Suppose one has a sample of high-frequency intraday discrete observations of a continuous time random process, such as foreign exchange rates and stock prices, and wants to test for the presence of jumps in the process. We show that the power of any test of this hypothesis depends on the frequency of observation. In particular, if the process is observed at intervals of length \(1/n\) and the instantaneous volatility of the process is given by \( \sigma _{t}\), we show that at best one can detect jumps of height no smaller than \(\sigma _{t}\sqrt{2\log (n)/n}\). We present a new test which achieves this rate for diffusion-type processes, and examine its finite-sample properties using simulations.
This implies that we consider only alternatives with a maximum number of jumps.
Chaboud et al. (2010) report that when the data are sampled at sufficiently high frequencies, in empirical practice most sampling intervals contain zero returns. At sufficiently high sampling frequencies, the likelihood of encountering two (or more) adjacent intervals with non-zero returns is very small. This feature of the data generating process imparts a downward bias on volatility estimates that are based on sums of products of absolute returns over adjacent time intervals.
For \(\widehat{V}_{p,k}\), ASJ suggest the following two estimators:
Our tests and other tests have some options to choose. So we consider several versions of those tests. LLP-4LN means our test where we choose four times log sample size as our averaging window. BNS-LIN means \(\hat{\tau }_{ BNS }^{ LIN }\) in Definition 1. ASJ-QV and ASJ-BPV means ASJ with \(\widehat{V}_{p,k}^{c}\quad \)and\(\quad \widetilde{V}_{p,k}^{c}\) in Definition 2, respectively. We choose \(p=4\) and \(k=2\) as recommended. LM-SQRT means LM test where we choose the square root of sample size as the averaging window.
The LM Test requires the averaging window the order of which is equal to or larger than the square root of sample size. We report the performance of LM-4LN and LM-2LN to show the problem of small averaging windows although they are not valid in their theory.
We are convinced that our result also holds for larger values of \( \varepsilon \). However, imposing this condition greatly simplifies one part of the proof.
References
Ait-Sahalia Y, Jacod J (2009) Testing for jumps in a discretely observed process. Ann Stat 37(1):184–222
Barndorff-Nielsen O, Shephard N (2002) Econometric analysis of realized volatility and its use in estimating stochastic volatility models. J R Stat Soc B 64(2):253–280
Barndorff-Nielsen O, Hansen P, Lunde A, Shephard N (2008) Designing realized kernels to measure the ex post variation of equity prices in the presence of noise. Econometrica 76(6):1481–1536
Chaboud A, Chiquoine B, Hjalmarsson E, Loretan M (2010) Frequency of observation and the estimation of integrated volatility in deep and liquid financial markets. J Empir Finance 17(2):212–240
We assume that the variance of the Wiener process \(W\) is known; without loss of generality, we may assume that this variance is equal to 1. Let \(P_{n}\) be the probability measure of \((X_{0},X_{1/n},X_{2/n},\dots ,X_{1})\) under the null, and \(Q_{n}\) be the measure under the alternative. Let the \(z_{i}\) be defined as
Because each \(z_{i}\) is standard normal, the expectation of each \(\exp \Big \{ (c_{n}\sqrt{n}\,)z_{i}-\frac{1}{2}(c_{n}\sqrt{n}\,)^{2}\Big \} \) term equals 1. Moreover,
Hence the variance of \(\ln (\mathrm d Q_{n}/\mathrm d P_{n})\) is smaller than \(\exp \left\{ (c_{n}\sqrt{n}\,)^{2}\right\} /n\), which converges to 0 if \(c_{n}\) satisfies Eq. (3).
