).

https://doi.org/10.1016/j.irfa.2023.102941
Received 9 November 2022; Received in revised form 10 February 2023; Accepted 14 September 2023
Available online 20 September 2023
1057-5219/© 2023 Elsevier Inc. All rights reserved.
J. Trifonov and B. Potanin International Review of Financial Analysis 91 (2024) 102941

1991) and GJR-GARCH (Glosten et al., 1993), have been focusing on 2009). The other possible alternative is EGARCH proposed by Nelson
the functional form of the asymmetry effect in the variance equa- (1991). According to the empirical study of Awartani and Corradi
tion. Modern studies mainly concentrate on developing more flexible (2005), EGARCH has demonstrated the highest accuracy on S&P 500
asymmetric specifications of variance equation (Hansen et al., 2012) index data compared to various asymmetric GARCH-family models,
or on the multivariate generalization of classical asymmetric GARCH including GJR-GARCH, TGARCH, and AGARCH. However, it is compli-
models (McAleer et al., 2009). However, the concept of capturing asym- cated to derive the theoretical properties of the EGARCH model since
metric responses of returns to conditional volatility, i.e., the asymmetry it is based on a complex random coefficient nonlinear moving average
in the mean equation, is poorly studied in the literature. Our paper process whose stationarity conditions are still unknown (McAleer &
fills this gap by introducing the leverage effect in the mean part of the Hafner, 2014). Moreover, according to Awartani and Corradi (2005),
GARCH-M model. GJR-GARCH performed as the second best (after EGARCH) when the
In this study, we propose a new GARCH-M-GJR-LEV model that author studied combinations of asymmetric GARCH models with a
accounts for the asymmetry effect not only in the variance equation classical GARCH-M. Therefore we prefer GJR-GARCH over other asym-
but also in the equation of returns. The model is constructed in the metric GARCH specifications as a starting point for our model because
framework of GJR-GARCH and is based on the idea that negative shocks it is relatively easy to derive theoretical properties of its modifications
in returns lead to the higher contribution of conditional volatilities (i.e., of GARCH-M-GJR-LEV), and it has demonstrated reasonably good
to asset returns. In other words, the model includes an additional performance on S&P 500 index that we use in the empirical part of our
parameter in the mean equation that reflects the asymmetric responses study.
of returns to volatility changes. This process is identical to the classical GARCH specification but al-
According to the simulated data analysis results, we have found lows for asymmetric variance responses to shocks in returns. Following
statistical evidence in favor of the significant advantage of the proposed the notations from the previous model, the GJR-GARCH(1,1) process is
method in comparison with other GARCH-type models. Moreover, us- specified as follows:
ing the log returns of the S&P 500 index, we have found evidence 𝑦𝑡 = 𝜇 + 𝜀𝑡 ,
in favor of an asymmetric risk premium existence over different time
𝜀𝑡 |𝑡−1 ∼  (0, 𝜎𝑡2 ),
intervals.
The paper proceeds as follows: In the next section, we briefly intro- 𝑉 𝑎𝑟(𝑦𝑡 |𝑡−1 ) = 𝜎𝑡2 = 𝜔 + 𝛼𝜀2𝑡−1 + 𝛾𝐼𝑡−1 𝜀2𝑡−1 + 𝛽𝜎𝑡−1
2
,
{
duce and discuss GARCH-M and GJR-GARCH specifications. Next, we 0, if 𝜀𝑡 ≥ 0 (3.1)
introduce our model, explain the intuition and derive the unconditional 𝐼𝑡 =
1, if 𝜀𝑡 < 0,
variance of returns. In Section 4, we provide the results of the simulated
data analysis. In Section 5, we apply our model to study the log returns 𝜉𝑡 ∼  (0, 1) 𝑖.𝑖.𝑑.,
of the S&P 500 index and discuss the results. Finally, we summarize our 𝜀𝑡 = 𝜎𝑡 𝜉𝑡 .
contribution and findings.
The crucial feature of this model is that volatility may react dif-
ferently to positive and negative shocks in returns. It is implemented
2. The GARCH-M model
by including an indicator variable 𝐼𝑡 that reflects the sign of the shock
in period 𝑡. Thus, it may be easily seen that if the shock is positive,
On the probability space (𝛺,  , P), let us define the 𝜎-algebra in 𝑡−1
the contribution to volatility will be made only by the 𝛼 parameter. In
period as 𝑡−1 , which represents the set of all available information up contrast, negative shocks reflect a sharper contribution to conditional
to period 𝑡 − 1, 𝑡 ∈ N. Let us denote 𝑦 as a 𝑇 × 1 vector of log returns volatility equal to 𝛼 + 𝛾.
and 𝜎 as a 𝑇 × 1 vector of conditional volatilities, where 𝑇 ∈ N. Then It is straightforward to show that the unconditional mean and
without loss of generality, we can specify the GARCH-M(1,1) (Engle variance of the process are specified as follows:
et al., 1987) process with the following system of equations:
E[𝑦𝑡 ] = 𝜇,
𝑦𝑡 = 𝜇 + 𝜆𝜎𝑡 + 𝜀𝑡 , ( ) (3.2)
1
𝑉 𝑎𝑟 𝑦𝑡 = 𝜎 2 = .
𝜀𝑡 |𝑡−1 ∼  (0, 𝜎𝑡2 ), 1 − 𝛼 − 𝛾∕2 − 𝛽
𝑉 𝑎𝑟(𝑦𝑡 |𝑡−1 ) = 𝜎𝑡2 = 𝜔 + 𝛼𝜀2𝑡−1 + 𝛽𝜎𝑡−1
2
, (2.1) By definition, the unconditional variance is always nonnegative. There-
𝜉𝑡 ∼  (0, 1) 𝑖.𝑖.𝑑., fore, we can easily impose the parameter constraints that ensure the
existence of a stationary solution:
𝜀𝑡 = 𝜎𝑡 𝜉𝑡 ,
𝜔 > 0,
where parameters 𝜔, 𝛼, and 𝛽 determine the dynamics of conditional (3.3)
variance and 𝑡 ∈ {1, … , 𝑇 }. The distinctive feature of the model is the 𝛼 + 𝛾∕2 + 𝛽 < 1.
presence of 𝜎𝑡 in the mean equation. Therefore, volatility is allowed The model has been developed to account for the stylized fact about
to influence returns directly, and, thus, 𝜆𝜎𝑡 is usually interpreted as the asymmetric relationship between volatility and shocks in asset re-
a risk premium. Based on the portfolio theories of Markowitz (1952) turns. According to Black (1976) and Zhang (2006), statistical evidence
and Sharpe (1964), 𝜆 is expected to be positive because the increase in has been found that volatility reacts more sharply to negative shocks
volatility leads to a higher return (Engle et al., 1987). in returns than positive ones. Therefore, the sign of the 𝛾 parameter,
Although the model allows for capturing a risk premium effect, the reflecting the leverage effect, is expected to be positive. The positive
specification does not allow to account for the asymmetric responses. estimate will reflect more considerable volatility growth when the ob-
In other words, returns react identically to changes in volatility without served shocks are negative. Several approaches exist in the literature to
considering the sign of the shocks that caused it. Therefore, volatility explain the causes of the asymmetry effect. According to Black (1976)
during both ‘‘bear’’ and ‘‘bull’’ markets has the same effect on returns. and Christie (1982), negative shocks lead to an increase in the financial
leverage of issuing companies. Consequently, higher leverage enlarges
3. The GJR-GARCH model the stocks’ level of risk and leads to an increase in their volatility.
Moreover, according to the prospect theory of Kahneman and Tversky
The GJR-GARCH (Glosten et al., 1993) model has been developed (1979), the asymmetry effect can arise in the framework of investors’
to capture the leverage effect in financial markets. We use the GJR- cognitive features. People tend to perceive losses more critically, and,
GARCH framework as the basis for our model since its theoretical therefore, negative shocks in returns may result in massive asset sales
properties are well studied even in the multivariate case (McAleer et al., by investors, thereby provoking an increase in volatility.

