Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.
127,586 characters · 5 sections · 2 citation commands
Inference from high-frequency data: A subsampling approach
\thispagestyle{empty}
\@startsection {section}{1} {0mm} {-\baselineskip}{0.2\baselineskip} {\bf}{Introduction}
\setcounter{page}{1}
Volatility is a key ingredient in the assessment and prediction of financial risk, be it in asset- and derivatives pricing \citep*[e.g.,][]{black-scholes:73a, sharpe:64a}, portfolio selection \citep*[e.g.,][]{markowitz:52a}, or risk management and hedging \citep*[e.g.,][]{jorion:06a}.
Around the turn of the millennium, the advent of financial high-frequency data led to a surge in the nonparametric measurement of volatility \citep*[see, e.g.,][]{andersen-bollerslev-diebold:10a, barndorff-nielsen-shephard:07a}. High-frequency data are recorded at the tick-by-tick level and store information about the time, price (i.e., a bid-ask quote or transaction price), and size of individual orders and executions. In theory, the harnessing of high-frequency information leads to a perfect, error-free measure of ex-post volatility via the realized variance; a sum of squared intraday log-returns \citep*[e.g.,][]{andersen-bollerslev:98a, barndorff-nielsen-shephard:02a}.
After the initial---pioneering---work, the literature turned towards addressing two inherent shortcomings of realized variance. Firstly, realized variance can only estimate quadratic variation, and it does not separate continuous, diffusive volatility from discontinuous jump risk. This motivated the development and application of estimators that can robustly measure very general functionals of volatility, also in the presence of jumps \citep*[e.g.,][]{ait-sahalia-jacod:12a, andersen-dobrev-schaumburg:12a, barndorff-nielsen-shephard:04b, barndorff-nielsen-graversen-jacod-podolskij-shephard:06a, christensen-oomen-podolskij:10a, corsi-pirino-reno:10a, mancini:09a, mykland-shephard-sheppard:12a}. Secondly, when applied to tick-by-tick data realized variance is severely biased by common sources of noise, which form an integral part of any realistic model for securities' prices \citep*[e.g.,][]{hansen-lunde:06b, zhou:96a}. This paved the way for the next cohort of estimators that were designed to be more resistant to noise, e.g., \citet*{ait-sahalia-mykland-zhang:05a, barndorff-nielsen-hansen-lunde-shephard:08a, jacod-li-mykland-podolskij-vetter:09a, podolskij-vetter:09a, zhang:06a}. As such, much progress has been made and today there is no shortage of estimators that can provide consistent estimates of volatility functionals in various contexts (either plain vanilla, or robustly to jumps or noise---or both).
The large battery of estimators at our disposal also brings with it an increasing demand for assessing estimation errors and drawing inference about volatility---e.g., in the form of confidence intervals or hypothesis tests. This is because whether the sample is small or large, as long as it is finite, there is necessarily some sampling error left in the estimate, and when confidence intervals are computed in practice, high-frequency estimators of volatility are often found to contain sizable errors \citep*[e.g.,][]{barndorff-nielsen-hansen-lunde-shephard:08a}. The distinction between a realized measure of volatility and its population target is critical, because failing to properly take sampling uncertainty into account can severely distort parameter estimation of stochastic volatility models and be detrimental to the construction and evaluation of forecasts of volatility \citep*[e.g.,][]{andersen-bollerslev:98a, andersen-bollerslev-meddahi:05a,andersen-bollerslev-meddahi:11a, hansen-lunde:06a,hansen-lunde:14a,patton:11a}.
There are several problems associated with drawing inference about volatility functionals in high-frequency data. The first and foremost is of course to figure out the relevant distribution theory. The next hurdle is then to find a good proxy for the asymptotic (conditional) variance of the estimator. This is a formidable challenge in practice, because the asymptotic variance often depends on parameters that are substantially more difficult to back out from the available sample of high-frequency data. The expression for the asymptotic variance typically also rests heavily on the properties of the data and it is bound to change depending on these. This is an unpleasant concern with real high-frequency data, which are contaminated by market microstructure frictions. While the noise is often assumed to be i.i.d. and independent of the efficient price, there is some empirical and theoretical support for a serially correlated, heteroscedastic, and, potentially, endogenous noise process at the tick level \citep*[e.g.,][]{ait-sahalia-mykland-zhang:11b,diebold-strasser:13a, hansen-lunde:06b,kalnina-linton:08a}. An estimator of the asymptotic variance designed for i.i.d. and independent noise can not be expected to give valid inference, if the underlying conditions are violated. In practice, it is not trivial to verify the conditions imposed on the noise \citep*[e.g.,][]{hautsch-podolskij:13a}, which makes it more pressing to find estimators that are robust against modeling criteria. Finally, in multivariate analysis, inference would at some stage require an estimate of the asymptotic covariance matrix. Here, the proposed estimator should ideally be positive semi-definite, while, in contrast, some existing estimators of the asymptotic covariance matrix in the high-frequency setting are not assured to be that. As we show in this paper, this runs smack into problems in practically relevant and realistic settings (see Table (ref) in Section (ref)).
In this paper, we propose to use subsampling for assessing the uncertainty embedded in high-frequency estimation of functionals of financial volatility. Subsampling is based on creating several---properly rescaled---estimates of the parameter(s) of interest using local stretches of sample data and then studying the sampling variation of these. It was originally developed in the context of stationary time series in the long-span domain \citep*[e.g.,][]{politis-romano:94a, politis-romano-wolf:99a}. The term appeared in the high-frequency literature in \citet*{zhang-mykland-ait-sahalia:05a}, who proposed a two-scale realized variance based on price subsampling. This is different from traditional subsampling and actually does not work for asymptotic variance estimation, because it leads to an overlapping samples problem in the subsampled returns, causing the subsample estimates to be too strongly correlated in large samples. This was pointed out by \citet*{kalnina-linton:07a} and \citet*{kalnina:11a}, who propose an inference strategy based on various alternative subsampling schemes, which lead to better asymptotic properties. \citet*{kalnina:15a} extends these ideas to inference about a multivariate parameter, while \citet*{ikeda:16a} and \citet*{varneskov:16a} consider subsample estimation of the asymptotic variance of the realized kernel.
