INNOVATIONS
MUHAMMAD SHERAZ
Communicated by the former editorial board
Recently, there has been a growing interest in the methods addressing volatility in computational finance and econometrics. Peiris et al. [8] have introduced doubly stochastic volatility models with GARCH innovations. Random coeffi- cient autoregressive sequences are special case of doubly stochastic time series.
In this paper, we consider some doubly stochastic stationary time series with GARCH and Threshold GARCH errors. Some general properties of process, like variance and kurtosis are derived.
AMS 2010 Subject Classification: 60Gxx, 60C10, 62M10, 91B84.
Key words:GARCH processes, TGARCH−(1,1), doubly stochastic time series, RCA(1), RCA-MA(1), Sign-RCA-MA(1).
1. INTRODUCTION
Recently, there has been a growing development in the use of nonlinear volatility models in financial literature. Black-Scholes option pricing model has some contradictory assumptions, such as constant volatility and normally dis- tributed log returns. In many financial time series, particular empirical studies reveal some facts such as stock returns and foreign exchange rates, exhibit leptokurtosis and stochastic volatility. Any distribution can be characterized by a number of features such as mean, variance, skewness and kurtosis. The measure of kurtosis is considered as whether the data are peaked or flat relative to the Normal distribution.
The concept of stochastic volatility for financial time series, the autore- gressive conditionally heteroskedastic (ARCH) model was first studied by Engle [4] and its generalization, the GARCH model, by Bollerslev [2]. The GARCH model assume symmetric effects on volatility that is, good and bad news have the same effect on volatility, which is shortfall of these models. The Thresh- old GARCH (TGARCH) model was introduced by Zakoian [12] and various other nonlinear GARCH extensions have been proposed to capture asymmetric effects [3, 5]. Random coefficient autoregressive (RCA) time series were intro- duced by Nicholls and Quinn [7] and some of their properties were studied by
MATH. REPORTS16(66),3(2014), 421–430
Appadooet al. [1]. In this paper, we derive the kurtosis for various classes of doubly stochastic models with GARCH and TGARCH innovations. In Section 2, some random coefficient autoregressive time series are discussed. In Section 3, some doubly stochastic models with GARCH and TGARCH innovations are introduced and we derive the formula for kurtosis in terms of model parameters.
2.RANDOM COEFFICIENT AUTOREGRESSIVE TIME SERIES
Random coefficient autoregressive sequences represent a special case of doubly stochastic time series. Some random coefficient autoregressive time series are given as follows:
(2.1) yt= (φ+bt)yt−1+εt , t∈Z (2.2) yt= (φ+bt)yt−1+εt+θεt−1, t∈Z (2.3) yt= (φ+bt+ Φst−1)yt−1+εt+θεt−1, t∈Z
From the equations (2.1)–(2.3) we have Random coefficient autoregres- sive (RCA-1), Random coefficient autoregressive-moving average (RCA-MA- 1), Sign-RCA-MA -1, respectively. The mean, variance and kurtosis for the above processes have been discussed in [8]. The two necessary and sufficient conditions for second order stationarity are given by:
1. bεt
t
∼
0 0
,
σb2 0 0 σε2
; 2. φ2+σb2<1.
We write forst as:
st=
+1 if yt>0 0 ifyt= 0
−1 ifyt<0
where {bt} and {εt} are errors in the model respectively, φ,Φ and st are pa- rameters also E(s2t) = 1. In order to calculate the kurtosis, E(s4t) = 1 [11]. A process of the following form was considered in [10]:
(2.4) bt+1 =abt+ (1 +bt)υt+1.
Recently, Peiris et al. [8] have studied a doubly stochastic model of the following form:
(2.5) yt= (φ+bt)yt−1+εt
(2.6) bt+1=abt+ (1 +bt)υt+1
Let yt be a stationary doubly stochastic time series satisfying (2.5) and (2.6). The moments for the model have been recently calculated in [8], whereεt
is an identically distributed independent sequence of variables with zero mean and varianceσε2t.
1. υεt
t
N
∼
0 0
,
συ2 0 0 σ2ε
2. 1−a2−συ2 <1
3. 1−a2−2σ2υ−φ2+φ2a2+φ2συ2 <1
In 2.5 and 2.6 the observed random variable is modeled in two steps.
Firstly, the distribution of the observed outcome is represented in a standard way and in the second step,bt is treated as being itself a random variable.