2
J. Trifonov and B. Potanin International Review of Financial Analysis 91 (2024) 102941

4. The GARCH-M-GJR-LEV model Also, when the parameters 𝜆2 and 𝛾 are equal to zero, then the
specification converges to the process that is very similar to the classical
The classical GARCH-M model allows capturing risk premium in as- GARCH-M:
2
sets’ returns. Albeit, the model does not allow for asymmetric responses 𝑦𝑡 = 𝜇 + 𝜆1 𝜎𝑡−1 + 𝜀𝑡 ,
of returns to risk premium fluctuations. Therefore, the process cannot 𝜀𝑡 |𝑡−1 ∼  (0, 𝜎𝑡2 ),
differentiate between signs of shocks that cause the volatility, i.e., it
𝑉 𝑎𝑟(𝑦𝑡 |𝑡−1 ) = 𝜎𝑡2 = 𝜔 + 𝛼𝜀2𝑡−1 + 𝛽𝜎𝑡−1
2
, (4.2)
equally perceives negative and positive shocks. Thus, the GARCH-M
assumes that investors demand equal risk premiums when the market 𝜀𝑡 = 𝜎𝑡 𝜉𝑡 ,
experiences ‘‘good’’ and ‘‘bad’’ volatility periods. In other words, the 𝜉𝑡 ∼  (0, 1) 𝑖.𝑖.𝑑.
model accounts only for the absolute value of shocks determining the Further, to impose the necessary stationarity conditions, we need
risk premium. Nevertheless, according to Bollerslev (2022), statistical to specify the unconditional variance of returns with Theorem 1. The
evidence has been found that GARCH-M may demonstrate insignificant complete proof of the Theorem is given in Appendix.
risk premiums in asset returns. Moreover, some studies have shown that
the model may provide negative risk premium estimates. Such results Theorem 1. Suppose there exists a stationary solution to the proposed
are inconsistent with the main portfolio theories of Markowitz (1952) system of equations. Then, the expression for the unconditional variance of
and Sharpe (1964) since they state an inverse risk-return relationship, returns in the GARCH-M-GJR-LEV model has the following form:
implying a negative correlation between volatility and returns. Follow- ( ) ( ) ( [ ] [ ]2 )
ing Bollerslev (2022), the problem may be explained by the necessity of 𝑉 𝑎𝑟 𝑦𝑡 = 𝜎 2 = 𝜆21 + 𝜆1 𝜆2 × E 𝜎𝑡4 − E 𝜎𝑡2
( [ ] ) (4.3)
differentiation between ‘‘good’’ and ‘‘bad’’ volatility measures. Rational 1 1 [ ]2 [ ]
+ 𝜆22 × E 𝜎𝑡4 − E 𝜎𝑡2 + E 𝜎𝑡2 ,
investors tend to demand a higher risk premium during the ‘‘bad’’ 2 2
volatility period (downside risk) than in the case of the ‘‘good’’ period [ 4] 4
where E 𝜎𝑡 is an expectation of 𝜎𝑡 and is defined as follows:
(upside potential). [ ]
This article proposes a new class of GARCH models that accounts for [ ] 𝜔2 + 𝜔E 𝜎𝑡2 × (2𝛼 + 2𝛽 + 𝛾)
E 𝜎𝑡4 = . (4.4)
an asymmetric relationship between returns and volatility. The distinc- 1 − 3𝛼 2 − 𝛽 2 − 32 𝛾 2 − 2𝛼𝛽 − 3𝛼𝛾 − 𝛽𝛾
tive feature of the proposed method is that it allows for differentiation [ ]
between ‘‘good’’ and ‘‘bad’’ volatility periods in the returns equation. Also, E 𝜎𝑡2 denotes the unconditional variance of 𝜀𝑡 :
We construct the model in the framework of the GJR-GARCH process. 𝜔
Thereby, it captures leverage effects both in the return and variance 𝜎𝜀2 = 𝑉 𝑎𝑟(𝜀𝑡 ) = . (4.5)
1 − 𝛼 − 𝛾∕2 − 𝛽
equations.
By the definition of variance, we need to ensure that the value
Following the notations from Sections 2 and 3, we specify the
of Eq. (4.3) is positive. Thus, the stationary solution exists if and only
proposed model as follows:
if:
2 2
𝑦𝑡 = 𝜇 + 𝜆1 𝜎𝑡−1 + 𝜆2 𝐼𝑡−1 𝜎𝑡−1 + 𝜀𝑡 , 𝜎 2 = 𝑉 𝑎𝑟(𝑦𝑡 ) > 0. (4.6)
𝜀𝑡 |𝑡−1 ∼  (0, 𝜎𝑡2 ),
Usually, GARCH models are estimated via the maximum likelihood
𝑉 𝑎𝑟(𝑦𝑡 |𝑡−1 ) = 𝜎𝑡2 = 𝜔 + 𝛼𝜀2𝑡−1 + 𝛾𝐼𝑡−1 𝜀2𝑡−1 + 𝛽𝜎𝑡−1
2
, estimator. The log-likelihood function of the GARCH-M-GJR-LEV model
{
0, if 𝜀𝑡 ≥ 0 (4.1) may be specified as follows, assuming a normal distribution of random
𝐼𝑡 = shocks:
1, if 𝜀𝑡 < 0,
1∑ 1 ∑ 𝜀𝑡
𝑇 𝑇 2
𝑇
𝜀 𝑡 = 𝜎𝑡 𝜉 𝑡 , ln 𝐿(𝜃, 𝜀) = − ln 2𝜋 − ln 𝜎𝑡2 − , (4.7)
2 2 𝑡=1 2 𝑡=1 𝜎 2
𝑡
𝜉𝑡 ∼  (0, 1) 𝑖.𝑖.𝑑.,
where 𝜃 = (𝜇, 𝜔, 𝛼, 𝛽, 𝜆1 , 𝛾, 𝜆2 )′ is a vector of estimated parameters, 𝜀 =
where parameters 𝜆1 and 𝜆2 define the influence of conditional variance (𝜀1 , … , 𝜀𝑇 )′ . The values of conditional volatilities 𝜎𝑡 and random shocks
on returns. The presence of 𝜆2 𝐼𝑡−1 𝜎𝑡−1 2 allows differentiating between 𝜀𝑡 may be calculated recursively, following the GARCH-M-GJR-LEV
‘‘good’’ and ‘‘bad’’ volatility periods while defining the risk premium. process:
Therefore, volatility caused by negative shocks in returns can lead to
2 2
a higher risk premium. As a consequence, we expect the estimate of 𝜀𝑡 |𝑡−1 = 𝑦𝑡 − 𝜇 − 𝜆1 𝜎𝑡−1 − 𝜆2 𝐼𝑡−1 𝜎𝑡−1 ,
(4.8)
𝜆2 to be higher than 𝜆1 . When 𝜆2 > 0, it reflects the idea that investors 𝜎𝑡2 |𝑡−1 = 𝜔 + 𝛼𝜀2𝑡−1 + 𝛾𝐼𝑡−1 𝜀2𝑡−1 + 𝛽𝜎𝑡−1
2
.
demand a higher risk premium when the volatility is caused by a ‘‘bear’’
Maximization of the log-likelihood function provides estimators for
market.
the unknown parameters 𝜇, 𝜔, 𝛼, 𝛽, 𝜆1 , 𝛾, and 𝜆2 .
The model combines the GARCH-M process with the GJR-GARCH
specification. Although, we use a slightly modified version of the
GARCH-M model, imposing two specific differences. First, we specify 5. Simulated data analysis
the risk premiums through the conditional variances as 𝜆𝑖 𝜎𝑡−1 2 , 𝑖 =

{1, 2}, unlike the classical GARCH-M process that defines it using the We conducted simulated data analysis to study the properties of
conditional volatility, i.e., 𝜆𝜎𝑡 . We specify it this way since the uncondi- the obtained estimators and compare them with alternatives. The main
tional variance of the proposed method may be explicitly derived only point under consideration was to determine whether it is crucial to
by including a conditional variance instead of conditional volatility. account for asymmetric responses in the mean equation or whether
Second, while the classical GARCH-M model uses the current volatility other models may provide accurate estimates of the parameters, volatil-
to specify a risk premium, we use the variance of the previous period, ities, and returns even under the GARCH-M-GJR-LEV data generating
2 . It is necessary since the last period of conditional variance is
i.e., 𝜎𝑡−1 process. The parameters of the simulated data analysis are described in
required to estimate the model. We may clearly see that 𝜀𝑡 defines the Table 1.
binary variable 𝐼𝑡 in the same period 𝑡. In other words, if we specify We considered two sets of parameter values: Set I and Set II. In
the model with 𝜎𝑡2 and 𝐼𝑡 𝜎𝑡2 (instead of 𝜎𝑡−1 2 2 ), it will be
and 𝐼𝑡−1 𝜎𝑡−1 Set I, we set the true values of parameters similar to those frequently
impossible to define 𝜀𝑡 in the same period and, therefore, estimate the observed in applications of GARCH models to stock returns. Parame-
model via the maximum likelihood method. ters responsible for leverage effects, both in the mean and variance