As an inferential tool, subsampling has several attractive features from a practical point of view. First, subsampling is intuitive and relatively easy to compute, because it does not require an extra set of estimators; it uses copies of the original statistic. Second, in the multivariate context, it leads to variance-covariance matrix estimates that are positive semi-definite by construction. And, third, subsampling does not explicitly take the structure of the asymptotic variance into account. It is to a large degree automatic and has an innate ability to adapt to the problem at hand, which makes it highly robust against design criteria, as shown by \citet*{kalnina:11a}. This type of analysis, where inference is effectively carried out by bypassing the asymptotic variance, is also emphasized by mykland-zhang:17a, who propose a so-called Observed Asymptotic Variance, which, as our approach, is based on the comparison of adjacent estimators.
This paper builds on these ideas. It contributes to extant literature in several directions. First, we propose to subsample bipower variation as a means to estimate the asymptotic variance-covariance matrix of this statistic. We devise an estimator, which involves fewer tuning parameters compared to \citet*{kalnina:11a}. Second, we derive an asymptotic theory within this framework in both frictionless and noisy markets. We show our estimator is consistent under weak assumptions on the data-generating process, accommodating jumps in the price and volatility, while allowing the noise to be either heteroscedastic or autocorrelated. Third, with stronger conditions, we provide a decomposition of the leading errors of the subsampler, from which we get insights about how to configure it by optimally choosing its tuning parameters (e.g., the number of subsamples). This yields a rate of convergence for our statistic; a result that has---to the best of our knowledge---not been derived in earlier work. It reveals that the robustness of subsampling is not free of charge, but leads to a loss of efficiency compared to existing estimators in the form of a slower rate of convergence. It implies a trade-off in that if, for example, one is prepared to use an estimator, which is not positive semi-definite, a better rate can potentially be achieved. Or, if prior knowledge about the asymptotic variance matrix is available or parametric assumptions can be verified from the data, it is typically better to construct estimators which attempt to exploit that information relative to doing subsampling.\footnote{To paraphrase \citet*{politis-romano-wolf:99a}, subsampling is “a robust starting point toward even more refined procedures.”} Still, in finite samples we show in a realistic setting with microstructure noise the subsampler produces convincing results compared to some available alternatives.
The rest of this paper goes as follows. Section 2 introduces the setting. In Section 3, we derive the theory first without and then with noise. In Section (ref), we do numerical simulations in order to inspect the finite sample performance of our estimator. In Section (ref), we confront our framework with some real high-frequency data, while the Appendix contains the proofs of our results.
\@startsection {section}{1} {0mm} {-\baselineskip}{0.2\baselineskip} {\bf}{Theoretical framework}
We consider a scalar process $X = (X_{t})_{t \geq 0}$, which represents the log-price of some financial security. It is defined on a filtered probability space $(\Omega, \mathcal{F}, (\mathcal{F}_{t})_{t\geq 0}, \mathbb{P})$ and adapted to $(\mathcal{F}_{t})_{t \geq 0}$. We assume that $X$ can be described by a continuous It\^{o} semimartingale (or stochastic volatility model), as expressed by the equation:
where $X_{0}$ is the starting price, $a = (a_{t})_{t \geq 0}$ is a predictable and locally bounded drift process, $\sigma = (\sigma_{t})_{t \geq 0}$ is an adapted, c\`{a}dl\`{a}g volatility process, while $W = (W_{t})_{t \geq 0}$ is a standard Brownian motion.\footnote{A basic result in financial economics states that if $X$ is drawn from an arbitrage-free, frictionless market, it necessarily has a semimartingale structure \citep*[e.g.,][]{back:91a,delbaen-schachermayer:94a}. If further $X$ has continuous sample paths, then weak assumptions ensure that $X$ can be represented as in Eq. (ref). Later on, in Section (ref), we provide further details about the robustness of our results, when $X$ exhibits jumps. We cover the noisy setting with market frictions in Section (ref).}
In these models, natural measures of ex-post volatility can be written, for some suitable function $f$,
i.e. integrated functions of the diffusion coefficient.
We also point out that $a$ and $\sigma$ are left unspecified in this paper, and our results are completely nonparametric (within this class of models). While we do not impose assumptions a priori, we sometimes need to add additional, weak regularity conditions on $\sigma$, which nevertheless allow for very complex dynamics in the volatility process.\\[-0.50cm]
Assumption (V): $\quad \sigma$ is of the form:
where $\sigma_{0}$ is its initial value, $\tilde{a} = (\tilde{a}_{t})_{t \geq 0}$, $\tilde{ \sigma} = (\tilde{ \sigma}_{t})_{t \geq 0}$ and $\tilde{v} = (\tilde{v}_{t})_{t \geq 0}$ are adapted, c\`{a}dl\`{a}g stochastic processes, while $B = (B_{t})_{t \geq 0}$ is a standard Brownian motion that is independent of $W$. Furthermore, $(E, \mathcal{E})$ is a Polish space, $\tilde{ \mu}$ is a random measure on $\mathbb{R}_{+} \times E$, which is independent of $(W,B)$ and has an intensity measure $\tilde{ \nu}( \text{d}s, \text{d}x) = \text{d}s \tilde{F}( \text{d}x)$, where $\tilde{F}$ is a $\sigma$-finite measure on $(E, \mathcal{E})$. Also, $\tilde{ \delta}: \Omega \times \mathbb{R}_{+} \times E \to \mathbb{R}$ is a predictable function and $(S_{k})_{k \geq 1}$ is a sequence of stopping times increasing to $\infty$ such that $| \tilde{ \delta}( \omega,s,x)| \wedge 1 \leq \tilde{ \psi}_{k}(x)$ for all $( \omega, s, z)$ with $s \leq S_{k}( \omega)$ and $\int_{E} \tilde{ \psi}^{2}_{k}(x) \tilde{F}( \text{d}x) < \infty$ for all $k \geq 1$. \\[-0.50cm]
We are in the high-frequency setting. We suppose that historical data of $X$ is available in the time frame $[0,1]$, i.e. we set $t = 1$. In this interval, we assume that $X$ is recorded at equidistant time points $t_{i} = i / n$, for $i = 0, 1, \ldots n$, so that $n + 1$ is the total number of log-price observations in the sample. We define the $n$ increments, or log-returns, of $X$ as:
The asymptotic theory we derive below is then infill, i.e. we are at some point going to let $n \to \infty$.
To maintain a streamlined exposition, we only study the univariate setting in this paper. All of our theoretical results extend directly and without any changes (apart from additional notation) to multivariate $X$, if the sampling times are equidistant, as above, and recorded synchronously across assets, see, e.g., \citet*{kalnina:15a} for related research in that direction. If the high-frequency data are randomly spaced and asynchronous, our results are probably still true using a previous-tick rule to align prices and under suitable assumptions on the regularity on the observation grid of individual assets, e.g., \citet*{christensen-podolskij-vetter:13a}.
