In this paper we define the fractional Cox–Ingersoll–Ross process as Xt:=Yt21{t<inf{s>0:Ys=0}}0:Y_{s}=0\}\}}$]]>, where the process Y={Yt,t≥0} satisfies the SDE of the form dYt=12(kYt−aYt)dt+σ2dBtH, {BtH,t≥0} is a fractional Brownian motion with an arbitrary Hurst parameter H∈(0,1). We prove that Xt satisfies the stochastic differential equation of the form dXt=(k−aXt)dt+σXt∘dBtH, where the integral with respect to fractional Brownian motion is considered as the pathwise Stratonovich integral. We also show that for k>00$]]>, H>1/21/2$]]> the process is strictly positive and never hits zero, so that actually Xt=Yt2. Finally, we prove that in the case of H<1/2 the probability of not hitting zero on any fixed finite interval by the fractional Cox–Ingersoll–Ross process tends to 1 as k→∞.
The classical Cox–Ingersoll–Ross (CIR) process, which was proposed and studied by Cox, Ingersoll and Ross in [4–6], is the process r={rt,t≥0} that satisfies the following stochastic differential equation:
drt=(k−art)dt+σrtdWt,a,k,σ>0.0.\]]]>
Here a corresponds to the speed of adjustment, k/a is called “the mean”, σ is “the volatility”, W={Wt,t≥0} is a Wiener process and r0>00$]]>.
The CIR process is widely used for short-term interest rate modeling as well as for stochastic volatility modeling in the Heston model [10]. Therefore, in most cases it is also assumed that 2k≥σ2 as, according to [9], if this condition holds, the process is strictly positive and never hits zero.
It is well known that the CIR process is ergodic and has a stationary distribution. Moreover, the distribution of its future values rt+T provided that rt is known is a noncentral chi-square distribution, and the distribution of the limit value r∞ is a gamma distribution.
However, the real financial models are often characterized by the so-called “memory phenomenon” (see [1, 2, 7, 18] for more detail), while the standard Cox–Ingersoll–Ross process does not display it. Therefore, for a better simulation of interest rates or stochastic volatility it is reasonable to consider a fractional generalization of the Cox–Ingersoll–Ross process. It should be noted that there are several approaches to definition of the fractional Cox–Ingersoll–Ross process. In [11, 12] the fractional CIR process is introduced as a time changed CIR process with inverse stable subordinator, the so-called “rough path approach” is described in [13]. Another way of defining the considered process is presented in [8] as part of the discussion on rough Heston models.
The definition of the fractional CIR process with k=0, based on the pathwise integration with respect to fractional Brownian motion, was also presented in [14] for the case H>2/32/3$]]>. In [16] it was shown that for such definition the fractional CIR process is the square of the fractional Ornstein–Uhlenbeck process until the first zero hitting (for definition and properties of the fractional Ornstein–Uhlenbeck process see [3]). Based on that, the fractional CIR process with k=0 was defined as the square of the fractional Ornstein–Uhlenbeck process before its first zero hitting. It was also shown that such process satisfies the stochastic differential equation of the form
dXt=aXtdt+σXt∘dBtH,t≥0,
where X0>0,a∈R,σ>0,H∈(0,1)0,a\in \mathbb{R},\sigma >0,H\in (0,1)$]]> and integral with respect to the fractional Brownian motion is the pathwise Stratonovich integral. However, due to positive probability of hitting zero, this process is not suitable for interest rate modeling.
In this paper we introduce a natural generalization of the above model. First, we consider the process Y={Yt,t≥0} which satisfies the SDE of the form
dYt=12(kYt−aYt)dt+σ2dBtH,Y0>0.0.\]]]>
Then, we define the fractional Cox–Ingersoll–Ross process as the square of Yt until the first zero hitting moment and show that it satisfies the SDE of the form
dXt=(k−aXt)dt+σXt∘dBtH,t≥0,
where X0=Y02>00$]]> and the integral with respect to the fractional Brownian motion is defined as the pathwise Stratonovich integral. We also show that for any k>00$]]> and for any Hurst parameter H>1/21/2$]]> the process is strictly positive and never hits zero.
Next, for the case of H<1/2, we prove that the probability that the fractional CIR process does not hit zero on any fixed finite interval tends to 1 as k→∞. As an auxiliary result, we prove the analogue of the comparison theorem.