Since we are interested primarily in small (but positive) values of \( \varepsilon \)—the smaller we choose \(\varepsilon \), the bigger are the permissible jumps—we may impose the condition thatFootnote 6
in probability. The term \(\mathrm d Q_{n}/\mathrm d P_{n}\) is a nonnegative random variable. Hence one may establish Eq. (13) by showing that its Laplace transform obeys
In the following, we will prove that this statement is correct. Let \(\Phi (\cdot )\) be the cumulative distribution function (cdf) of the standard normal. Then the cdf of \(\exp \left\{ z_{i}\sqrt{2(1-\varepsilon )\ln n} \,\right\} \) is given by
$$\begin{aligned} y=\frac{s}{n^{2-\varepsilon }}\,x,\hbox { and thus }\mathrm d x=\frac{ n^{2-\varepsilon }}{s}\,\mathrm d y\hbox { and }\ln x=\ln y+(2-\varepsilon )\ln n-\ln s \end{aligned}$$
Label the third term \(C_{n}\). Observe that \(C_{n}\) does not depend on \(y\); hence, when evaluating Eq. (17), it may be taken outside the integral. It is straightforward to rearrange \(C_{n}\) as
$$\begin{aligned} S_{n}=\frac{1}{\sqrt{2\pi }}\frac{n}{\sqrt{2(1-\varepsilon )\ln n}}\,\frac{1 }{s}\,C_{n}\int \limits _{0}^{\infty }y^{H_{n}}\frac{1-e^{-y}}{y}\exp \left\{ -\frac{ (\ln y)^{2}}{4(1-\varepsilon )\ln n}\right\} \mathrm d y. \end{aligned}$$
(19)
In order to evaluate the integral on the right-hand side of Eq. (19), we start by rewriting it as
$$\begin{aligned} \int \limits _{0}^{\infty }y^{H_{n}}\frac{1-e^{-y}}{y}\exp \left\{ -\frac{(\ln y)^{2} }{4(1-\varepsilon )\ln n}\right\} \mathrm d y \end{aligned}$$
We will show in the following that the integral labeled \(\mathcal{A }\) is asymptotically negligible relative to that labeled \(\mathcal{B }\). With the substitution
As the denominator in Eq. (21) diverges to infinity as \( n\rightarrow \infty \) for all fixed \(s>0\), we have established that the integral \(\mathcal{B }\)—which is also the numerator in Eq. (21)—diverges to infinity as \(n\rightarrow \infty \).
Next, we show that the integral \(\mathcal{A }\) above is \(O(1)\). First, observe that
Therefore, the limit of \(H_{n}\) lies between \(-2\) and \(-1\). Hence, there exist constants \(\alpha \) and \(\beta \), with \(-2<\alpha <\beta <-1\), such that for all but finitely many \(n\)
is easily seen to be uniformly bounded for all \(y\ge 1\), i.e., for all \( y\ge 1\) there exists an \(M\) such that
$$\begin{aligned} \left| \psi (y)\right| \le M \hbox { for all } y \ge 1. \end{aligned}$$
(24)
For the case \(0\le y\le 1\), we may use the power series representation for \(\exp (\cdot )\) to find that \(\psi (\cdot )\) is an analytic function as well and that \(\psi (0)=0\). Since any analytic function has derivatives that are bounded on any compact set, we deduce that, for \(0\le y\le 1\),
$$\begin{aligned} \left| \psi (y)\right| \le C y \end{aligned}$$
(25)
for some universal constant \(C\), i.e., that \(\left| \psi (y)\right| \) is bounded by a linear function in \(y\).
Finally, the third term in the integrand of the integral \(\mathcal{A }\) above, viz.,
may easily be shown to be smaller than 1 in absolute value for all values of \(y\) and \(n>1\).
We now combine the results for the three terms that make up the integrand of integral \(\mathcal{A }\). To evaluate the integral, one needs to consider separately the regions \(0\le y\le 1\) and \(y\ge 1\).
of the density ratio converges to \(e^{-s}\), which is the Laplace transform of a measure concentrated at 1. We have thus shown that the density ratio \( \mathrm d Q_{n}/\mathrm d P_{n}\) converges in distribution to a constant, viz., 1. By (Feller (1971), Theorem 2, p. 431), it also converges in probability to 1. Put differently, for any arbitrary \(\eta >0\),
Because \(A_{n}\) is an arbitrary sequence of events, we further deduce that the total variation between \(P_{n}\) and \(Q_{n}\) converges to 0. Hence, for all measurable functions \(\varphi _{n}\) with \(0\le \varphi _{n}\le 1\), we have
$$\begin{aligned} \int \varphi _{n}\mathrm d P_{n}-\int \varphi _{n}\mathrm d Q_{n}\rightarrow 0. \end{aligned}$$
But this is exactly what we had to show: For every sequence of tests, their power under the null, \(P_{n}\), is asymptotically equal to that under the alternative, \(Q_{n}\). \(\square \)
It may be seen immediately that the \(\tau _{i}\) are \(\mathcal{F }_{i}-\) measurable. We will repeatedly apply the optional sampling theorem for various stopping times. Let \(\varepsilon >0\) be arbitrary and let \( M(\varepsilon )\) be defined by Eqs. (9)–(11) in Appendix 1.