\@startsection {subsection}{1} {0mm} {-\baselineskip}{0.2\baselineskip} {\bf}{Bipower variation}
The econometric challenge is that the objects appearing in Eq. (ref) are latent, but they can be estimated from the available sample of high-frequency data in Eq. (ref). A popular statistic, which is well-suited to do this, is the bipower variation of \citet*{barndorff-nielsen-shephard:04b}.\footnote{Note that bipower variation is nested within a broader class of multipower variations, which adds additional lags in Eq. (ref), see, e.g, \citet*{barndorff-nielsen-shephard:04b}. Our theoretical results extend to this framework.} Here, we adopt the more general definition of bipower variation from \citet*{barndorff-nielsen-graversen-jacod-podolskij-shephard:06a}:\footnote{As we frequently cite this paper, it will henceforth be referred to as BGJPS6.}
where $f = (f_{1}, \ldots, f_{m})'$ and $g = (g_{1}, \ldots, g_{m})'$ are $\mathbb{R}^{m}$-valued functions. Note that in Eq. (ref) the multiplication is understood to be done element-by-element. Moreover, the components of $f = (f_{1}, \dots, f_{m})'$ and $g = (g_{1}, \ldots, g_{m})'$ are assumed to fulfill a condition, which we state with a generic function $h$.\\[-0.50cm]
Assumption (K): $\quad$ $h: \mathbb{R} \mapsto \mathbb{R}$ is even and continuously differentiable. Moreover, both $h$ and its derivative $h'$ are at most of polynomial growth.\\[-0.50cm]
Assumption (V) and (K) are standard conditions for the validity of central limit theorems for classical high-frequency statistics; see, e.g., BGJPS6. The following proposition, which is adapted from that paper, then describes the limiting properties of bipower variation.
The proposition provides the foundation for making inference about bipower variation. It shows that $V(f,g)^{n}$ is consistent for $V(f,g)$. The asymptotic distribution of $V(f,g)^{n}$ is mixed normal, i.e. it has a random variance-covariance matrix $\Sigma$, which is independent from the randomness of the normal distribution. To transform this into a probabilistic statement based on the standard normal distribution, it would be tempting to look at Eq. (ref) and deduce that, assuming $\Sigma$ is invertible, $\Sigma^{-1/2} \sqrt{n} \Bigl( V(f,g)^{n} - V(f,g) \Bigr) \overset{d}{\to} \text{\upshape{N}}(0, I_{m})$, where “$\overset{d}{\to}$” is convergence in law and $I_{m}$ is the $m$-dimensional identity matrix. We cannot readily jump to this conclusion from the regular definition of convergence in law, however, unless $\Sigma$ is constant. In our setting, where $\Sigma$ is stochastic, we therefore exploit an alternative, stronger type of convergence, which recovers the above feature.
More formally, let $Z_{n}$ be a sequence of random variables. Also, let $Z$ be a random variable, which is defined on some appropriate extension $(\Omega', \mathcal{F}', \mathbb{P}')$ of the original space $(\Omega, \mathcal{F}, \mathbb{P})$. $Z_{n}$ is then said to converge stably to $Z$, and we shall write $Z_{n} \overset{d_{s}}{\to} Z$, if for any bounded, continuous function $h$ and any bounded, $\mathcal{F}$-measurable random variable $Y$, the convergence $\mathbb{E} \bigl[Yh(Z_{n}) \bigr] \to \mathbb{E}' \bigl[Yh(Z) \bigr]$ holds, as $n \to \infty$. It follows that stable convergence implies convergence in law by taking $Y = 1$. We refer to \citet*{jacod-protter:12a} for further details of this concept.
Proposition (ref) is based on Assumption (K), which assumes $f$ and $g$ are differentiable. It can be extended to non-differentiable functions, given Assumption (H') and (K') from BGJPS6. Assumption (H') says $\sigma$ should be bounded away from zero, while Assumption (K') puts restrictions on the set, where $f$ and $g$ are not differentiable. We refer to BGJPS6 for a concise, mathematical statement of these conditions. Throughout the paper, we explain how this weaker setting affects our results.
A prime example that falls in the latter group is the original bipower variation of \citet*{barndorff-nielsen-shephard:04b}. It is based on the summation of products of the absolute value of adjacent high-frequency returns and sets $f_{k}(x) = |x|^{q_{k}}$ and $g_{k}(x) = |x|^{r_{k}}$, for $1 \leq k \leq m$ and $q_{k}, r_{k} \geq 0$. As it is extensively used in applied work, we sometimes restrict attention to this class of estimators below. To distinguish this “pure” bipower variation from the general one, we write this version as $V(q,r)^{n}$, where $q = (q_{1}, \ldots, q_{m})'$ and $r = (r_{1}, \ldots, r_{m})'$ are $m$-dimensional vectors, whose coordinates index the powers:
where $\mu_{q} = \mathbb{E}[|Z|^{q}]$ and $Z \sim \text{N}(0,1)$. Note that while $f_{k}(x) = |x|^{q_{k}}$ and $g_{k}(x) = |x|^{r_{k}}$ are not differentiable at $x = 0$, if $q_{k} \leq 1$ or $r_{k} \leq 1$, Proposition (ref) is nevertheless still true, if $\sigma > 0$.
Here, $\Sigma$ has the form:
and the link between Eq. (ref) and Eq. (ref) is made apparent.
\@startsection {section}{1} {0mm} {-\baselineskip}{0.2\baselineskip} {\bf}{Subsample estimation of $\Sigma$}
In order to do inference about bipower variation based on Proposition (ref), we also need to estimate $\Sigma$. Although there are some existing estimators out there, they are typically not general enough, or else they are inherently flawed, as outlined below.
For example, assume the goal is to estimate the asymptotic conditional variance-covariance matrix of the pure bipower variation estimator $V(q_{k},r_{k})^{n}$. The structure of $\Sigma$ unveiled by Eq. (ref) implies that here we can do estimation componentwise with a second bipower variation:
where $c_{ij}$ is a constant that can be determined from Eq. (ref) -- (ref). This is a direct element-by-element calculation, which also works for the general case of Eq. (ref) with some minor modifications. However, this procedure is difficult to adapt to the noisy setting.