3
J. Trifonov and B. Potanin International Review of Financial Analysis 91 (2024) 102941

Table 1 Table 2
Parameters of simulations. Accuracy metrics of coefficient estimates (Set I).
Parameter Set I Set II Metric/Model GARCH-M GARCH-M-GJR GARCH-M-GJR-LEV
𝜇 0.01 0.05 𝑅𝑀𝑆𝐸(𝜇)̂ 7.091 6.056 4.616
𝜔 0.1 0.05 𝑅𝑀𝑆𝐸(𝜔)̂ 3.390 3.111 2.609
𝛼 0.1 0.05 𝑅𝑀𝑆𝐸(𝛼)
̂ 10.106 4.013 4.144
𝛽 0.7 0.8 𝑅𝑀𝑆𝐸(𝛽)̂ 6.147 5.885 5.403
𝜆1 0.2 −0.05 𝑅𝑀𝑆𝐸(𝜆̂ 1 ) 21.438 7.657 6.639
𝛾 0.15 0.2 𝑅𝑀𝑆𝐸(̂𝛾 ) – 22.917 6.513
𝜆2 0.5 0.2 𝑅𝑀𝑆𝐸(𝜆̂ 2 ) – – 8.934
Sample size 1000
Number of simulations 100
Table 3
Accuracy metrics of coefficient estimates (Set II).
Metric/Model GARCH-M GARCH-M-GJR GARCH-M-GJR-LEV
equations, were set to the values that significantly change the data- 𝑅𝑀𝑆𝐸(𝜇)̂ 4.196 3.283 3.383
generating process. In Set II, parameters are assigned to the values that 𝑅𝑀𝑆𝐸(𝜔)̂ 3.150 1.694 1.700
𝑅𝑀𝑆𝐸(𝛼)
̂ 12.824 3.092 3.168
replicate our results of the GARCH-M-GJR-LEV model application to 𝑅𝑀𝑆𝐸(𝛽)̂ 7.106 4.095 3.965
the S&P 500 index stock returns (see Section 6). Next, we decided to 𝑅𝑀𝑆𝐸(𝜆̂ 1 ) 14.072 26.764 25.596
use a sample size of 1000 observations. We chose such a sample size 𝑅𝑀𝑆𝐸(̂𝛾 ) – 15.310 23.970
to make the experiment similar to the real data application problems 𝑅𝑀𝑆𝐸(𝜆̂ 2 ) – – 6.997

that usually provide a long enough time series of returns. At the same
time, larger samples are generally unavailable due to structural breaks.
Consequently, the experiment is designed to replicate practical appli- where 𝜃 = {𝜇, 𝜔, 𝛼, 𝛽, 𝜆1 , 𝛾, 𝜆2 } denotes a set of estimated parameters,
cations as closely as possible. Besides that, such a sample size allows and 𝑚 is an index of each simulation.
for studying the large sample properties of estimators. Finally, the total The values of RMSE metrics for every estimate in each model are
number of simulations equals 100 since the share of simulations where presented in Tables 2 and 3 (for Set I and Set II correspondingly).
the proposed method has demonstrated a significant advantage over Based on the obtained results for Set I, we can see that the GARCH-
the analogs is significantly high (see Tables 4–5). At the same time, a M-GJR-LEV model provides significantly lower values of the RMSE
higher number of simulations would require extensive computational metric for all coefficient estimates except 𝛼, ̂ which is slightly more ac-
and time resources. curate in the case of the GARCH-M-GJR model. Consequently, we have
The simulated data analysis proceeds as follows. First, we gener- found statistical evidence in favor of the advantage of the proposed
ate a pseudo-random sample following the GARCH-M-GJR-LEV data- method in terms of the accuracy of estimates. Following the results in
generating process. Second, we use the generated data to estimate Table 2, if the true data generating process differs from the one assumed
three models: GARCH-M, GARCH-M-GJR, and GARCH-M-GJR-LEV.1 By by the model, the obtained estimators may become inaccurate without
obtaining estimates of each model, we compare their accuracy and accounting for the leverage effect. The results for Set II (Table 3)
properties by using information criteria and accuracy metrics. reveal that the proposed method allows estimating the 𝜆1 parameter
more accurately by capturing the leverage effect in the risk premium.
It is important to note that the likelihood function in the GARCH-
Furthermore, according to the RMSE value for 𝜆̂ 2 , the proposed model
M-GJR-LEV model may not be necessary concave. Therefore, to reduce
tends to estimate 𝜆2 accurately (even more precisely than 𝜆1 ). This
the risks that the optimization procedure may converge to a local
evidence justifies the development of the GARCH-M-GJR-LEV model
maximum, we use 100 iterations of the hybrid genetic algorithm of
because if the observed data demonstrate an asymmetric relationship
global optimization with BFGS local optimizer that takes place each
between risk premium and volatility, then existing methods are likely
iteration with a probability of 0.1.2 To get an initial point for the to provide inefficient and inaccurate estimates.
optimization routine for the GARCH-M model, we set 𝜆1 = 0 and take Although parameter estimates are vital, the accuracy of predicted
classical GARCH estimates for other parameters. Similarly, we take volatilities and returns plays a more important role in practical applica-
GARCH-M estimates as initial points for GARCH-M-GJR, setting 𝛾 = 0. tions. Therefore, we calculated RMSE values for predicted conditional
Finally, we take GARCH-M-GJR estimates as the initial point for the volatilities and returns.4 We evaluated RMSE metrics for in-sample pre-
GARCH-M-GJR-LEV model, setting 𝜆2 = 0. dictions for each simulation in every model. To aggregate the results,
We calculated RMSE 3 values based on the coefficient estimates we calculated the mean values of RMSE among all simulations using
to compare the estimation accuracy among all four models. We use the following formulas:
the following formula to calculate RMSE metrics for every coefficient √
100 √
1 ∑√ √ 1 ∑ (𝜎 − 𝜎̂ )2 ,
𝑇
estimate:
√ 𝑅𝑀𝑆𝐸(𝜎) ̂ =
√ 100 𝑚=1 𝑇 𝑡=1 𝑡 𝑡
√ 1 ∑ 100
( )2
𝑅𝑀𝑆𝐸(𝜃) = √
̂ 𝜃 − 𝜃̂𝑚 , (5.1) √
100 √
(5.2)
100 𝑚=1 𝑚 1 ∑√ 1 ∑(
𝑇
)2
𝑅𝑀𝑆𝐸(𝑦) ̂ = √ 𝑦 − 𝑦̂𝑡 .
100 𝑚=1 𝑇 𝑡=1 𝑡

1
The mean values of RMSE for conditional volatilities and returns
We use the modified version of the GARCH-M model, defining the risk
2 are presented in Tables 4–5 for Set I and Set II correspondingly. Besides
premiums as 𝜆𝜎𝑡−1 (see Section 4). Such modification is applied for the
straightforward comparison with the GARCH-M-GJR-LEV model since, in this
that, we provide shares of simulations where each specific model has
case, GARCH-M is nested inside the GARCH-M-GJR-LEV process. demonstrated the lowest values of the RMSE criterion for conditional
2
During the preliminary analysis, we found that after 100 iterations volatilities and returns (Victories %). Finally, the average values of
genetic algorithm rarely provides notable improvements to the log-likelihood Akaike (AIC) and Bayesian (BIC) information criteria are calculated for
value for the models under consideration. Although, sometimes, these 100 every model.
iterations provide significant improvements in terms of log-likelihood value
and estimation accuracy.
3 4
For convenience, we indicate RMSE values multiplied by 100. Please, see We also considered the MAE and MSE metrics. However, we moved them
Appendix B for alternative metrics. into the Appendix since they indicated similar results.

4
J. Trifonov and B. Potanin International Review of Financial Analysis 91 (2024) 102941

Table 4 Table 6
Accuracy metrics and information criteria (Set I). S&P 500 estimation results for the period 2004–2021.
Metric/Model GARCH-M GARCH-M-GJR GARCH-M-GJR-LEV Parameters GARCH-M GARCH-M-GJR GARCH-M-GJR-LEV
𝑅𝑀𝑆𝐸(𝜎) ̂ 12.345 10.295 6.142 𝜇 0.0499*** 0.0225 0.0347**
𝑉 𝑖𝑐𝑡𝑜𝑟𝑖𝑒𝑠𝜎 % 0% 4% 96% (0.0141) (0.0138) (0.0140)
𝑅𝑀𝑆𝐸(𝑦) ̂ 94.581 94.210 90.568 𝜔 0.0286*** 0.0287*** 0.0291***
𝑉 𝑖𝑐𝑡𝑜𝑟𝑖𝑒𝑠𝑦 % 0% 0% 100% (0.0024) (0.0020) (0.0012)
𝛼 0.1347*** 0.0129*** 0.0162***
AIC 2573.072 2559.644 2500.384
(0.0084) (0.0042) (0.0030)
BIC 2597.610 2589.091 2534.738
𝛽 0.8389*** 0.8589*** 0.8553***
(0.0092) (0.0080) (0.0005)
Table 5 𝜆1 0.0265* 0.0135 −0.0237***
Accuracy metrics and information criteria (Set II). (0.0137) (0.0128) (0.0091)
𝛾 – 0.1903*** 0.1871***
Metric/Model GARCH-M GARCH-M-GJR GARCH-M-GJR-LEV
(0.0120) (0.0091)
𝑅𝑀𝑆𝐸(𝜎) ̂ 13.290 6.938 5.804 𝜆2 – – 0.0760***
𝑉 𝑖𝑐𝑡𝑜𝑟𝑖𝑒𝑠𝜎 % 0% 19% 81% (0.0003)
𝑅𝑀𝑆𝐸(𝑦) ̂ 100.393 100.285 98.766
AIC 11 791.596 11 643.245 11 629.996
𝑉 𝑖𝑐𝑡𝑜𝑟𝑖𝑒𝑠𝑦 % 0% 4% 96%
Note: *** – 𝑝 < 0.01, ** – 𝑝 < 0.05, * – 𝑝 < 0.1; st.errors in parentheses.
AIC 2619.222 2598.067 2585.704
BIC 2643.761 2627.514 2620.058