The paper is organized as follows. In Section 2 we define the fractional CIR process and show that it satisfies the SDE (7) with the pathwise Stratonovich integral. In Section 3 we prove that the fractional CIR process is strictly positive for k>00$]]> and H>1/21/2$]]>. In Section 4 we prove the analogue of the comparison theorem and show that in the case of H<1/2 the probability that the fractional CIR process does not hit zero on any fixed finite interval tends to 1 as k→∞. In Appendix there are simulations that illustrate the results of the paper.
Definition of the fractional Cox–Ingersoll–Ross process
Consider the process Y={Yt,t≥0} that satisfies the following SDE until its first zero hitting:
dYt=12(kYt−aYt)dt+σ2dBtH,Y0>0,0,\]]]>
where a,k∈R,σ>00$]]> and {BtH,t≥0} is a fractional Brownian motion with the Hurst parameter H∈(0,1).
Let H∈(0,1) be an arbitrary Hurst index, {Yt,t≥0} be the process that satisfies the equation (5) and τ be the first moment of reaching zero by the latter. The fractional Cox–Ingersoll–Ross process is the process {Xt,t≥0} such that for all t≥0,ω∈Ω:
Xt(ω)=Yt2(ω)1{t<τ(ω)}.
Before moving to the main result of this section, let us give the definition of the pathwise Stratonovich integral.
Let {Xt,t≥0}, {Yt,t≥0} be random processes. The pathwise Stratonovich integral∫0TXs∘dYs is a pathwise limit of the following sums
∑k=1nXtk+Xtk−12(Ytk−Ytk−1),
as the mesh of the partition 0=t0<t1<t2<⋯<tn−1<tn=T tends to zero, in case if this limit exists.
Letτ:=inf{s>0:Ys=0}0:Y_{s}=0\}$]]>. For0≤t≤τthe fractional CIR process from Definition1satisfies the following SDE:dXt=(k−aXt)dt+σXt∘dBtH,whereX0=Y02>00$]]>and the integral with respect to the fractional Brownian motion is defined as the pathwise Stratonovich integral.
Let us fix an ω∈Ω and consider an arbitrary t<τ(ω).
According to (5) and (6),
Xt=Yt2=(X0+12∫0t(kYs−aYs)ds+σ2BtH)2.
Consider an arbitrary partition of the interval [0,t]:
0=t0<t1<t2<⋯<tn−1<tn=t.
Using (8), we get
Xt=∑i=1n(Xti−Xti−1)+X0=∑i=1n([X0+12∫0ti(kYs−aYs)ds+σ2BtiH]2−[X0+12∫0ti−1(kYs−aYs)ds+σ2Bti−1H]2)+X0.
Factoring each summand as the difference of squares, we get:
Xt=X0+∑i=1n[2X0+12(∫0ti(kYs−aYs)ds+∫0ti−1(kYs−aYs)ds)+σ2(BtiH+Bti−1H)]×[12∫ti−1ti(kYs−aYs)ds+σ2(BtiH−Bti−1H)].
Expanding the brackets in the last expression, we obtain:
Xt=X0+∑i=1nX0∫ti−1ti(kYs−aYs)ds+14∑i=1n(∫0ti(kYs−aYs)ds+∫0ti−1(kYs−aYs)ds)×∫ti−1ti(kYs−aYs)ds+σ4∑i=1n(BtiH+Bti−1H)∫ti−1ti(kYs−aYs)ds+σX0∑i=1n(BtiH−Bti−1H)+σ24∑i=1n(BtiH−Bti−1H)(BtiH+Bti−1H)+σ4∑i=1n(∫0ti(kYs−aYs)ds+∫0ti−1(kYs−aYs)ds)(BtiH−Bti−1H).
Let the mesh Δt of the partition tend to zero. The first three summands
∑i=1nX0∫ti−1ti(kYs−aYs)ds+14∑i=1n(∫0ti(kYs−aYs)ds+∫0ti−1(kYs−aYs)ds)×∫ti−1ti(kYs−aYs)ds+σ4∑i=1n(BtiH+Bti−1H)∫ti−1ti(kYs−aYs)ds→∫0t(kYs−aYs)(X0+12∫0s(kYu−aYu)du+σ2BsH)ds=∫0t(k−aYs2)ds=∫0t(k−aXs)ds,Δt→0,
and the last three summands
σX0∑i=1n(BtiH−Bti−1H)+σ24∑i=1n(BtiH−Bti−1H)(BtiH+Bti−1H)+σ4∑i=1n(∫0ti(kYs−aYs)ds+∫0ti−1(kYs−aYs)ds)(BtiH−Bti−1H)→σ∫0t(X0+12∫0s(kYu−aYu)du+σ2BsH)∘dBsH=σ∫0tYs∘dBsH=σ∫0tXs∘dBsH,Δt→0.