Another idea, which is more generally applicable, is to write
and set
Then, $\tilde{ \Sigma}^{n} \overset{p}{ \to} \Sigma$ as given by Eq. (ref). In contrast to Eq. (ref), the latter estimator is “automatic,” as it does not depend on explicit knowledge of the asymptotic covariance matrix. It works, because the summands $\gamma_{i}^{n}(k)$ are asymptotically 1-dependent, and the estimator implicitly incorporates covariances up to lag one (hence the index bounds in the last summation). Indeed, we compare a vectorized version of $\tilde{ \Sigma}^{n}$ to our subsampler in Section (ref). But, while both of the above estimators are $\sqrt{n}$-consistent, neither is positive semi-definite in finite samples. In the noisy setting, this type of construction has been proposed by \citet*{podolskij-vetter:09a}, which we thoroughly analyze in Section (ref).
In this paper, we propose to estimate $\Sigma$ by subsampling of high-frequency data. This was suggested for the asymptotic variance of realized variance in \citet*{kalnina-linton:07a} and its noise-robust two-scale version in \citet*{kalnina:11a}, following earlier work in the classical time series literature \citep*[e.g.,][]{politis-romano:94a,politis-romano-wolf:99a}. Subsampling is an attractive procedure. First, all it does is to compute rescaled copies of the original statistic based on local stretches of high-frequency data, and then it studies the sampling variation of these. This makes it highly intuitive and trivial to implement. Second, it leads to estimators of $\Sigma$ that are positive semi-definite by construction. This is critical in practice, as shown in Section (ref). And third, it has an innate robustness against the statistical properties of microstructure noise, as discussed in Section (ref).
\@startsection {subsection}{1} {0mm} {-\baselineskip}{0.2\baselineskip} {\bf}{Subsampling for power variation}
In order to develop a complete theory for positive semi-definite estimation of the asymptotic, conditional covariance matrix $\Sigma$ of bipower variation using subsampling, we will start by analyzing the power variation case, i.e. $f = (f_{1}, \ldots, f_{m})'$. This should not only help to build intuition for the general case, but it turns out the results are interesting in their own right, because the rates of convergence differ according to whether we consider power variation, or multipower variation of order bigger than one.
Thus, we define the power variation estimator:
Then, we propose to set:
where, assuming $L$ divides $n,$
In Figure (ref), we illustrate the construction of the subsamples for power variation. As the figure shows, we start by splitting the full sample of high-frequency data into $L$ smaller subsamples. We assign log-returns successively to each subsample, starting from the first and going back to it after the $L$th subsample has been reached. We continue this process until the entire sample is exhausted. The $l$th subsample therefore consists of the increments $\bigl( \Delta_{(i - 1)L + l}^{n} X \bigr)_{i = 1, \ldots, n/L}$.
We then compute the statistics $V_{l}(f)^{n}$, which is the power variation estimator based on the $l$th subsample. To estimate $\Sigma$, we look at the sampling variation of the $V_{l}(f)^{n}$'s, after they have been suitably demeaned and scaled, with a standard sample covariance matrix.
As we prove in a moment, we can choose the number of subsamples $L$, such that $\hat{ \Sigma}_{n}$ is a consistent estimator of $\Sigma$. Intuitively, under suitable conditions on $L$, $V_{l}(f)^{n} \overset{p}{\to} V(f)$, where $V(f) = \int_{0}^{1} \rho_{\sigma_{s}} (f) \text{d}s$, and its asymptotic distribution (more or less) follows from Proposition (ref), except that its rate of convergence is $(n/L)^{-1/2}$, i.e. $\displaystyle \sqrt{\frac{n}{L}} \Bigl(V_l(f)^{n} - V(f) \Bigr) \overset{d_{s}}{\to} \text{\upshape{MN}}(0,\Sigma)$. Moreover, as each subsample is based on non-overlapping increments, the $V_{l}(f)^{n}$'s are, asymptotically, conditionally independent. This suggests that by averaging the sum of outer products of $\displaystyle \sqrt{\frac{n}{L}} \Bigl( V_{l}(f)^{n} - V(f) \Bigr)$, we should get a consistent estimator of $\Sigma$, which is then positive semi-definite by construction.
While $V_{l}(f)^{n}$ should in principle be centered around $V(f)$, the whole problem to begin with is of course that $V(f)$ is latent. So it has to be replaced by a consistent estimator to get a feasible estimator of $\Sigma$ that can be computed from data. In Eq. (ref), we plug in the full sample power variation estimate $V(f)^{n}$. This does not affect the asymptotics, because $V(f)^{n}$ converges much faster than $V_{l}(f)^{n}$.
We pause here for a moment to reflect a bit on the setup. It turns out that the assumptions of Proposition (ref) are, by construction, necessary to prove the consistency of $\hat{ \Sigma}_{n}$. Moreover, as we show later, they are also sufficient, if $L \rightarrow \infty$ and $n/L \rightarrow \infty$. Indeed, an underlying principle of this paper is that, as long as an associated central limit theorem holds, subsampling can be used to consistently estimate the asymptotic covariance matrix under minimal conditions on the tuning parameters. To be able to derive a convergence rate for $\hat{ \Sigma}_{n}$ and provide an optimal choice of the parameter $L$, however, we need some additional structure, which is not standard in the high-frequency literature.
First, we are going to assume that all of the driving terms in both $X$ and $\sigma$ can be modeled as Brownian semimartingales, i.e. we shall require that: \\[0.50cm] Assumption (H): $\quad$ $\sigma$ is continuous and follows Assumption (V), and each of $a$, $\tilde{a}$, $\tilde{ \sigma}$ and $\tilde{v}$ is continuous of the form in Eq. (ref).\\[-0.50cm]
The second condition we impose is a highly technical requirement that concerns the Malliavin smoothness of the random variables appearing in Assumption (H). We summarize some selected elements and notations from Malliavin calculus, which are relevant to our paper, in Section (ref) at the end of the Appendix, while the list of necessary conditions for the main text are comprised as Assumption (M).\\[-0.50cm]
Assumption (M): $\quad$ It holds that for any $0 \leq t \leq r \leq s$: $\sigma_{s}$, $\tilde{ \sigma}_{s}$, $\tilde{v}_{s}$, $D_{t}( \sigma_{s})$, $D_{t}( \tilde{ \sigma}_{s})$, $D_{t}( \tilde{v}_{s}) \in \mathbb{D}_{1,2}$ with
Moreover, $f \in C^{3}( \mathbb{R})$, while $f$, $f'$, $f''$ and $f'''$ exbibit polynomial growth. \\[0.50cm] Note that if the involved processes are solutions to stochastic differential equations, then condition (ref) is fulfilled given a sufficient smoothness of the corresponding drift term and volatility function. We refer to Eq. (ref) for the computation of the Malliavin derivative in this case.