To ensure the robustness of the results, we estimated models for

the entire sample and specific periods. It is essential to consider that
The results clearly show that the proposed method provides signif-
long financial time series may be characterized by structural breaks.
icantly lower values of the RMSE criterion. For Set I, The GARCH-M-
Therefore, we decided to estimate the model on a set of samples that
GJR-LEV model demonstrates about two times the lower value of RMSE
include observations of 3 years. The main goal of the analysis is to ex-
for conditional volatilities than alternative models. Moreover, the share
amine the S&P 500 index for the asymmetric responses of risk premium
of lower RMSE values is much higher for the proposed method: the
to conditional volatility and compare the results of the GARCH-M-
GARCH-M-GJR-LEV has provided more accurate conditional volatility
GJR-LEV model with alternatives by the Akaike information criterion.
forecasts in 92% cases and more accurate forecasts of returns in 98%
The estimation results for the whole sample and specific periods are
cases correspondingly. The results for Set II similarly reveal that the
presented in Tables 6–13. The covariance matrix was estimated via the
developed model estimate volatilities and returns more precisely than
gradient outer product (GOP).
its counterparts. Thus, the results of simulated data analysis suggest
a significant advantage of the proposed method and justify its devel-
6.1. Whole sample analysis
opment and application. In other words, if the data generating process
involves the asymmetric relationship between risk premium and volatil-
In this subsection, we provide estimation results for the entire
ity, the existing methods may demonstrate inefficient estimators and
sample. To interpret the results meaningfully, it is necessary to evaluate
significantly lower accuracy of forecasts for volatilities and returns.
all three models: GARCH-M, GARCH-M-GJR, and GARCH-M-GJR-LEV.
Finally, we have tested the robustness of the maximum likelihood
Let us recall that the GARCH-M model accounts for the risk premium
estimator of the GARCH-M-GJR-LEV model to distributional assump-
effect, and GARCH-M-GJR captures the risk premium effect along with
tion violations.5 It is well known that a classical GARCH process esti-
the leverage effect in the volatility equation. Finally, the proposed
mator (based on the normality assumption) may preserve consistency
GARCH-M-GJR-LEV model allows accounting for the asymmetry effect
even if random shocks’ distribution deviates from normality (Berkes &
both in volatility and risk premium responses to shocks in returns.
Horváth, 2004). Unfortunately, establishing a similar formal result for
Based on the results presented in Table 6, all parameters in the
the GARCH-M-GJR-LEV model is a technically complicated task beyond
GARCH-M-GJR-LEV model are statistically significant at any reasonable
this paper’s scope. So, instead, we test the finite sample robustness of
level. The parameters of most interest are 𝜆1 and 𝜆2 . The estimate of
the GARCH-M-GJR-LEV estimator by replicating the analysis for Set
𝜆1 has a negative sign which is not intuitive since returns appear to
1, alternately assuming that 𝜉𝑡 follows the Student’s 𝑡 and noncentral
react negatively to volatility increases. Although this coincides with the
Student’s 𝑡-distributions (standardized to zero mean and unit variance).
results of previous research by Bollerslev (2022), such evidence may
To ensure that tails are rather heavy, we use 5 degrees of freedom
be caused by structural breaks. Therefore, we analyze this effect more
for both distributions. Noncentrality parameter for the noncentral Stu-
carefully by estimating smaller samples in the next subsections. The
dent’s 𝑡-distribution also has been set to 5, which makes this distribution
estimate of 𝜆2 is positive and demonstrates that risk premium increases
positively skewed. The results are presented in the Appendix (Ta-
more sharply when volatility rises due to the negative shocks in returns.
bles C.18–C.21) and suggest that distributional assumption violation
The asymmetric relationship between the risk premium and pre-
provides just a slight increase in RMSE values. It indicates in favor of
vious shocks in returns is visualized in Fig. 2. Since the estimate
the finite sample robustness of the GARCH-M-GJR-LEV estimator.
of 𝜆1 is negative and the estimate of 𝜆2 is positive, risk premium
only rises when the observed shocks in returns are negative. This
6. Application to the S&P 500 index returns
result clearly illustrates the above-mentioned contradictory findings
and demonstrates that during ‘‘good’’ volatility periods, investors tend
After studying the properties of the proposed method, we applied
to demand a discount instead of a premium.
it to the actual data, represented by the log returns of the S&P 500
Comparing the results with the other two models, we see that esti-
market index. The sample consists of 4531 observations and covers the
mate of 𝜆 in the GARCH-M model is positive and statistically significant.
period from 01.01.2004 to 31.12.2021. Fig. 1 represents the dynamics
of the S&P 500 log returns. The data was retrieved from the Bloomberg Although, the absolute value of the effect is much smaller than in the
Terminal. GARCH-M-GJR model. Consequently, by estimating GARCH-M instead
of the GARCH-M-GJR-LEV process, a researcher may underestimate
the risk premium. Moreover, applying the GARCH-M-GJR model, we
5
In the case of distributional assumption violation, it becomes a identify a significant leverage effect in the volatility equation (𝛾), while
quasi-maximum likelihood estimator (QMLE). the risk premium effect becomes insignificant and small. Therefore,

5
J. Trifonov and B. Potanin International Review of Financial Analysis 91 (2024) 102941

Fig. 1. Dynamics of the S&P 500 market index log returns.

Fig. 2. Risk premium responses to shocks in returns.

accounting for the leverage effect in the variance equation, one can Table 7
S&P 500 estimation results for the period 2019–2021.
misidentify the absence of the risk premium in returns since the lever-
Parameters GARCH-M GARCH-M-GJR GARCH-M-GJR-LEV
age coefficient in the volatility equation takes on all the influence.
Consequently, to correctly identify the risk premium and leverage 𝜇 0.0994*** 0.0762** 0.0801**
(0.032) (0.0319) (0.0327)
effects, the proposed GARCH-M-GJR-LEV model should be used because
𝜔 0.0656*** 0.0718*** 0.0711***
of the presence of an asymmetric risk premium effect in the observed (0.0108) (0.0115) (0.0115)
data. Finally, the Akaike (AIC) criterion identifies the method as the 𝛼 0.3034*** 0.1623*** 0.1616***
best of all three options. (0.0321) (0.0231) (0.023)
𝛽 0.6777*** 0.6569*** 0.6574***
(0.0296) (0.0297) (0.0298)
6.2. Analysis of 3-year samples: 2019–2021, 2018–2020, 2017–2019 𝜆1 0.0325 0.0149 −0.0114
(0.0219) (0.019) (0.0271)
This subsection provides results for the last three sequential 3-year 𝛾 – 0.3176*** 0.3205***
(0.0682) (0.0682)
samples. The estimation results are presented in Tables 7–9 corre- 𝜆2 – – 0.0644
spondingly for each period. We combine these three periods into one (0.0405)
subsection since they demonstrate similar results. AIC 2058.326 2046.602 2046.535
All three periods are characterized by statistically insignificant and
Note: *** – 𝑝 < 0.01, ** – 𝑝 < 0.05, * – 𝑝 < 0.1; st.errors in parentheses.
small estimates of 𝜆1 and 𝜆2 parameters in the GARCH-M-GJR-LEV
model. Consequently, we have found no statistical evidence of risk pre-
mium effects in the S&P 500 returns on the analyzed periods. We should
note that this finding coincides with the results of the GARCH-M and

6
J. Trifonov and B. Potanin International Review of Financial Analysis 91 (2024) 102941

Table 8 Table 10
S&P 500 estimation results for the period 2018–2020. S&P 500 estimation results for the period 2016–2018.
Parameters GARCH-M GARCH-M-GJR GARCH-M-GJR-LEV Parameters GARCH-M GARCH-M-GJR GARCH-M-GJR-LEV
𝜇 0.1017*** 0.0780** 0.0829** 𝜇 0.0598** 0.0301 0.0470
(0.0314) (0.0328) (0.0328) (0.0297) (0.0304) (0.0301)
𝜔 0.0550*** 0.0533*** 0.0522*** 𝜔 0.0394*** 0.0370*** 0.0344***
(0.0092) (0.0094) (0.0093) (0.0057) (0.0054) (0.0051)
𝛼 0.2467*** 0.1390*** 0.1399*** 𝛼 0.2146*** 0.0507*** 0.0581***
(0.0289) (0.0208) (0.0210) (0.0213) (0.0143) (0.0171)
𝛽 0.7227*** 0.7228*** 0.7242*** 𝛽 0.7382*** 0.7634*** 0.7701***
(0.0274) (0.0261) (0.0260) (0.0297) (0.0284) (0.0288)
𝜆1 0.0185 0.0060 −0.0156 𝜆1 0.0424 0.0319 −0.0749
(0.0242) (0.0224) (0.0097) (0.0573) (0.0544) (0.0525)
𝛾 – 0.2116*** 0.2077*** 𝛾 – 0.2556*** 0.2527***
(0.0437) (0.0426) (0.0298) (0.0398)
𝜆2 – – 0.0489 𝜆2 – – 0.1914***
(0.0317) (0.0483)
AIC 2145.537 2132.454 2133.173 AIC 1575.468 1557.305 1553.541

Note: *** – 𝑝 < 0.01, ** – 𝑝 < 0.05, * – 𝑝 < 0.1; st.errors in parentheses. Note: *** – 𝑝 < 0.01, ** – 𝑝 < 0.05, * – 𝑝 < 0.1; st.errors in parentheses.