Let us define the stopping time \(\nu \) as the first index \(m\le n-1\) such that either
Observe that \(\nu \) is indeed a stopping time adapted to \(\mathcal{F }_{i}\): For any \(i\le m+1\), both \(E\left( (\tau _{i}\le K_{n}^{*})\mid \mathcal{F }_{i-1}\right) \) and \(\hat{\sigma }_{m+1}^{2}\) are \(\mathcal{F }_{m}\) -measurable, and hence we also find that the event
$$\begin{aligned} \sum _{i=\ell +1}^{n}Y_{i}\rightarrow c \end{aligned}$$
(34)
in probability. By the definition of \(K_{n}^{*}\), \(EY_{i}=c/n\). Moreover, we know that \(\hat{\sigma }_{i}^{2}\) is distributed according to a scaled \(\chi ^{2}\) distribution with \(\ell \) degrees of freedom. Hence it is an elementary exercise to show that \(EY_{i}^{2}=O(1/n^{2})\) and that \(Y_{j}\) and \(Y_{k}\) are independent if
Moreover, Eq. (30) implies that \(P\left\{ \sum _{i=\ell +1}^{\nu }Y_{i}=\sum _{i=\ell +1}^{n}Y_{i}\right\} \rightarrow 1\). Therefore, \(\sum _{i=\ell +1}^{\nu }Y_{i}\rightarrow c\) as well. Hence it can be seen that Eqs. (36) and (37) allow us to deduce from Eq. (35) that
Repeating the preceding argument with \(-\int _{a}^{b}\!f\mathrm d W\) completes the proof. \(\square \)
Returning to the Proof of Theorem 3 , we have
$$\begin{aligned} d\mu _{t}=A_{t}\mathrm d t+B_{t}\mathrm d V_{t}^{(1)} \end{aligned}$$
and
$$\begin{aligned} d(\log \sigma _{t})=C_{t}\mathrm d t+D_{t}\mathrm d V_{t}^{(2)}, \end{aligned}$$
where \(A_{t}\), \(B_{t}\), \(C_{t}\), and \(D_{t}\) are continuous processes and \( V_{t}^{(1)}\) and \(V_{t}^{(2)}\) are standard-scale Wiener processes. We begin by demonstrating that we may assume without loss of generality that the processes \(A_{t}\), \(B_{t}\), \(C_{t}\), \(D_{t}\), \(\mu _{t}\), and \(\log \sigma _{t}\) are uniformly bounded. Because \(A_{t}\), \(B_{t}\), \(C_{t}\), \(D_{t}\), \( \mu _{t}\), and \(\ln \sigma _{t}\) are continuous, for every \(\varepsilon >0\) there exists an \(M=M(\varepsilon )\) such that
We define the stopping time \(\tau ^{(\varepsilon )}\) as the first value of \( t \) when one of \(A_{t}\), \(B_{t}\), \(C_{t}\), \(D_{t}\), \(\mu _{t}\), and \(\ln \sigma _{t}\) becomes larger in absolute value than \(M(\varepsilon )\); if the absolute values of these processes remain below \(M(\varepsilon )\) all the time, we set \(\tau ^{(\varepsilon )}=1\). Then
By definition, \(\rho _{n}\) and \(\xi _{n}\) are our test statistics applied to \(X_{i/n}\) and \(W_{i/n}\), respectively. Moreover, Eq. (38) guarantees that
Hence it is sufficient to show that, for all \(\varepsilon >0\), the difference between \(\xi _{n}^{(\varepsilon )}\) and \(\rho _{n}^{(\varepsilon )}\) converges to zero. To show this, observe that
Observe that \(\ln (\sigma _{i/n}^{2})-\ln (\sigma _{(i-k)/n}^{2})=\int _{(i-k)/n}^{i/n}C_{t}\mathrm d t+D_{t}\mathrm d V_{t}^{(2)}\). For \(i<\tau ^{(\varepsilon )}\),
we apply Lemma 5. Since \(\sigma _{u}\) is a diffusion process with drift and diffusion coefficients that are assumed to be bounded, we may deduce that for all \(\alpha >0\) there exists an \(M\) such that for all \(i\) and all \(u\in [(i-1)/n,i/n]\),
To apply Lemma 5, we need to guarantee that the integral
$$\begin{aligned} \int \limits _{(i-1)/n}^{i/n}\left( \sigma _{u}-\sigma _{(i-1)/n}\right) ^{2}\mathrm d u \end{aligned}$$
is uniformly bounded. But the existence of such a bound may be established using a stopping time argument. Put \(i/n\ge S\ge (i-1)/n\). We stop the process at time \(S\) if for the first time
Hence it is sufficient to give estimates for \(\int _{(i-1)/n}^{i/n}(\sigma _{u}^{*}-\sigma _{(i-1)/n}^{*})\mathrm d W_{u}\). For this task, however, we may apply Lemma 5. We deduce that
Lee, T., Loretan, M. & Ploberger, W. Rate-optimal tests for jumps in diffusion processes.
Stat Papers54, 1009–1041 (2013). https://doi.org/10.1007/s00362-013-0541-y