The next result then leads to a rate of convergence for $\hat{ \Sigma}_{n}$.
Theorem (ref) presents the leading errors inherent in $\hat{ \Sigma}_{n}$. The first term, $1/\sqrt{L}$, intuitively follows from a central limit theorem result, because $\hat{ \Sigma}_{n} $ is an empirical mean of $L$ asymptotically, conditionally independent statistics. However, it is not easy to apply this relationship for a formal derivation of the error rate. The second error is more subtle. It comes from freezing the volatility process at the beginning of a subblock of length $L/n$. If volatility is assumed to be H\"older continuous of order $\alpha \in (0,1]$, a rough estimate implies an error rate of $(L/n)^{ \alpha}$. However, due to the semimartingale structure of $\sigma$, we can improve this to $L/n$ by applying a more refined estimation technique. As such, we should point out that the proof of Theorem (ref) is much more complex compared to subsampling of i.i.d. observations.
We find the fastest rate of convergence by balancing both errors. This requires:
such that
In a general model (e.g., where $\sigma$ is a diffusion process), we believe this rate is sharp and cannot be improved within the context of subsampling high-frequency data, but we shall not attempt to prove it. Of course, if we impose stricter, parametric assumptions, such as $\sigma_{t} = \sigma$ is constant and $\mu_{t} = 0$, the rate is faster and gets arbitrarily close to $n^{-1/2}$, as here the additional blocking error in Theorem (ref) drops out, and then we can take $L = O(n^{1-\epsilon})$, with $\epsilon > 0$ arbitrarily small.
We can combine the consistency of $\hat{ \Sigma}_{n}$ from Theorem (ref) with the convergence in distribution in Eq. (ref). If we then appeal to the properties of stable convergence, we get the feasible result:
which can be used to construct confidence intervals for $V(f)$ or do hypothesis testing. If the convergence had not been stable in law, this result would not follow in general.
Theorem (ref) is proved under the assumption that $\sigma$ is continuous---as are all coefficients of the model---and $f$ is differentiable. We note again the stable central limit theorem of Proposition (ref) is also valid for a non-differentiable function $f$, and possibly discontinuous volatility process, given Assumption (H') and (K') from BGJPS6, but it appears out of reach to derive a convergence rate for $\hat{ \Sigma}_{n}$ here. We can nonetheless show that $\hat{ \Sigma}_{n}$ still converges in probability to $\Sigma$ under these weaker conditions, which is relevant for applied work.
\@startsection {subsection}{1} {0mm} {-\baselineskip}{0.2\baselineskip} {\bf}{Subsampling for bipower variation}
In the previous section, we presented a subsampling estimator for the asymptotic conditional covariance matrix of power variation. If we are interested in bipower (or multipower) variation, the theory derived there does not readily apply. This is because the summands in Eq. (ref) are, asymptotically, 1-dependent, which the subsampling approach shown in Figure (ref) does not adequately capture.
In order to consistently estimate $\Sigma$ in the bipower case, we use an intuitive blocking approach, which is described next. We define the $i$th block of high-frequency data by taking:
where $p \geq 2$ is an integer, and $i \geq 1$.
$B_{i}(p)$ is composed of the observation index associated with the sample of adjacent log-price observations $X_{(i-1)p/n}, \ldots, X_{i p /n}$. From this, we can compute $p$ consecutive returns $\Delta_{(i-1)p+1}^{n}X, \ldots, \Delta_{i p}^{n}X$. Therefore, $B_{i}(p)$ plays the role of the interval $[(i-1)/n,i/n]$ for power variation, which was used to compute a single return $\Delta_{i}^{n}X$. The only change is that we need to make this interval longer, such that we can consistently estimate the covariance structure of $V(f,g)^{n}$. As the $B_{i}(p)$'s are based on non-overlapping increments, it still holds that bipower variations computed from different subsamples are, asymptotically, conditionally independent.
We reset $\hat{ \Sigma}_{n}$ as follows:
where, assuming $L p$ divides $n$,
Note that $n/L p$ is the number of blocks assigned to each subsample, and that the subsample statistic $v_{i}(f,g)^{n}$ is computed only from data within the $i$th block $B_{i}(p)$. As in the above, we definitely require $n \to \infty$, $p \to \infty$, $L \to \infty$, and $n/pL \to \infty$ to prove the asymptotic theory for $\hat{ \Sigma}_{n}$. It turns out, however, we need a slightly stronger condition for the last part to ensure consistency. This is because the rate $\displaystyle \sqrt{ \frac{n}{L}}$ in the definition of Eq. (ref) corresponds to the martingale part of $V_{l} (f,g)^{n} - V(f,g)^{n}$, while the statistic $V_{l} (f,g)^{n} - V(f,g)^{n}$ also has a bias term, which is of order $Lp/n$. Thus, to make the bias negligible with respect to the martingale part, we need $n/Lp^2 \to \infty$. Hence, our “minimal” assumptions are based on this condition.
The first two errors in Eq. (ref) can be interpreted as to those in Theorem (ref), except the second is also affected by the block size $p$. Meanwhile, the decomposition of $\hat{ \Sigma}_{n} - \Sigma$ in Theorem (ref) has an extra error of order $O_{p}(1/p)$. The additional term, which emerges from the computation of the conditional variance of $v_{i}(f,g)^{n}$, has an intuitive interpretation, if we recall that in the current setting of bipower variation, the summands in Eq. (ref) (or Eq. (ref)) are asymptotically 1-dependent.
Consider the following stylized example. Assume that $(Z_{i})_{i \geq 1}$ is a sequence of stationary 1-dependent random variates. Then,
as $p \to \infty$.
This calculation shows that the finite sample variance on the left-hand side is not equal to, but converges towards, the asymptotic variance. The difference, i.e. the bias, is the term $-2 \text{cov} (Z_{1},Z_{2})/p$, which has order $O(1/p)$. This example also helps to illustrate that Theorem (ref) does not change, and in particular the convergence rate of $\hat{ \Sigma}_{n}$ is unaffected, if we were to compute a higher order multipower variation statistic. Then there would be more covariance terms in Eq. (ref), but the bias in each of them would still be $O(1/p)$.