Definition 2.1 ([12]).Let zt be an independently identically distributed (iid) sequence of random variables such thatE(zt) = 0 andV ar(zt) = 1 then εt is called a Threshold GARCH−(p, q) process if it satisfies:
σt = ω+
p
X
i=1
αi,+ε+t−i−αi,−ε−t−i+
q
X
j=1
βjσt−j
εt = σtzt
where ω,αi,+,αi,− and β are real numbers and through the coefficients αi,+, αi,− the current volatility depends on both the modulus and the sign of the past returns. The positive and negative components of εt are given by
ε+t = max (εt,0), ε−t = min (εt,0) andεt=ε+t +ε−t .
3. DOUBLY STOCHASTIC MODELS WITH THRESHOLD GARCH ERRORS
In this section, we present our results for doubly stochastic volatility models with Threshold GARCH errors. We consider equations (2.1) to (2.3) and provide the moments of doubly stochastic models with TGARCH−(1,1).
Theorem 3.1. Consider a doubly stochastic model of the following form with TGARCH−(1,1):
yt = (φ+bt)yt−1+εt
bt+1 = abt+ (1 +bt)υt+1
εt = σtzt
σt = ω+α1,+ε+t−1−α1,−ε−t−1+β1σt−j
Then we have the following results:
1. E(yt2) = ω2σz2(1 +a1)(1−a2−συ2)
(1−a1)(1−a2)(1−a2−2συ2−φ2+φ2a2+φ2συ2) 2. E(yt4) = f1
f2
, where
f1 = 3σz4ω4(1 + 3a1+ 5a2+ 3a1a2+ 3a3+ 5a1a3+ 3a2a3+a1a2a3) (1−a1)(1−a2)(1−a3)
+ 6ω4σz4(1 +a1)2[φ2(1−a2) +συ2(1−φ2)]
(1−a1)2(1−a2)2(1−a2−2σ2υ−φ2+φ2a2+φ2συ2)
f2 = 1−φ4−3συ4[ (−1−a2−5σ2υ+a3+a5−16a3σ2υ+ 3aσυ2−9aσυ4 (−1 +a3+ 3aσυ2)(−1 +a2+σ2υ)(−1 + 6a2σ2υ+a4+ 3συ4)]
−6φ2[ σ2υ
(1−a2−σ2υ)]−4φ6a[ σ4υ
(1−a3−3aσυ2)(1−a2−συ2)] 3. The Kurtosis of the process follows the definition K(y)= E(y4t)
[E(yt2)]2. Proof. 1. First we calculate the value of E(yt2)
E(yt2) =E[(φ+bt)yt−1+εt]2=φ2E(y2t−1) +E(b2t)E(yt−12 ) +E(ε2t) E(yt2)−φ2E(y2t−1)−E(b2t)E(yt−12 ) =E(σt2)σ2z
Using the second order stationarity condition and by putting values of E(σ2t) andE(b2t) we get:
E(y2t) = E(σ2t)σz2 (1−φ2−E(b2t))
E(y2t) = ω2σz2(1 +a1)(1−a2−συ2)
(1−a1)(1−a2)(1−a2−2συ2−φ2+φ2a2+φ2σ2υ) 2. Now we findE(y4t)
E(yt4) =E[(φ+bt)yt−1+εt]4
E(y4t) =φ4E(yt−14 )+E(b4t)E(yt−14 )+E(ε4t)+6φ2E(b2t)E(y4t−1)+4φE(b3t)E(yt−14 ) + 6φ2E(ε2t)E(y2t−1) + 6E(ε2t)E(b2t)E(yt−12 ) + 12φE(εt)E(bt)E(y2t−1)
E(y4t) = 3σ4zE(σt4) + 6σ2zE(b2t)E(σ2t)E(yt−12 ) + 6σ2zφ2E(σt2)E(yt−12 ) 1−φ4−E(b4t)−6φ2E(b2t)−4φE(b3t)
E(y4t) = 3σ4zE(σt4) + 6σ2zE(σt2)[E(b2t) +φ2]E(yt−12 ) 1−φ4−E(b4t)−6φ2E(b2t)−4φE(b3t)
Using the second order stationarity condition and by putting values of E(σ2t), E(σ2t) and E(b2t), E(b3t) and E(b4t), we get the required value of E(yt4) as given in the statement of the theorem and
E(σ2t) = ω2(1 +a1) (1−a1)(1−a2)
E(σ4t) = ω4(1 + 3a1+ 5a2+ 3a1a2+ 3a3+ 5a1a3+ 3a2a3+a1a2a3) (1−a1)(1−a2)(1−a3)
a1= 1
√
2π(α1,++α1,−)+β1, a2 = 1
2(α21,++α21,−)+ 2
√
2πβ1(α1,++α1,−) +β12 a3=
r2
π(α31,++α31,−) +3
2β1(α21,++α21,−) + 3
√2πβ12(α1,++α1,−) +β31 3. We can getK(y)= E(yt4)
[E(yt2)]2. The Kurtosis of the process follows from its definition ofK(y).