Note that the left-hand side of (9) does not depend on the partition and the limit in (10) exists as the pathwise Riemann integral, therefore the corresponding pathwise Stratonovich integral exists and the passage to the limit in (11) is correct.
Thus, the fractional Cox–Ingersoll–Ross process, introduced in Definition 1, satisfies the SDE of the form
Xt=X0+∫0t(k−aXs)ds+σ∫0tXs∘dBsH,
where ∫0tXs∘dBsH is the pathwise Stratonovich integral. □
In the case of k=0, the process (5) is the fractional Ornstein–Uhlenbeck process and the definition coincides with the one given in [16].
Hitting zero by the fractional CIR process with positive “mean” and H>1/2\mathbf{1}/\mathbf{2}$]]>
The next natural question regarding the fractional CIR process is finiteness of its zero hitting time moment. It is obvious that it coincides with the respective moment of the process {Yt,t≥0}, defined by the equation (5).
Before formulating the main result of the section let us give a well-known property of trajectories of fractional Brownian motion (see, for example, [15]).
Let{BtH,t≥0}be a fractional Brownian motion with the Hurst index H. Then,∃Ω′⊂Ω,P{Ω′}=1, such that∀ω∈Ω′,∀T>00$]]>,∀δ>00$]]>,∀0≤s≤t≤T∃C=C(T,ω,δ)∈R:|BtH−BsH|≤C|t−s|H−δ.
Letk>0,H>1/20,H>1/2$]]>. Then the process{Yt,t≥0}, defined by the equation (5), is strictly positive a.s.
The proof is by contradiction.
Let Ω′ be the same as in Proposition 1. First, assume that a>00$]]> and let for some ω∈Ω′, τ(ω)=inf{t>0:Xt=0}=inf{t>0:Yt=0}<∞0:X_{t}=0\}=\inf \{t>0:Y_{t}=0\}<\infty $]]>.
For all ε∈(0,min(Y0,ka)) (the condition ε<ka provides the inequality kε−aε>00$]]>) let us introduce the last moment of hitting the level of ε before the first zero reaching:
τε:=sup{t∈(0,τ):Yt=ε}.
Consider δ>00$]]> such that the inequality H−δ>1/2⇔1+δ−H<1/21/2\Leftrightarrow 1+\delta -H<1/2$]]> holds. According to the definitions of τ,τε and Y, the following equality is true:
−ε=Yτ−Yτε=12∫τετ(kYs−aYs)ds+σ2(BτH−BτεH).
The process Ys∈(0,ε) on the interval (τε,τ), hence ∀s∈(τε,τ):
kYs−aYs≥kε−aε.
From this and Proposition 1, it follows that ∃C=C(τ(ω),ω,δ):
σ2C|τ−τε|H−δ≥σ2|BτH−BτεH|≥12(kε−aε)(τ−τε)+ε.
It is clear that there exists ε˜>00$]]> such that ∀ε<ε˜: kε−aε>k2ε\frac{k}{2\varepsilon }$]]>. Then, by choosing an arbitrary ε<ε˜, we have:
σ2C|τ−τε|H−δ≥k4ε(τ−τε)+ε.
For x≥0 consider the function
Fε(x)=k4εx−σ2CxH−δ+ε.
Let us show that there exists ε∗∈(0,ε˜) such that, for all ε<ε∗ and for all x≥0, Fε(x)>00$]]>. It is easy to check that Fε(0)=ε>00$]]> and Fε is convex on R+∖{0} (its second derivative is strictly positive on this set), so it is enough to examine the sign of the function in its critical points.
F′(x˜)=k4ε−σ(H−δ)2Cx˜H−δ−1=0⟹x˜=(k2σεC(H−δ))1/(H−δ−1)=(2σC(H−δ)k)1/(1+δ−H)ε1/(1+δ−H).