The fastest rate is again found by balancing the errors, which means taking:
for which
While we do not offer a formal proof, we again believe that in a general diffusion model this rate is optimal within the framework of subsampling high-frequency data, as elaborated above. Moreover, the consistency only result of $\hat{ \Sigma}_{n}$ holds under weaker assumptions that do not require Assumption (K), (H) and (M), while an extra condition $L/p \to \infty$ is necessary to deal with an additional bias term.
To end this section, we should point out that for the power variation estimator covered by Theorem (ref) in the previous subsection, it follows from the work of BGJPS6 that there exits another consistent, positive semi-definite estimator of $\Sigma$:
$\hat{S}_{n}$ has a better rate of convergence $n^{-1/2}$ compared to $n^{-1/3}$ derived in the previous section for $\hat{\Sigma}_{n}$. $\hat{S}_{n}$ is therefore more efficient for power variation, but it does not work for bi- or multipower variation.
In an efficient market, equilibrium prices should adjust instantly to new information about fundamentals. If this leads to a significant revision of the fair value of the asset, the price has to move sharply and, potentially, discretely. This feature of price formation is not captured by the previous setup, where $X$ has continuous sample paths. In this section, we therefore add a jump term to $X$ and develop a framework for jump-robust inference about volatility based on subsampling truncated bipower variation \citep*[e.g.,][]{jacod-protter:12a,mancini:09a}. Accordingly, we assume that:\\[-0.50cm]
Assumption (J): $X$ is of the form:
where $X_{0}$, $a = (a_{t})_{t \geq 0}$, $\sigma = ( \sigma_{t})_{t \geq 0}$ and $W = (W_{t})_{t \geq 0}$ are defined as in Eq. (ref), while $(E, \mathcal{E})$ is a Polish space, $\mu$ is a random measure on $\mathbb{R}_{+} \times E$ with compensator $\nu( \text{d}s, \text{d}x) = \text{d}s F( \text{d}x)$, where $F$ is a $\sigma$-finite measure on $(E, \mathcal{E})$. Also, $\delta: \Omega \times \mathbb{R}_{+} \times E \to \mathbb{R}$ is a predictable function and $(S_{k})_{k \geq 1}$ is a sequence of stopping times increasing to $\infty$ such that $| \delta( \omega,s,x)| \wedge 1 \leq \psi_{k}(x)$ for all $( \omega, s, z)$ with $s \leq S_{k}( \omega)$ and $\int_{E} \psi^{ \beta}_{k}(x) F( \text{d}x) < \infty$ for all $k \geq 1$ and $\beta \in [0,1)$. \\[-0.50cm]
$\beta$ relates to the activity index of the price jump process. The condition imposed on $\beta$ implies that the jumps in $X$ are (absolutely) summable, i.e. we restrict attention to jump processes with paths of finite variation, but, possibly, infinite activity.\footnote{As explained earlier, for the subsampling estimator to be consistent for the asymptotic conditional variance-covariance matrix, we typically require a central limit theorem to hold for the underlying statistic of interest. In this respect, the restriction on $\beta$ is a standard condition in the high-frequency volatility literature.}
Although the theory derived here should work with a general $f$ and $g$, it requires a lot of notation. To develop ideas and maintain a streamlined exposition, we focus on the class of pure bipower variations in this section. The $k$th coordinate of the truncated bipower variation $\check{V}(q,r)^{n}$ is therefore:
where $\Delta_{i}^{n} \check{X} = \Delta_{i}^{n} X \cdot \mathbbm{1}_{ \{| \Delta_{i}^{n} X | \leq u_{n} \}}$ is the increment after jump-trucation and the threshold level $u_{n} = \alpha n^{- \check{ \omega}}$ with $\alpha >0$ and $\check{ \omega}\in (0,1/2)$. By excluding the largest increments of $X$, the bipower variation statistic is, asymptotically, merely based on those high-frequency returns that are compatible with a continuous sample path model.
First, we recall the central limit theorem for $\check{V}(q,r)^{n}$.
In the jump-diffusion setting, we define the subsample estimator of $\Sigma$ as:
where, assuming $L p $ divides $n$,
and $B_{i}(p)$ is given as in Eq. (ref).
Finally, we are ready to state a consistency result.
\@startsection {subsection}{1} {0mm} {-\baselineskip}{0.2\baselineskip} {\bf}{Microstructure noise} In practice, assets are not traded within a frictionless market. The recorded data constitute a discrete sample of transactions or bid-ask quotes, whose prices are affected by common sources of market imperfections, such as bid-ask spreads, price discreteness, and so forth \citep*[e.g.,][]{black:86a,niederhoffer-osborne:66a,roll:84a}. Even if these were small enough to be ignored, high-frequency data are also corrupted by outliers (due to bugs in the data transmission, fat-finger errors, etc.) and subject to other irregularities (e.g., quote stuffing, screen fighting, etc.). The combination of these effects leads to marked differences between real data and those generated by a diffusion model.
To accommodate this, we need to modify the setup. We are going to take the observed price as the true, underlying price perturbed by an additive noise term, i.e.
where $X$ is defined as in Eq. (ref), while $\epsilon = ( \epsilon_{t})_{t \geq 0}$ is a noise process. We impose the following: \\[-0.50cm]
Assumption (N): (i) $\epsilon$ is i.i.d. with $\mathbb{E}[ \epsilon_{t}] = 0$ and $\text{var}( \epsilon_{t}) = \omega^{2}$ for all $t \geq 0$, (ii) $\epsilon$ is independent of $X$, (iii) the distribution of $\epsilon$ is symmetric around 0, and (iv) $\mathbb{E} \big[ | \epsilon_{t}|^{s} \big] < \infty$ for some $s > 0$.