Theorem 3.2. Consider the following doubly stochastic time series with given conditions:
yt = (φ+bt)yt−1+εt+θεt−1
bt+1 = abt+ (1 +bt)υt+1
a) υεt
t
∼N
0 0
,
σ2υ 0 0 σε2
b) 1−a2−συ2 <1
c) 1−a2−2συ2−φ2+φ2a2+φ2συ2 <1
where εt is an identically distributed independent sequence of variables with mean zero and varianceσε2. Then we have results forE(yt2), E(yt4)and Kurtosis as follows:
1. E(yt2) = ω2σ2ε(1 +θ2)(1−a2−συ2) (1−a2−2συ2−φ2+φ2a2+φ2σ2υ)
2. E(yt4) = 3σ4ε(1 + 6θ2+θ4) + 6σε2(1 +θ2)[(φ2+E(b2t)]E(y2t−1) 1−φ4−E(b4t)−6φ2E(b2t)−4φE(b3t)
3. The Kurtosis of the process follows the definition K(y)= E(y4t) [E(yt2)]2. Proof. 1. First we calculate the value of E(yt2):
E(yt2) =E[(φ+bt)yt−1+εt+θεt−1]2
E(yt2) = φ2E(y2t−1) +E(b2t)E(yt−12 ) +E(ε2t) +θ2E(ε2t−1) +2E(φ+bt)E(yt−1)E(εt) + 2θE(εt)E(εt−1) +2θE(φ+bt)E(yt−1)E(εt−1)
E(yt2)−φ2E(y2t−1)−E(b2t)E(y2t−1) =E(ε2t) +θ2E(ε2t−1)
Using the second order stationarity condition and by replacing the values of E(σ2t) and E(b2t) we get:
E(yt2) = σε2(1 +θ2)
(1−φ2−E(b2t)) = σε2(1 +θ2) (1−φ2−(1−aσ22υ−συ2)) E(yt2) = σε2(1 +θ2)(1−a2−συ2)
(1−a2−2σ2υ−φ2+φ2a2+φ2συ2) 2. Now we calculate E(yt4)
E(yt4) =E[(φ+bt)yt−1+εt+θεt−1]4
Because the odd moments ofytare zero, we can write the above expression by expanding the above equation and applying E we get:
E(y4t) =Eh
(φ+bt)4yt−14 + 6(φ+bt)2y2t−1(εt+θεt−1)2+ (εt+θεt−1)4i Using the stationarity condition and E(εt) =E(εt−1) = 0
E(yt4) = E(φ+bt)4E(yt−14 ) +E(ε4t) +θ4E(ε4t) + 6θ2E(ε4t)
+6E(φ+bt)2E(ε2t)E(yt−12 ) + 6θ2E(φ+bt)2E(ε2t)E(y2t−1) E(yt4) = 3σε4(1 + 6θ2+θ4) + 6σ2ε(1 +θ2)[(φ2+E(b2t)]E(yt−12 )
1−φ4−E(b4t)−6φ2E(b2t)−4φE(b3t) 3. The Kurtosis of the process follows from its definition
K(y) = E(yt4) [E(y2t)]2,
K(y)=
3σ4ε(1 + 6θ2+θ4) + 6σε2(1 +θ2)[(φ2+E(b2t)]E(y2t−1) 1−φ4−E(b4t)−6φ2E(b2t)−4φE(b3t)
σε2(1 +θ2)(1−a2−σ2υ) (1−a2−2συ2−φ2+φ2a2+φ2συ2)
2 .