After some calculations we get
F(x˜)=12(2(H−δ)k)H−δ1+δ−H(σC)11+δ−H(H−δ−1)εH−δ1+δ−H+ε.
From the choice of δ it follows that H−δ1+δ−H>11$]]>, hence ∀K∈R∃ε∗>00$]]>:
ε−KεH−δ1+δ−H>0,∀ε<ε∗.0,\hspace{1em}\forall \varepsilon <{\varepsilon }^{\ast }.\]]]>
Choosing the corresponding ε∗ for
K:=−12(2(H−δ)k)H−δ1+δ−H(σC)11+δ−H(H−δ−1),
and choosing an arbitrary ε<min{ε˜,ε∗} we obtain that
Fε(x)>0∀x>0.0\hspace{1em}\forall x>0.\]]]>
However, from (14) it follows that
Fε(τ−τε)≤0.
The contradiction obtained proves the theorem for a>00$]]>. If a≤0, instead of (13) the following bound can be used:
kYs−aYs≥kε.
□
Hitting zero by the fractional CIR process in the case of H<1/2
The condition of H>1/21/2$]]> is essential for Theorem 2, as if H<1/2, the condition (16) does not hold. However, it is possible to obtain another result concerning zero hitting by the fractional CIR process in the case of H<1/2.
Let {BtH,t≥0} be the fractional Brownian motion with H<1/2 and let a∈R, σ>00$]]> be fixed. Consider the set of processes
Y:={Y(k)={Yt(k),t≥0},k>0},0\big\},\]]]>
each element of which starts from the same level Y0>00$]]>, satisfies the SDE of the form (5) before hitting zero and remains in zero after that moment:
Yt(k)(ω)=Y0+12∫0t(kYs(k)(ω)−aYs(k)(ω))ds+σ2dBtH(ω),ift<τ(k)(ω),0,ift≥τ(k)(ω),
where τ(k):=inf{t≥0∣Yt(k)=0}.
Letk1<k2. Then∀ω∈Ω,∀t≥0:
τ(k1)(ω)≤τ(k2)(ω);
Yt(k1)(ω)≤Yt(k2)(ω), and the inequality is strict fort∈(0,τ(k2)(ω)).
This lemma holds for an arbitrary Hurst index H∈(0,1).
Let ω∈Ω be fixed (we will omit ω in brackets in further formulas). Consider the function δ on the interval [0,min{τ(k1),τ(k2)}), such that δ(t)=Yt(k2)−Yt(k1). It is obvious that δ is differentiable, δ(0)=0 and
δ+′(0)=12(k2Y0−aY0)−12(k1Y0−aY0)=k2−k12Y0>0.0.\]]]>
As δ(t)=δ+′(0)t+o(t), t→0+, it is easy to see that there exists the maximal interval (0,t∗)⊂(0,min{τ(k1),τ(k2)}) such that δ(t)>00$]]> for all t∈(0,t∗). It is also clear that
t∗=sup{t∈(0,min{τ(k1),τ(k2)})∣∀s∈(0,t):δ(s)>0}.0\big\}.\]]]>
Assume that t∗<min{τ(k1),τ(k2)}. According to the definition of t∗, δ(t∗)=0. Hence Yt∗(k2)=Yt∗(k1)=Y∗>00$]]> and
δ′(t∗)=k2−k12Y∗>0.0.\]]]>
As δ(t)=δ′(t∗)(t−t∗)+o(t−t∗), t→t∗, there exists ε>00$]]> such that δ(t)<0 for all t∈(t∗−ε,t∗), that contradicts the definition of t∗.
Now it is easy to show that (i) holds: indeed, if τ(k1)>τ(k2){\tau }^{(k_{2})}$]]>, then
0=Yτ(k2)(k2)<Yτ(k2)(k1).
This means that ∃t∗<τ(k2) such that Yt(k2)<Yt(k1) for all t∈(t∗,τ(k2)), which contradicts (19).
Finally, as Yt(k2)>Yt(k1){Y_{t}^{(k_{1})}}$]]> for all t∈(0,τ(k1)), Yt(k2)>Yt(k1)=0{Y_{t}^{(k_{1})}}=0$]]> for all t∈[τ(k1),τ(k2)) and Yt(k2)=Yt(k1)=0 for all t≥τ(k2), (ii) also holds. □
Now let us move to the main result of the section.