To alleviate the impact of noise, we make use of the notion that as $X$ is continuous and $\epsilon$ is i.i.d., we can locally smooth $Y_{i/n}$ in the vicinity of $i/n$ to retrieve an estimate, say $\bar{Y}_{i/n}$, which tends to be close to $X_{i/n}$, because the noise is largely averaged away \citep*[e.g.,][]{jacod-li-mykland-podolskij-vetter:09a, podolskij-vetter:09a,podolskij-vetter:09b}. Averaging our discrete sample of noisy high-frequency data this way leads to a new set of increments, $\Delta \bar{Y}_{i}^{n}$, based on pre-averaged prices.\footnote{In the context of volatility estimation, there are several tools at our disposal to handle microstructure noise, including the realized kernel; based on auto-covariance corrections \citep*[see, e.g.,][]{barndorff-nielsen-hansen-lunde-shephard:08a, barndorff-nielsen-hansen-lunde-shephard:11a}, and the two- or multi-scale realized variance; based on price subsampling \citep*[see, e.g.,][]{zhang:06a,zhang-mykland-ait-sahalia:05a}. Out of these, pre-averaging is the most general approach, as it is applicable to a large number of estimation problems.}
To implement pre-averaging, we need some extra notation. We choose a sequence $k_{n}$ of integers (the pre-averaging window) and a scalar $\theta > 0$, such that
We also need a weight function $w: \mathbb{R} \mapsto \mathbb{R}$ to do averaging. We assume $w$ is continuous on $[0,1]$ and piecewise continuously differentiable with a piecewise Lipschitz derivative $w'$. Moreover, we assume that $w(0) = w(1) = 0$ and $\int_{0}^{1} (w(t))^{2} \text{d}t > 0$. The following numbers and functions are associated with $w$:
where $w_{j}^{n} = w(j/k_{n})$. In passing, we note that
which means that in the asymptotic theory only $\psi_{1}$ and $\psi_{2}$ appear. Still, it is recommendable to use $\psi_{1}^{n}$ and $\psi_{2}^{n}$ for simulations and empirical work, as it entails better finite sample properties.
The return series, following pre-averaging, is:
The addition of microstructure noise creates further complications for inference procedures from high-frequency data. To make our framework analytically tractable, yet practically relevant, we therefore again restrict attention to the class of pure bipower variations.
The $k$th coordinate of $V^{*}(q,r)^{n}$ is defined as:
The intuition behind this construction is that pre-averaging induces some autocorrelation (of order $k_{n}$) in the pre-averaged price series, which is broken by multiplying pre-averaged returns that are $k_{n}$ terms apart. In essence, this leads to a lower, effective sample of size $n - 2k_{n} + 2$.
\citet*{podolskij-vetter:09a} show that
where
Thus, pre-averaging slows down the rate of convergence, but $n^{-1/4}$ is nonetheless the fastest rate in noisy diffusion models \citep*{gloter-jacod:01a,gloter-jacod:01b}.
In the above, $\Sigma^{*}$ is the $m \times m$ conditional covariance matrix of $V^{*}(q,r)^{n}$.\footnote{ We do not state the expression of $\Sigma^{*}$ here, but it can be found in \citet*{podolskij-vetter:09a}. In general, $\Sigma^{*}$ has a complicated structure (even with i.i.d., independent noise), and it is typically not known in closed-form. An exception, where $\Sigma^{*}$ can be computed analytically, is if $(q,r)$ consists of even non-negative integers, as in Theorem (ref). As an example, which is used in the simulations, take the pre-averaged realized variance. It sets $(q,r) = (2,0)$ and has $\Sigma^{*}(2,0) = 4 \int_{0}^{1} \Bigl( \theta^{3} \Phi_{22} \sigma_{s}^{4} + 2 \theta \Phi_{12} \sigma_{s}^{2} \omega^{2} + \frac{1}{\theta} \Phi_{11} \omega^{4} \Bigr)\text{d}s$.} \citet*{podolskij-vetter:09a} go on to develop a consistent estimator of $\Sigma^{*}$ (defined in Eq. (ref) in the simulation section), but it is based on element-by-element estimation. The disadvantage of this approach is that it does not ensure that the whole covariance matrix estimate is positive definite in finite samples and, as our simulations and empirical analysis show, a large fraction of such estimates in fact fail to be positive definite. Moreover, even if this property does hold, the estimate is often near-singular, resulting in an ill-conditioned and highly unstable covariance matrix.
To construct our estimator in the noisy setting, we follow the procedure from before by splitting the full sample of noisy high-frequency data into subsamples using a blocking approach.
We redefine:
where $p \geq 3$ is an integer and $i \geq 1$.
$B_{i}(p)$ is now the $i$th block of noisy high-frequency data. As readily seen, the only change compared to the noiseless setting is that $B_{i}(p)$ uses a larger block size. This implies we can do a sufficient amount of averaging within each block in order to diminish the noise, while still preserving enough of an effective sample size to estimate the correlation structure of $V^{*}(q,r)^{n}$.
Then, we set
where, assuming $L p k_n$ divides $n$,
As in the above, we should note that the summands $v_{i}(q_{k},r_{k})^{n}$ in the subsample estimates $V^{*}_{l}(q_{k},r_{k})^{n}$ exploit data solely from $B_{i}(p)$. Therefore, pre-averaging has to be done locally within the block, so that there is no overlap in the pre-averaged returns across the various blocks.
As in the previous subsection, the minimal assumptions we need to prove consistency are $n \to \infty$, $p \to \infty$, $L \to \infty$ and $\sqrt{n}/Lp^2 \to \infty$. The last condition again ensures that a bias term of the statistic $V^{*}_{l}(q,r)^{n} - V^{*}(q,r)^{n}$ is negligible with respect to its martingale part.
Now, we achieve the best rate
by choosing
Thus, the existence of microstructure frictions also adversely affects the speed of convergence of $\hat{ \Sigma}_{n}^{*}$.
Theorem (ref) is binding, because it restricts the choice of the powers $q_{k}$ and $r_{k}$ to even non-negative integers. But we can actually prove a weaker consistency result for any pre-averaged bipower variation, which is useful for practical work. First, we recall that allowing for general powers $q_{k}, r_{k} \geq 0$ by itself requires some stronger assumptions to prove the underlying central limit theorem in Eq. (ref). In particular, \citet*{podolskij-vetter:09a} impose that the noise distribution is (i) symmetric with (ii) $\mathbb{E}[|\epsilon|^{a}] < \infty$ for $a \in (-1,0)$ (their Assumption (A)) and that the noise distribution fulfills Cramer's condition, i.e. $\lim \sup_{|t| \to \infty} \chi(t) < 1$, where $\chi$ is the characteristic function of $\epsilon$ (their Assumption (A')). Of course, we also need this. Note, however, that we can again dispense with Assumption (M) for consistency.