Table 9
S&P 500 estimation results for the period 2017–2019. rather than positive ones. This evidence supports the ‘‘bad’’ and ‘‘good’’
Parameters GARCH-M GARCH-M-GJR GARCH-M-GJR-LEV volatility differentiation hypothesis of Bollerslev (2022). That is, in-
𝜇 0.0911*** 0.0616** 0.0566* vestors tend to demand a higher risk premium during the ‘‘bear’’
(0.0300) (0.0291) (0.0301)
market, i.e., driven by ‘‘bad’’ volatility than in the case of ‘‘good’’
𝜔 0.0306*** 0.0287*** 0.0298***
(0.0051) (0.0045) (0.0049) volatility periods.
𝛼 0.2022*** 0.0106 0.0134 According to Table 10, the parameter 𝜆1 in the GARCH-M model is
(0.0258) (0.0204) (0.0218) statistically insignificant, while the GARCH-M-GJR-LEV model provides
𝛽 0.7561*** 0.7910*** 0.7808***
a significant 𝜆2 parameter. It is an important result since it demon-
(0.0303) (0.0277) (0.0291)
𝜆1 0.0390 0.0064 0.0243 strates that the GARCH-M model could not capture a risk premium,
(0.0579) (0.0594) (0.0753) while GARCH-M-GJR-LEV has identified it. Considering that the param-
𝛾 – 0.2856*** 0.2928*** eter 𝜆1 in the proposed method is also statistically insignificant, we can
(0.0357) (0.0387)
conclude that the risk premium in the S&P 500 returns on this period
𝜆2 – – −0.0266
(0.0789)
reacts only to ‘‘bad’’ volatility. In contrast, ‘‘bull’’ market volatility
fluctuations do not increase the risk premium. In other words, investors
AIC 1538.754 1500.615 1502.444
demand risk premiums only during the drops in financial markets. In
Note: *** – 𝑝 < 0.01, ** – 𝑝 < 0.05, * – 𝑝 < 0.1; st.errors in parentheses.
contrast, high market turbulence during periods of sharp growth does
not stimulate an investor to require a higher risk premium. This pattern
demonstrates the irrationality of investors: one appears to perceive
GARCH-M-GJR models because they do not capture a significant risk negative news more dramatically in comparison with positive ones.
premium effect either. The result is also reflected in the slight difference Such evidence is consistent with the prospect theory of Black (1976),
in the AIC criterion between GARCH-M-GJR and GARCH-M-GJR-LEV Kahneman and Tversky (1979), Nelson (1991) and Zhang (2006).
models. Nevertheless, it is crucial to account for the leverage effect Besides that, it is essential to note that the data is also characterized
in the variance equation since both GARCH-M-GJR and GARCH-M- by asymmetric volatility responses to shocks in returns since GARCH-
GJR-LEV models provide significant and positive estimates of 𝛾, which M-GJR and GARCH-M-GJR-LEV models demonstrate significant and
evidences that volatility appears to react more sharply on negative
positive estimates of 𝛾. We may also see it from the lower values of the
shocks in returns than on positive ones. This conclusion is supported
AIC criterion in asymmetric models compared to the GARCH-M model.
by the significant difference in the AIC criterion between the GARCH-M
Finally, the proposed GARCH-M-GJR-LEV model appears to be the
model and the two other methods. Overall, the analyzed three periods
best based according to the Akaike information criteria. In other words,
are characterized by the absence of risk premium. Still, they include the
the model allowed to capture both leverage effects and, consequently,
leverage effect in the volatility equation, and therefore, GARCH-M-GJR
ensure a higher estimation accuracy.6 Furthermore, since the GARCH-
and the proposed GARCH-M-GJR-LEV model provide similar results.
M model did not capture a significant risk premium effect, one may
misidentify the presence of risk premium in the returns without apply-
6.3. Analysis of 3-year samples: 2016–2018, 2015–2017, 2014–2016,
2013–2015 ing the GARCH-M-GJR-LEV model.
Further, Tables 11–13 represent similar results. The only difference
In this subsection, we discuss results obtained from the analysis in these periods is that the 𝜆1 parameter in the GARCH-M model is
of four sequential 3-year samples, covering the period from 2013 to statistically significant in all three tables. This means that the GARCH-
2018. The estimation results of all periods are presented in Tables 10– M model captured a risk premium effect in contrast to the 2016–2018
13. As with the previous case, this section merges periods with similar period. Besides that, it is crucial to analyze the estimation results
patterns. provided by the GARCH-M-GJR model. It is seen that among all three
Based on the presented results, all estimated periods demonstrate tables, after accounting for the asymmetry in the variance equation
statistically significant estimates of 𝜆2 parameter in the GARCH-M- (through the application of the GARCH-M-GJR model), the 𝜆1 parame-
GJR-LEV model. The result provides evidence that returns include ter becomes statistically insignificant. That is, a researcher may not be
the asymmetric relationship between the risk premium and shocks in able to capture the risk premium by estimating the GJR specification.
returns, and the estimate is positive among all the estimated peri-
ods. Consequently, this clearly represents that risk premium increases
6
more sharply when volatility rises due to negative shocks in returns Leverage effects in returns and volatility equations.

7
J. Trifonov and B. Potanin International Review of Financial Analysis 91 (2024) 102941

Table 11 Table 13
S&P 500 estimation results for the period 2015–2017. S&P 500 estimation results for the period 2013–2015.
Parameters GARCH-M GARCH-M-GJR GARCH-M-GJR-LEV Parameters GARCH-M GARCH-M-GJR GARCH-M-GJR-LEV
𝜇 0.0004 0.0011 0.0065 𝜇 0.0042 0.0041 0.0174
(0.0313) (0.0312) (0.0164) (0.0514) (0.0387) (0.0282)
𝜔 0.0464*** 0.0492*** 0.0497*** 𝜔 0.0750*** 0.0683*** 0.0631***
(0.0068) (0.0086) (0.0085) (0.0194) (0.0159) (0.0144)
𝛼 0.2401*** 0.0543*** 0.0589*** 𝛼 0.1973*** 0.0002 0.0003
(0.0218) (0.0134) (0.0125) (0.0374) (0.0379) (0.0069)
𝛽 0.6996*** 0.7107*** 0.7003*** 𝛽 0.6854*** 0.6954*** 0.7039***
(0.0329) (0.0403) (0.0398) (0.0524) (0.0493) (0.0461)
𝜆1 0.1664** 0.0892 0.0134 𝜆1 0.1671* 0.0794 0.0123
(0.0650) (0.0594) (0.0324) (0.0877) (0.0625) (0.0715)
𝛾 – 0.3065*** 0.3098*** 𝛾 – 0.3996*** 0.3915***
(0.0352) (0.0351) (0.0588) (0.0411)
𝜆2 – – 0.1792*** 𝜆2 – – 0.1477***
(0.0486) (0.0476)
AIC 1553.335 1531.684 1523.906 AIC 1700.135 1651.677 1650.606

Note: *** – 𝑝 < 0.01, ** – 𝑝 < 0.05, * – 𝑝 < 0.1; st.errors in parentheses. Note: *** – 𝑝 < 0.01, ** – 𝑝 < 0.05, * – 𝑝 < 0.1; st.errors in parentheses.