For allT>00$]]>:P(τ(k)>T)→1,k→∞.T\big)\to 1,\hspace{1em}k\to \infty .\]]]>
The proof is by contradiction.
Assume that ∃T∗>00$]]>, ∃{kn,n≥1},kn↑∞ as n→∞ such that:
P(τ(kn)≤T∗)→α>0,n→∞.0,\hspace{1em}n\to \infty .\]]]>
Let us consider the case of a>00$]]>. Let Ω′ be from Proposition 1, and for all ε∈(0,min(Y0,1,k12a)) denote τε(kn):=sup{t∈(0,τ):Yt(kn)=ε} and
DT∗(kn):={ω∈Ω′∣τ(kn)≤T∗}.
According to Lemma 1, ∀n≥1:DT∗(kn+1)⊂DT∗(kn), so
P(⋂n≥1DT∗(kn))=limn→∞P(DT∗(kn))=α>0.0.\]]]>
Just like in Theorem 2, ∀n≥1, ∀ω∈DT∗(kn):
−ε=Yτ(kn)(kn)−Yτε(kn)(kn)=12∫τε(kn)τ(kn)(knYs(kn)−aYs(kn))ds+σ2(Bτ(kn)H−Bτε(kn)H).
The process Ys(kn)∈(0,ε) on the interval (τε(kn),τ(kn)), hence
knYs(kn)−aYs(kn)≥knε−aε,∀s∈(τε(kn),τ(kn)).
Let δ>00$]]> satisfy the condition 0<H−δ<1/2.
According to Proposition 1, ∃Cω=C(T∗,ω,δ), ∀0<s<t<T∗:
|BtH−BsH|≤Cω|t−s|H−δ.
As ε<k12a, the following inequality is true:
knε−aε>kn2ε∀n≥0,\frac{k_{n}}{2\varepsilon }\hspace{1em}\forall n\ge 0,\]]]>
so just like in the proof of Theorem 2 we can obtain that ∀n≥1, ∀ω∈DT∗(kn):
σ2Cω(τ(kn)−τε(kn))H−δ≥kn4ε(τ(kn)−τε(kn))+ε.
According to (22), ∀n≥1, ∀ω∈⋂n≥0DT∗(kn):
σ2Cω(τ(kn)−τε(kn))H−δ>ε.\varepsilon .\]]]>
However, it is easy to see from (22) that
σ2Cω(τ(kn)−τε(kn))H−δ>kn4ε(τ(kn)−τε(kn)).\frac{k_{n}}{4\varepsilon }\big({\tau }^{(k_{n})}-{\tau _{\varepsilon }^{(k_{n})}}\big).\]]]>
According to [17], E(|Cω|p)<∞ for all p∈[1,∞), so Cω is finite a.s.
Therefore, as P(⋂n≥0DT∗(kn))=α>00$]]>, ∃M>00$]]>, ∃E⊂⋂n≥0DT∗(kn), P(E)>00$]]> such that ∀ω∈E:
Cω<M.
Hence, as ε<1,
σ2Cω(τ(kn)−τε(kn))H−δ<C˜kn−H−δ1−H+δCω11−H+δεH−δ1−H+δ<C˜kn−H−δ1−H+δM11−H+δ<ε,
if kn>(C˜M11−H+δε)1−H+δH−δ{(\frac{\tilde{C}{M}^{\frac{1}{1-H+\delta }}}{\varepsilon })}^{\frac{1-H+\delta }{H-\delta }}$]]>, which contradicts (23).
If a<0, the following inequality can be used instead of (21):
knYs(kn)−aYs(kn)≥knε.
□
Appendix: Simulations of the fractional Cox–Ingersoll–Ross process
Theorems 2 and 3 can be illustrated by numerical simulations.
10000 sample paths of the fractional Cox–Ingersoll–Ross process were simulated on the interval [0,10] as the square of the process Y defined in (5). The Euler approximation of Y was used until the first zero hitting by the latter with the mesh of the partition of Δt=0.001:
Ytn=Ytn−1+12(kYtn−1−aYtn−1)Δt+σ2ΔBtnH,ifYtn−1>0,0,ifYtn−1≤0,Xtn=Ytn2.0,\\{} 0,\hspace{1em}& \text{if}\hspace{2.5pt}Y_{t_{n-1}}\le 0,\end{array}\right.\\{} \displaystyle X_{t_{n}}& \displaystyle ={Y_{t_{n}}^{2}}.\end{array}\]]]>
There were no zero hitting for 10000 trajectories simulated for four cases that satisfy the conditions of Theorem 2 (see Fig. 1, 2; the amount of trajectories on these and further figures is reduced in order to make them more convenient for the reader).