Theorem 3.3. Consider a doubly stochastic model of the following form with TGARCH−(1,1)
yt = (φ+bt)yt−1+εt+θεt−1
bt+1 = abt+ (1 +bt)υt+1 εt = σtzt
σt = ω+α1,+ε+t−1−α1,−ε−t−1+β1σt−j
Then we have the following results:
1. E(yt2) = ω2σz2(1 +θ2)(1 +a1)(1−a2−συ2)
(1−a1)(1−a2)(1−a2−2συ2−φ2+φ2a2+φ2συ2) 2. E(yt4) = f1
f2
,where
f1=σz4ω4(1+ 3a1+ 5a2+ 3a1a2+ 3a3+ 5a1a3+ 3a2a3+a1a2a3)
(1−a1)(1−a2)(1−a3) (1+6θ2+θ4) + 6ω4σz4(1 +θ2)2(1 +a1)2[φ2(1−a2) +συ2(1−φ2)]
(1−a1)2(1−a2)2(1−a2−2σ2υ−φ2+φ2a2+φ2συ2) f2= 1−φ4−3σ4υ[ (−1−a2−5σ2υ+a3+a5−16a3σ2υ+ 3aσυ2−9aσυ4
(−1 +a3+ 3aσυ2)(−1 +a2+συ2)(−1 + 6a2συ2+a4+ 3συ4)]
−6φ2[ σ2υ
(1−a2−σ2υ)]− −4φ6a
συ4
(1−a3−3aσ2υ)(1−a2−σ2υ)
3. The Kurtosis of the process follows from its definition K(y)= E(yt4) [E(y2t)]2 Proof. 1. First we calculate the value of E(yt2)
E(yt2) =E[(φ+bt)yt−1+εt+θεt−1]2
E(y2t)=φ2E(yt−12 )+E(b2t)E(y2t−1)+E(ε2t)+θ2E(ε2t−1)+2E(φ+bt)E(yt−1)E(εt) +2θE(εt)E(εt−1) + 2θE(φ+bt)E(yt−1)E(εt−1)
Using the second order stationarity condition and by replacing the values of E(σ2t) and E(b2t) we get:
E(y2t) = σz2(1 +θ2)E(σ2t) (1−φ2−E(b2t))
E(y2t) = ω2σz2(1 +θ2)(1 +a1)(1−a2−συ2)
(1−a1)(1−a2)(1−a2−2συ2−φ2+φ2a2+φ2σ2υ)
2. Now we calculate E(yt4)
E(yt4) =E[(φ+bt)yt−1+εt+θεt−1]4 Taking expectation E we get:
E(yt4) = E(φ+bt)4E(yt−14 ) +E(ε4t) +θ4E(ε4t) + 6θ2E(ε4t)
+6E(φ+bt)2E(ε2t)E(yt−12 ) + 6θ2E(φ+bt)2E(ε2t)E(y2t−1) Using the second order stationarity condition and by replacing the values of E(ε4t) and E(ε2t)
E(yt4) = 3σz4E(σ4t)(1 + 6θ2+θ4) + 6σ2zE(σt2)(1 +θ2)[(φ2+E(b2t)]E(y2t) 1−φ4−E(b4t)−6φ2E(b2t)−4φE(b3t)
Now by replacing the values of E(σt4), E(σt2) andE(b2t), E(b3t), E(b4t) we obtain the required value ofE(y4t) as:
E(y4t) = f1
f2, wheref1 and f2are given in the statement of the theorem.
Theorem 3.4. Consider a doubly stochastic model of the following form with TGARCH(1,1) errors as:
yt = (φ+bt+ Φst)yt−1+εt+θεt−1
bt+1 = abt+ (1 +bt)υt+1 εt = σtzt
σt = ω+α1,+ε+t−1−α1,−ε−t−1+β1σt−j
Then we have the following results:
1. E(yt2) = ω2σz2(1 +θ2)(1 +a1)(1−a2−συ2)
(1−a1)(1−a2)(1−a2−2συ2−Φ2+Φ2a2+Φ2σ2υ−φ2+φ2a2+φ2σ2υ) 2. E(yt4) =f1
f2
,where
f1 = 3σ4zω4(1 +θ4+ 6θ2)(1+3a1+5a2+3a1a2+3a3+5a1a3+ 3a2a3+a1a2a3) (1−a1)(1−a2)(1−a3)
+ 6σz4ω4(1 +θ2)2(1 +a1)2
(φ2+ Φ2)[(1−a2−συ2) +συ2 (1−a1)2(1−a2)2(1−a2−2σ2υ−Φ2+Φ2a2+Φ2συ2−φ2+φ2a2+φ2σ2υ) f2 = 1−φ4−3σ4υ[ (−1−a2−5συ2+a3+a5−16a3συ2+ 3aσυ2−9aσ4υ
(−1 +a3+ 3aσ2υ)(−1 +a2+συ2)(−1 + 6a2συ2+a4+ 3σ4υ)]
−6(φ2+ Φ2)[ σ2υ
(1−a2−2σ2υ)]−Φ4−6φ2Φ2
3. The Kurtosis of the process follows from its definition K(y)= E(yt4) [E(y2t)]2 Proof. 1. First we calculate the value of E(yt2)
E(yt2) =E[(φ+bt+ Φst)yt−1+εt+θεt−1]2
E(y2t) = φ2E(y2t−1) + Φ2E(yt−12 ) +E(b2t)E(yt−12 ) +E(ε2t) +θ2E(ε2t−1) +2E(φ+bt+ Φst)E(yt−1)E(εt) + 2θE(εt)E(εt−1)
+ 2θE(φ+bt+ Φst)E(yt−1)E(εt−1)