Table 12
S&P 500 estimation results for the period 2014–2016. time, negative shocks in returns influence the premium more sharply,
Parameters GARCH-M GARCH-M-GJR GARCH-M-GJR-LEV i.e., investors demand a higher risk premium during the ‘‘bear’’ market.
𝜇 0.0004 0.0004 0.0004 To summarize, statistical evidence has been found that the ob-
(0.0422) (0.0397) (0.0419) served data exhibit the asymmetry effect in the variance equation and
𝜔 0.0691*** 0.0703*** 0.0679***
asymmetric responses of the risk premium to volatility changes. The
(0.0123) (0.0137) (0.0139)
𝛼 0.2499*** 0.0233 0.0058 proposed GARCH-M-GJR-LEV method allows capturing both of these
(0.0262) (0.0152) (0.0174) effects simultaneously. Because of this integral feature, the model was
𝛽 0.6874*** 0.6996*** 0.7172*** able to provide a better estimation quality, which is justified by the
(0.0399) (0.0454) (0.0461) lower values of the AIC information criterion.
𝜆1 0.1435** 0.0578 −0.0049
(0.0656) (0.0577) (0.0635)
7. Results & conclusion
𝛾 – 0.3630*** 0.3711***
(0.0429) (0.0402)
𝜆2 – – 0.1826** In this study, we have proposed the GARCH-M-GJR-LEV model.
(0.0747) Its integral feature is to simultaneously account for asymmetric re-
AIC 1753.369 1720.883 1715.978 sponses in the returns and variance equations. According to Bollerslev
Note: *** – 𝑝 < 0.01, ** – 𝑝 < 0.05, * – 𝑝 < 0.1; st.errors in parentheses.
(2022), investors perceive ‘‘good’’ and ‘‘bad’’ volatility periods differ-
ently, demanding different risk premiums. While well-known asym-
metric GARCH models (e.g., EGARCH and GJR-GARCH) capture asym-
metric responses of variance to shocks in returns, they either do not
Consequently, this may lead to a substantial misinterpretation of the account for a risk premium or assume that it reacts equally to negative
results. and positive return fluctuations. In contrast, while modeling the risk
Further, as with the 2016–2017 results, the 𝜆1 parameter in the premium, the proposed method distinguishes between volatility caused
GARCH-M-GJR-LEV model is statistically insignificant (Tables 11–13). by the ‘‘bear’’ and ‘‘bull’’ markets.
This evidences again that investors demand a risk premium only during We derived the analytical expression of the unconditional variance
‘‘bad’’ volatility periods when variance rises due to the sharp falls in for the proposed method and studied the properties via the simulated
returns. data analysis. According to the results, the GARCH-M-GJR-LEV model
The leverage effect (𝛾) in the variance equation is significant among demonstrates a significant advantage compared with other methods
all periods. This result is consistent between the GARCH-M-GJR and in case an asymmetric relationship between the risk premium and
GARCH-M-GJR-LEV models and across periods. Consequently, the S&P volatility changes characterizes the data generating process. Therefore
500 index returns exhibit the asymmetry effect, which should be cap- applying the model is necessary to obtain accurate estimates of GARCH
tured appropriately to provide a better estimation quality. model parameters, conditional volatilities, and returns.
In Fig. 3, we present the dependence of the risk premium on previ- Further, the introduced model was applied to study the volatility
ous values of shocks in returns for each period. All periods demonstrate of the S&P 500 market index. As a result, we have found statistical
evidence favoring the leverage effect in the mean equation over various
an asymmetric relationship between the risk premium and volatility
periods. Consequently, investors demand a higher risk premium in case
changes, based on the sign of shocks that cause volatility rises. The
volatility is determined by negative shocks in returns rather than posi-
graph for the period 2016–2018 represents a similar pattern to the
tive ones. Moreover, the results indicate that only negative shocks form
whole sample dependence (Fig. 2), while three other periods exhibit
a risk premium during some time intervals. In other words, investors
a different structure. The difference is explained only by the sign of
demanded a risk premium only during ‘‘bad’’ volatility periods. This
the 𝜆1 estimate. Because it was negative and high for the period 2016– evidence coincides with different empirical results (Bollerslev et al.,
2018 (relative to other periods), thus the plot illustrates that investors 2006; Rossi & Timmermann, 2015), and the prospect theory of Kahne-
demand a risk premium only during ‘‘bad’’ volatility periods, while man and Tversky (1979). Furthermore, we have found that a researcher
‘‘good’’ volatility may even produce a discount. The period 2015–2017 may not be able to correctly estimate a risk premium without applying
also illustrates that the risk premium is negative (risk discount) when the GARCH-M-GJR-LEV model in case the data demonstrate a leverage
shocks in returns are positive, albeit the absolute discount value is effect in the equation of returns. This problem may lead to a significant
significantly lower compared to 2016–2018. It is due to the lower misinterpretation of results. Finally, the Akaike information criterion
value of 𝜆̂ 1 . The remaining two periods demonstrate that risk premium indicates that the proposed model is the best among alternatives during
rises during both ‘‘good’’ and ‘‘bad’’ volatility periods. At the same most periods.

8
J. Trifonov and B. Potanin International Review of Financial Analysis 91 (2024) 102941

Fig. 3. Risk premium responses to shocks in returns.

CRediT authorship contribution statement distribution, so:

[ ] [ ] 1 [ ] [ ] 1 [ ]
E 𝐼𝑡 𝜀2𝑡 = E 𝜎𝑡2 𝜉𝑡2 |𝜉𝑡 < 0 ×𝑃 (𝜉 < 0) = E 𝜎𝑡2 × E 𝜉𝑡2 = E 𝜎𝑡2 ,
Juri Trifonov: Development of the main idea, Derivation ⏟⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏟ 2 2
of the unconditional variance, Drafting the manuscript, Software independent
implementation of the proposed method. Bogdan Potanin: Drafting (A.7)
the manuscript, Software implementation of the proposed method. [ ] [ ] 1 [ ] [ ] 3 [ ]
E 𝐼𝑡 𝜀4𝑡 = E 𝜎𝑡4 𝜉𝑡4 |𝜉𝑡 < 0 ×𝑃 (𝜉 < 0) = E 𝜎𝑡4 × E 𝜉𝑡4 = E 𝜎𝑡4 ,
⏟⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏟ 2 2
Declaration of competing interest independent
(A.8)
[ ] [ ] [ ] ( )
The authors declare that they have no known competing financial E 𝐼𝑡 𝜎𝑡2 𝜀2𝑡 = E 𝐼𝑡 𝜎𝑡4 𝜉𝑡2 = E 𝜎𝑡4 𝜉𝑡2 |𝜉𝑡 < 0 × 𝑃 𝜉𝑡 < 0
interests or personal relationships that could have appeared to 1 [ ] [ ] 1 [ ] (A.9)
= E 𝜎𝑡4 × E 𝜉𝑡2 = E 𝜎𝑡4 .
influence the work reported in this paper. 2 2
Using the formulas above, it is straightforward to show that:
[ 4] [( )2 ]
Appendix A. Derivation of the unconditional variance E 𝜎𝑡 = E 𝜔 + 𝛼𝜀2𝑡−1 + 𝛽𝜎𝑡−1 2
+ 𝛾𝐼𝑡−1 𝜀2𝑡−1
[ ] [ 4 ]
= 𝜔2 + 𝛼 2 E 𝜀4𝑡−1 + 𝛽 2 E 𝜎𝑡−1
Here, we derive an analytical expression for the unconditional [ [ ] [ 2 ]] [ ]
( )
variance of returns, i.e., 𝑉 𝑎𝑟 𝑦𝑡 . + 𝛾 2 E 𝐼𝑡−1 2
𝜀4𝑡−1 + 2𝜔𝛼E 𝜀2𝑡−1 + 2𝜔𝛽E 𝜎𝑡−1 + 2𝜔𝛾E 𝐼𝑡−1 𝜀2𝑡−1
[ ] [ ] [ ]
+ 2𝛼𝛽E 𝜀2𝑡−1 𝜎𝑡−1 2
+ 2𝛼𝛾E 𝐼𝑡−1 𝜀4𝑡−1 + 2𝛽𝛾E 𝐼𝑡−1 𝜎𝑡−1 2
𝜀2𝑡−1
Proof of Theorem 1. First, let us derive some preliminary results. ( ) [ ]
3
= 𝜔2 + 3𝛼 2 + 𝛽 2 + 𝛾 2 + 2𝛼𝛽 + 3𝛼𝛾 + 𝛽𝛾 × E 𝜎𝑡4
Remind that from Glosten et al. (1993): [ ] 2
( ) [ ] [ ] 𝜔 + 𝜔E 𝜎𝑡2 × (2𝛼 + 2𝛽 + 𝛾) .
𝑉 𝑎𝑟 𝜀𝑡 = E 𝜀2𝑡 = E 𝜎𝑡2 = . (A.1)
1 − 𝛼 − 𝛽 − 12 𝛾 (A.10)
[ ]
Notice that 𝐼𝑡 = 𝐼𝑡𝑠 for any 𝑠 ∈ 𝑁 and 𝜎𝑡 is independent of 𝜉𝑡 , so: Solving for E 𝜎𝑡4 , we get:
[ ]
[ ] [ ] [ ] ( ) 1 [ ] [ ] 𝜔2 + 𝜔E 𝜎𝑡2 × (2𝛼 + 2𝛽 + 𝛾)
E 𝐼𝑡2 𝜎𝑡2
= E 𝐼𝑡 𝜎𝑡2 = E 𝜎𝑡2 |𝜉𝑡 < 0 ×𝑃 𝜉𝑡 < 0 = E 𝜎𝑡2 , (A.2) E 𝜎𝑡4 = . (A.11)
2
⏟⏞⏞⏞⏟⏞⏞⏞⏟ 1 − 3𝛼 2 − 𝛽 2 − 32 𝛾 2 − 2𝛼𝛽 − 3𝛼𝛾 − 𝛽𝛾
independent
[ ] [ ] [ ] ( ) 1 [ ] Now we are ready to derive the expression for the unconditional
E 𝐼𝑡2 𝜎𝑡4 = E 𝐼𝑡 𝜎𝑡4 = E 𝜎𝑡4 |𝜉𝑡 < 0 ×𝑃 𝜉𝑡 < 0 = E 𝜎𝑡4 . (A.3) variance of 𝑦𝑡 :
⏟⏞⏞⏞⏟⏞⏞⏞⏟ 2
( )
independent 𝜎 2 = 𝑉 𝑎𝑟(𝑦𝑡 ) = 𝑉 𝑎𝑟 𝜇 + 𝜆1 𝜎𝑡−1 2 2
+ 𝜆2 𝐼𝑡−1 𝜎𝑡−1 + 𝜀𝑡
( 2 2
) ( )
Using the first and the fourth moments of standard normal distribu- = 𝑉 𝑎𝑟 𝜆1 𝜎𝑡−1 + 𝜆2 𝐼𝑡−1 𝜎𝑡−1 + 𝑉 𝑎𝑟 𝜀𝑡
tion, we get: ( 2 2
)
+ 2 𝐶𝑜𝑣 𝜆1 𝜎𝑡−1 + 𝜆2 𝐼𝑡−1 𝜎𝑡−1 , 𝜀𝑡
[ ] [ ] [ ] [ ] ⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟
E 𝜀2𝑡 = E 𝜎𝑡2 𝜉𝑡2 = E 𝜎𝑡2 × E 𝜉𝑡2 = E[𝜎𝑡2 ], (A.4) zero because of independence
( 2 ) ( )
[ ] [ ] [ ] [ ] [ ] = 𝜆21 𝑉 𝑎𝑟 𝜎𝑡−1 + 𝜆22 𝑉 𝑎𝑟 𝐼𝑡−1 𝜎𝑡−1
2
E 𝜀4𝑡 = E 𝜎𝑡4 𝜉𝑡4 = E 𝜎𝑡4 × E 𝜉𝑡4 = 3E 𝜎𝑡4 . (A.5) ( 2 2
) [ ] (A.12)
+ 2𝜆1 𝜆2 𝐶𝑜𝑣 𝜎𝑡−1 , 𝐼𝑡−1 𝜎𝑡−1 + E 𝜎𝑡2
Simply expanding the expression of shocks, one gets: ( [ ] [ 2 ]2 )
[ ] [ ] [ ] = 𝜆21 × E 𝜎𝑡−14
− E 𝜎𝑡−1
E 𝜀2𝑡 𝜎𝑡2 = E 𝜎𝑡2 𝜉𝑡2 𝜎𝑡2 = E 𝜎𝑡4 . (A.6) ( [ ] 1 [ 2 ]2 )
1
+ 𝜆22 × E 𝜎𝑡−1 4
− E 𝜎𝑡−1
𝜉𝑡2 2 ( [ 2
The distribution of does not depend on the event 𝜉𝑡 < 0 if
4
] [ 2 ]2 ) [ ]
𝜉𝑡 is symmetric around zero, which is the case of a standard normal + 𝜆1 𝜆2 × E 𝜎𝑡−1 − E 𝜎𝑡−1 + E 𝜎𝑡2 .