Case of a=1, k=1, σ=1, H=0.6, X0=1
Case of a=1, k=1, σ=1, H=0.8, X0=1
Case of a=1, k=0.5, σ=1, H=0.4, X0=1. 17% of sample paths hit zero
Case of a=1, k=1, σ=1, H=0.4, X0=1. Less than 1% of sample paths hit zero
Case of a=1, k=1, σ=2, H=0.4, X0=1. 86% of sample paths hit zero
Case of a=1, k=3, σ=2, H=0.4, X0=1. Less than 1% of sample paths hit zero
Case of a=1, k=1, σ=1, H=0.2, X0=1. 43% of sample paths hit zero
Case of a=1, k=3, σ=1, H=0.2, X0=1. Less than 1% of sample paths hit zero
However, the behavior of the fractional Cox–Ingersoll–Ross process is not completely clear for the situation of k>0,H<1/20,H<1/2$]]>. Simulations for different parameters k,H,σ (see Figures 3–8) show that in this case the process may hit zero with positive probability, however, as stated in Theorem 3, as k gets bigger, the amount of trajectories that hit zero tends to zero. Moreover, it seems that the less H and the bigger σ are, the bigger is the probability of reaching zero.
ReferencesAnh, V., Inoue, A.: Financial markets with memory I: Dynamic models. 23(2), 275–300 (2005). MR2130350Bollerslev, T., Mikkelsen, H.O.: Modelling and pricing long memory in stock market volatility. 73(1), 151–184 (2005). MR0054814Cheridito, P., Kawaguchi, H., Maejima, M.: Fractional Ornstein-Uhlenbeck processes. 8(1), 1–14 (2003).Cox, J.C., Ingersoll, J.E., Ross, S.A.: A re-examination of traditional hypotheses about the term structure of interest rates. 36, 769–799 (1981).Cox, J.C., Ingersoll, J.E., Ross, S.A.: An intertemporal general equilibrium model of asset prices. 53(1), 363–384 (1985).Cox, J.C., Ingersoll, J.E., Ross, S.A.: A theory of the term structure of interest rates. 53(2), 385–408 (1985). MR0785475Ding, Z., Granger, C.W., Engle, R.F.: A long memory property of stock market returns and a new model. 1(1), 83–106 (1993).Euch, O., Rosenbaum, M.: The characteristic function of rough Heston models. https://arxiv.org/pdf/1609.02108.pdf. Accessed 24 Sep 2017. arXiv:1609.02108Feller, W.: Two singular diffusion problems. 54, 173–182 (1951).Heston, S.L.: A closed-form solution for options with stochastic volatility with applications to bond and currency options. 6(2), 327–343 (1993).Leonenko, N., Meerschaert, M., Sikorskii, A.: Correlation structure of fractional Pearson diffusion. 66(5), 737–745 (2013). MR3089382Leonenko, N., Meerschaert, M., Sikorskii, A.: Fractional Pearson diffusion. 403(2), 532–546 (2013).Marie, N.: A generalized mean-reverting equation and applications. 18, 799–828 (2014). MR3334015Melnikov, A., Mishura, Y., Shevchenko, G.: Stochastic viability and comparison theorems for mixed stochastic differential equations. 17(1), 169–188 (2015). MR3306678Mishura, Y.: . Springer, Berlin (2008).Mishura, Y., Piterbarg, V., Ralchenko, K., Yurchenko-Tytarenko, A.: Stochastic representation and pathwise properties of fractional Cox-Ingersoll-Ross process (ukr). 97, 157–170 (2017).Nualart, D., Rascanu, A.: Differential equations driven by fractional Brownian motion. 53(1), 177–193 (2002).Yamasaki, K., Muchnik, L., Havlin, S., Bunde, A., Stanley, H.E.: Scaling and memory in volatility return intervals in financial markets. 102(26), 9424–9428 (2005).