Using the second order stationarity condition and by replacing the values of E(σ2t) and E(b2t) we get:
E(y2t) = σz2(1 +θ2)E(σ2t) (1−φ2−Φ2−E(b2t))
E(y2t) = ω2σz2(1 +θ2)(1 +a1)(1−a2−συ2)
(1−a1)(1−a2)(1−a2−2σ2υ−Φ2+Φ2a2+Φ2συ2−φ2+φ2a2+φ2σ2υ) 2. Now we calculate E(yt4)
E(yt4) =E[(φ+bt+ Φst)yt−1+εt+θεt−1]4
By applying expectation E, the simplification of the above step is the same as the one in the previous theorem and we can write:
E(y4t) = 3ε4t(1 + 6θ2+θ4) + 6ε2t(1 +θ2)[(φ2+ Φ2+E(b2t)]E(yt−12 ) 1−φ4−3E(b4t)−6(φ2+ Φ2)E(b2t)−Φ4−6φ2Φ2
By replacing the values of ε4t, ε2t, E(b2t), E(b4t) and then using the second order stationarity condition and the values ofE(σ4t), E(σ2t) andE(y2t) we obtain the final expression of E(yt4) in terms of f1/f2 ,where f1 and f2 are given in the statement of the theorem.
3. The Kurtosis of the process follows from its definitionK(y)= E(y4t)
[E(yt2)]2.
4.CONCLUSIONS
In this paper, Kurtosis of doubly stochastic models with TGARCH in- novations are derived. Statistical inferences to these doubly stochastic models with other asymmetric GARCH errors and state space modeling can be viewed as a special case for nonlinear time series.
REFERENCES
[1] S.S. Appadoo, M. Ghahramani and A. Thavaneswaran,Moment properties of sometime series models. Math. Sci. 30(2005),1, 50–63.
[2] T. Bollerslev,Generalized autoregressive conditional heteroskedasticity. J. Econometrics 31(1986), 307–327.
[3] F. Christian and Z.M. Jean,GARCH Models. John Wiley and Sons, 2010.
[4] R.F. Engle,Autoregressive conditional heteroscedasticity with estimates of the variance of U.K, inflation. Econometrica50(1982), 987–1008.
[5] L. George, Computational Finance: Numerical Methods for Pricing Financial Instru- ments. Butterworth-Heinemann, Elsevier, 2004.
[6] F. Jurgen, H.K. Wolfgang and H.M. Christian, Statistics of Financial Markets. An Introduction, Springer, 2011.
[7] D.F. Nicholls and B.G. Quinn,Random Coefficient Autoregressive Models, An Introduc- tion. Lector Notes in Statistics11, Springer-New York, 1982.
[8] S. Peiris, A. Thavaneswaran and S. Appadoo,Doubly stochastic models with GARCH innovations. Appl. Math. Lett. 24 (2011), 1768–1773.
[9] V. Preda, S. Dedu and M. Sheraz, New measure selection for Hunt-Devolder semi- Markov regime switching interest rate models. Phys. A407(2014), 350–359.
[10] A.N. Shiryaev,Probability (2nd Edition). Springer-Verlag, New York Inc., 2005.
[11] A. Thavaneswaran, S.S. Appadoo and C.R. Bector, Recent developments in volatility modeling and applications. J. Appl. Math. Decis. Sci.,2006(2006), 1–23.
[12] J.M. Zakoian, Threshold heteroskadasticity models. J. Econom. Dynam. Control 15 (1994), 931–955.
Received 5 July 2012 University of Bucharest,
Faculty of Mathematics and Computer Science, Str. Academiei 14,
0100144 Bucharest, Romania