9
J. Trifonov and B. Potanin International Review of Financial Analysis 91 (2024) 102941

[ 4 ] [ ]
Note that since the process is stationary, then E 𝜎𝑡−1 = E 𝜎𝑡4 and Table B.17
[ 2 ] [ 2] Additional accuracy metrics of conditional volatilities and return predictions (Set II).
E 𝜎𝑡−1 = E 𝜎𝑡 . Applying this and simplifying the expression, the final
Metric/Model GARCH-M GARCH-M-GJR GARCH-M-GJR-LEV
specification of the unconditional variance in the GARCH-M-GJR-LEV
𝑀𝐴𝐸(𝜎)
̂ 9.161 4.478 3.957
process takes the following form:
𝑀𝑆𝐸(𝜎)
̂ 1.766 0.481 0.337
( ) ( ) ( [ ] [ ]2 ) 𝑀𝐴𝐸(𝑦)
̂ 74.851 74.763 74.032
𝑉 𝑎𝑟 𝑦𝑡 = 𝜆21 + 𝜆1 𝜆2 × E 𝜎𝑡4 − E 𝜎𝑡2
𝑀𝑆𝐸(𝑦)
̂ 100.787 100.572 97.548
( [ ] ) (A.13)
1 1 [ ]2 [ ]
+ 𝜆22 × E 𝜎𝑡4 − E 𝜎𝑡2 + E 𝜎𝑡2 ,
2 2
[ ] [ ]
where E 𝜎𝑡2 and E 𝜎𝑡4 can be calculated by the formulas (A.1) and Appendix C. Robustness to normality assumption violation
(A.11) correspondingly. □
Here we provide the results for the QMLE estimators. Tables C.18–
Additionally, using the formulas mentioned above, it is possible to
C.19 present the results assuming the Student’s 𝑡-distribution, while Ta-
show that the unconditional variance of 𝜎𝑡2 is specified as follows: bles C.20–C.21 reflect the results considering the noncentral Student’s
( ( [ ] [ ]2 ) ( [ ] [ ]2 ) 𝑡-distribution.
𝑉 𝑎𝑟(𝜎𝑡2 ) = 𝛼 2 × 3E 𝜎𝑡4 − E 𝜎𝑡2 + 2𝛼𝛽 × E 𝜎𝑡4 − E 𝜎𝑡2 Table C.18
( ) ( [ ]2 )
3 1 [ ]2 Accuracy metrics of coefficient estimates (Student’s 𝑡-distribution).
+ 𝛾2 × E[𝜎𝑡4 ] − E 𝜎𝑡2 + 𝛼𝛾 × 3E[𝜎𝑡4 ] − E 𝜎𝑡2
2 4 Metric/Model GARCH-M GARCH-M-GJR GARCH-M-GJR-LEV
( [ ]2 ) ) ( )−1
𝑅𝑀𝑆𝐸(𝜇)
̂ 8.961 7.596 5.076
+ 𝛽𝛾 × E[𝜎𝑡4 ] − E 𝜎𝑡2 × 1 − 𝛽2 .
𝑅𝑀𝑆𝐸(𝜔)
̂ 5.962 3.881 3.275
(A.14) 𝑅𝑀𝑆𝐸(𝛼)
̂ 10.866 4.709 3.935
̂
𝑅𝑀𝑆𝐸(𝛽) 8.679 6.944 6.245
𝑅𝑀𝑆𝐸(𝜆̂ 1 ) 20.727 8.462 7.262
Appendix B. Alternative metrics for model comparison 𝑅𝑀𝑆𝐸(̂𝛾 ) – 22.989 7.666
𝑅𝑀𝑆𝐸(𝜆̂ 2 ) – – 11.212
Table C.19
In this Section, we provide the calculated values of MAE and Accuracy metrics and information criteria (Student’s 𝑡-distribution).
MSE accuracy metrics for parameter estimates, conditional volatilities, Metric/Model GARCH-M GARCH-M-GJR GARCH-M-GJR-LEV
and returns. Tables B.14–B.17, similarly to RMSE, demonstrate the 𝑅𝑀𝑆𝐸(𝜎) ̂ 14.342 11.423 7.229
advantage of the GARCH-M-GJR-LEV model over alternative methods. 𝑉 𝑖𝑐𝑡𝑜𝑟𝑖𝑒𝑠𝜎 % 2% 11% 87%
𝑅𝑀𝑆𝐸(𝑦) ̂ 92.934 92.082 87.910
Table B.14
𝑉 𝑖𝑐𝑡𝑜𝑟𝑖𝑒𝑠𝑦 % 0% 1% 99%
Additional accuracy metrics of coefficient estimates (Set I).
AIC 2484.250 2468.050 2413.541
Metric/Model GARCH-M GARCH-M-GJR GARCH-M-GJR-LEV BIC 2508.789 2497.497 2447.896
𝑀𝐴𝐸(𝜇)
̂ 5.047 4.247 2.608 Table C.20
𝑀𝐴𝐸(𝜔)
̂ 2.501 2.281 2.092 Accuracy metrics of coefficient estimates (Noncentral Student’s 𝑡-distribution).
𝑀𝐴𝐸(𝛼)
̂ 9.239 3.014 2.908
̂ Metric/Model GARCH-M GARCH-M-GJR GARCH-M-GJR-LEV
𝑀𝐴𝐸(𝛽) 4.481 4.120 4.143
𝑀𝐴𝐸(𝜆̂ 1 ) 19.959 5.473 5.340 𝑅𝑀𝑆𝐸(𝜇)
̂ 9.309 8.144 5.804
𝑀𝐴𝐸(̂𝛾 ) – 21.676 5.341 𝑅𝑀𝑆𝐸(𝜔)
̂ 7.633 7.206 6.827
𝑀𝐴𝐸(𝜆̂ 2 ) – – 5.879 𝑅𝑀𝑆𝐸(𝛼)
̂ 13.954 11.561 10.734
̂
𝑅𝑀𝑆𝐸(𝛽) 14.588 13.559 13.519
𝑀𝑆𝐸(𝜇)̂ 0.500 0.364 0.211
𝑅𝑀𝑆𝐸(𝜆̂ 1 ) 24.855 24.556 21.107
𝑀𝑆𝐸(𝜔)̂ 0.114 0.096 0.067
𝑅𝑀𝑆𝐸(̂𝛾 ) – 28.285 9.677
𝑀𝑆𝐸(𝛼)
̂ 1.020 0.159 0.170
̂ 𝑅𝑀𝑆𝐸(𝜆̂ 2 ) – – 15.128
𝑀𝑆𝐸(𝛽) 0.374 0.343 0.289
𝑀𝑆𝐸(𝜆̂ 1 ) 4.589 0.581 0.438 Table C.21
𝑀𝑆𝐸(̂𝛾 ) – 5.246 0.421 Accuracy metrics and information criteria (Noncentral Student’s 𝑡-distribution).
𝑀𝑆𝐸(𝜆̂ 2 ) – – 0.790 Metric/Model GARCH-M GARCH-M-GJR GARCH-M-GJR-LEV
Table B.15 𝑅𝑀𝑆𝐸(𝜎) ̂ 13.147 12.676 10.230
Additional accuracy metrics of coefficient estimates (Set II). 𝑉 𝑖𝑐𝑡𝑜𝑟𝑖𝑒𝑠𝜎 % 19% 16% 65%
Metric/Model GARCH-M GARCH-M-GJR GARCH-M-GJR-LEV 𝑅𝑀𝑆𝐸(𝑦) ̂ 83.456 83.335 80.786
𝑀𝐴𝐸(𝜇)
̂ 3.537 2.766 2.754 𝑉 𝑖𝑐𝑡𝑜𝑟𝑖𝑒𝑠𝑦 % 5% 0% 95%
𝑀𝐴𝐸(𝜔)
̂ 1.725 1.357 1.377 AIC 2318.284 2311.030 2270.023
𝑀𝐴𝐸(𝛼)
̂ 12.245 2.528 2.580 BIC 2342.823 2340.477 2304.377
̂
𝑀𝐴𝐸(𝛽) 4.772 3.174 3.299
𝑀𝐴𝐸(𝜆̂ 1 ) 13.063 26.252 25.080
𝑀𝐴𝐸(̂𝛾 ) – 14.681 23.254
𝑀𝐴𝐸(𝜆̂ 2 ) – – 5.472 References
𝑀𝑆𝐸(𝜇)
̂ 0.176 0.108 0.114
𝑀𝑆𝐸(𝜔)
̂ 0.099 0.029 0.029 Awartani, B. M., & Corradi, V. (2005). Predicting the volatility of the S&P-500
𝑀𝑆𝐸(𝛼)
̂ 1.645 0.096 0.100 stock index via GARCH models: the role of asymmetries. International Journal of
̂
𝑀𝑆𝐸(𝛽) 0.505 0.168 0.157 Forecasting, 21(1), 167–183. http://dx.doi.org/10.1016/j.ijforecast.2004.08.003.
𝑀𝑆𝐸(𝜆̂ 1 ) 1.980 7.163 6.552 Berkes, I., & Horváth, L. (2004). The efficiency of the estimators of the parameters in
𝑀𝑆𝐸(̂𝛾 ) – 2.344 5.745 GARCH processes. The Annals of Statistics, 32, 633–655. http://dx.doi.org/10.1214/
𝑀𝑆𝐸(𝜆̂ 2 ) – – 0.490 009053604000000120.
Table B.16 Black, F. (1976). Studies of stock price volatility changes. In Proceedings of the 1976
Additional accuracy metrics of conditional volatilities and return predictions (Set I). meeting of the business and economic statistics section (pp. 177–181). American
Statistical Association.
Metric/Model GARCH-M GARCH-M-GJR GARCH-M-GJR-LEV
Bollerslev, T. (1986). Generalized autoregressive conditional heteroskedasticity. Journal
𝑀𝐴𝐸(𝜎)
̂ 7.655 6.237 3.694 of Econometrics, 31, 307–327. http://dx.doi.org/10.1016/0304-4076(86)90063-1.
𝑀𝑆𝐸(𝜎)
̂ 1.524 1.060 0.377
Bollerslev, T. (2022). Realized semi(co)variation: Signs that all volatilities are not
𝑀𝐴𝐸(𝑦)
̂ 71.677 71.453 69.117
created equal. Journal of Financial Econometrics, 20, 219–252. http://dx.doi.org/
𝑀𝑆𝐸(𝑦)
̂ 89.456 88.755 82.025
10.1093/jjfinec/nbab025.

10
J. Trifonov and B. Potanin International Review of Financial Analysis 91 (2024) 102941

Bollerslev, T., Litvinova, J., & Tauchen, G. (2006). Leverage and volatility feedback Kahneman, D., & Tversky, A. (1979). Prospect theory: An analysis of decision under
effects in high-frequency data. Journal of Financial Econometrics, 4, 353–384. http: risk. Econometrica, 47, 263–292. http://dx.doi.org/10.2307/1914185.
//dx.doi.org/10.1093/jjfinec/nbj014. Markowitz, H. (1952). Portfolio selection. The Journal of Finance, 7, 77–91. http:
Christie, A. (1982). The stochastic behavior of common stock variances value, leverage //dx.doi.org/10.1111/j.1540-6261.1952.tb01525.x.
and interest rate effects. Journal of Financial Economics, 10, 407–432. http://dx.doi. McAleer, M., & Hafner, C. M. (2014). A one line derivation of EGARCH. Econometrics,
org/10.1016/0304-405X(82)90018-6. 2(2), 92–97. http://dx.doi.org/10.3390/econometrics2020092.
Engle, R. F. (1982). Autoregressive conditional heteroscedasticity with estimates of the McAleer, M., Hoti, S., & Chan, F. (2009). Structure and asymptotic theory for
variance of United Kingdom inflation. Econometrica, 50, 987–1007. http://dx.doi. multivariate asymmetric conditional volatility. Econometric Reviews, 28(5), 422–440.
org/10.2307/1912773. http://dx.doi.org/10.1080/07474930802467217.
Engle, R. F., Lilien, D. M., & Robins, R. P. (1987). Estimating time varying risk Miralles-Marcelo, J. L., Miralles-Quirós, J. L., & del Mar Miralles-Quirós, M. (2013).
premia in the term structure: The arch-m model. Econometrica, 55, 391–407. Multivariate GARCH models and risk minimizing portfolios: The importance of
http://dx.doi.org/10.2307/1913242. medium and small firms. The Spanish Review of Financial Economics, 11, 29–38.
Glosten, L. R., Jagannathan, R., & Runkle, D. E. (1993). On the relation between the http://dx.doi.org/10.1016/j.srfe.2013.03.001.
expected value and the volatility of the nominal excess return on stocks. The Journal Nelson, D. B. (1991). Conditional heteroskedasticity in asset returns: A new approach.
of Finance, 48, 1779–1801. http://dx.doi.org/10.1111/j.1540-6261.1993.tb05128.x. Econometrica, 59, 347–370. http://dx.doi.org/10.2307/2938260.
Hansen, P. R., Huang, Z., & Shek, H. H. (2012). Realized GARCH: a joint model for Rossi, A. G., & Timmermann, A. (2015). Modeling covariance risk in merton’s ICAPM.
returns and realized measures of volatility. Journal of Applied Econometrics, 27(6), The Review of Financial Studies, 28, 1428–1461. http://dx.doi.org/10.1093/rfs/
877–906. http://dx.doi.org/10.1002/jae.1234. hhv015.
Hong, S. Y., & Linton, O. (2020). Nonparametric estimation of infinite order regression Sharpe, W. F. (1964). Capital asset prices: A theory of market equilibrium under
and its application to the risk-return tradeoff. Journal of Econometrics, 219, 389–424. conditions of risk. The Journal of Finance, 19, 425–442. http://dx.doi.org/10.1111/
http://dx.doi.org/10.1016/j.jeconom.2020.03.009. j.1540-6261.1964.tb02865.x.
Zhang, X. F. (2006). Information uncertainty and stock returns. The Journal of Finance,
61, 105–137. http://dx.doi.org/10.1111/j.1540-6261.2006.00831.x.

11