The i.i.d. framework on the microstructure noise $\epsilon$ is a convenient outset, but it is hard to defend at the tick frequency, both in theory and practice \citep*[e.g.,][]{diebold-strasser:13a}. An intriguing ability of the subsampling estimator $\hat{ \Sigma}_{n}^{*}$ is that it tends to be robust against the intricate features of the noise process, as long as an associated central limit theorem holds. kalnina:11a studied subsampling in the presence of both autocorrelated and heteroscedastic noise for the two-scale realized variance. In this subsection, we show how our theoretical results adapt to such models, allowing for more general structure in the noise process.
1. Dependent noise
Autocorrelation in tick-by-tick returns can extend beyond the first lag, depending a bit on how you gather the data \citep*[e.g.,][]{hansen-lunde:06b,ait-sahalia-mykland-zhang:11b}. This cannot be captured by independent noise, so we start by weakening this assumption to so-called $m$-dependent noise. Thus, we now assume that the noise process $( \epsilon_t)_{t \geq 0}$ is stationary and that the random variables $\epsilon_{i/n}$ and $\epsilon_{j/n}$ are independent, only if $|i - j| > m$. \citet*{hautsch-podolskij:13a} prove a central limit theorem for pre-averaging in this setup (based on the estimator $V^{*}(2,0)^{n}$). As indicated by their results, the law of large numbers and the central limit theorem for the pre-averaged bipower variation estimator $V^{*}(q,r)^n$ do not change, except that the noise variance $\omega^{2}$ has to be replaced by the expression $\rho^{2} = \omega^{2} + 2 \sum_{j = 1}^{m} \text{cov}( \epsilon_{i/n}, \epsilon_{(i+j)/n})$, i.e.
Here, the form of $\Sigma^{*}$ changes.\footnote{In particular, \citet*{hautsch-podolskij:13a} show that the asymptotic variance of the pre-averaged realized variance $V^{*}(2,0)^{n}$ presented in Footnote (ref) is unchanged, apart from replacing $\omega$ with $\rho$ everywhere.} Our proposed estimator $\hat{ \Sigma}_{n}^{*}$ is still consistent though, as the change in the bias caused through replacing $\omega^2$ by $\rho^2$ is corrected by construction in Eq. (ref). This is because $\hat{ \Sigma}_{n}^{*}$ imitates the underlying covariation, irrespective of the true microstructure model. This implies that the asymptotic results of Theorem (ref) and (ref) are also true for the $m$-dependent noise model.
2. Heteroscedastic noise
The market microstructure reveals itself, for example, via the bid-ask spread. It has been noticed in many empirical asset price series that such margins are not constant through time, but tend to vary systematically within the day in the form of a U-shape. Thus, in the equity market spreads are typically larger in the morning and afternoon than during the middle of the day. To accommodate this, another setup that has been studied is heteroscedastic noise \citep*[e.g.,][]{bandi-russell:06a,kalnina-linton:08a}. Here, we follow the exposition in \citet*{jacod-li-mykland-podolskij-vetter:09a} by assuming that
while, conditional on $X$, $\epsilon_{t}$ and $\epsilon_{s}$ are independent for any $t \neq s$. By construction, this model exhibits a time-varying variance structure of the noise, which depends on the efficient price $X$.\footnote{An example of this model is additive, uniform noise plus rounding: $Y_{t} = \gamma \big[ (X_{t} + u_{t}) / \gamma \big]$, where $u = (u_{t})_{t \geq 0}$ is an i.i.d. $\mathcal{U}([0, \gamma])$-distributed process that is independent of $X$, and $\gamma > 0$ is a fixed rounding level. In this setting, the conditional variance of the noise process is given by: $\omega_{t}^{2} = \gamma^{2} \bigg( \Big\{ \frac{X_{t}}{ \gamma} \Big\} - \Big\{ \frac{X_{t}}{ \gamma} \Big\}^{2} \bigg)$, with $\{x\} = x - [x]$ denoting the fractional part of $x$.} Note that these conditions do not contradict unconditional dependence in $\epsilon$. In this case, the consistency result translates to:
Again, the estimator $\hat{ \Sigma}_{n}^{*}$ automatically adapts to the new environment, which, in particular, implies that the consistency result of Theorem (ref) is still true. In order to maintain unchanged error rates as in Theorem (ref), however, we need to also impose identical assumptions on the process $( \omega_{t})_{t \geq 0}$ as for the volatility $(\sigma_{t})_{t \geq 0}$. This is because the role of both processes are identical in all asymptotic expansions.
\@startsection {section}{1} {0mm} {-\baselineskip}{0.2\baselineskip} {\bf}{Simulations}
In this section, we conduct a small Monte Carlo study. It takes a closer look at the finite sample properties of covariance matrix estimation by subsampling. Throughout, we restrict attention to the noisy setting and estimation of $\Sigma^{*}$. We examine the ability of our proposed estimator $\hat{ \Sigma}_{n}^{*}$ to assist in drawing feasible inference about the pre-averaged bipower variation.
The efficient log-price $X$ is simulated as:
where $W_{t}$ and $B_{t}$ are independent standard Brownian motions, while $\kappa, \sigma^{2}, \xi$ and $\rho$ are parameters. The process adopted for $\sigma_{t}^{2}$ is a \citet*[][]{heston:93a} model, which is mean-reverting; features square-root volatility; and accommodates a leverage effect.\footnote{The leverage effect describes a negative correlation between an asset's return and volatility \citep*[e.g.,][]{black:76a,christie:82a}. Thus, if a leverage effect is present, one would expect $\rho$ to be negative.}
To get a version of the model from which we can actually simulate data, we apply a standard Euler approximation to the continuous time formulation in Eq. (ref). We then simulate 10,000 independent sample path realizations of the discretized system of bivariate equations.\footnote{To avoid a systematic effect from an assumed initial condition of volatility, $\sigma_{0}^{2}$, we restart the variance process in each simulation by drawing at random from its stationary distribution, $\sigma_{t}^{2} \sim \text{Gamma}(2 \kappa \sigma^{2} \xi^{-2}, 2 \kappa \xi^{-2})$.} We use two different sample sizes of $n$ = 2,340 and 23,400. In our empirical investigation, we look at high-frequency equity data from NYSE. With a US stock exchange trading session running from 9:30am to 4:00pm---or 6.5 hours---these sample sizes translate into receiving a new price update every ten and one second(s). Our sample sizes are therefore representative of more frequently traded securities.
We assume that the parameter values in the volatility equation are $\kappa = 5$, $\sigma^{2} = 0.04$, $\xi = 0.50$ and $\rho = -0.50$, which is broadly consistent with prior work \citep*[e.g.,][]{ait-sahalia-mykland-zhang:11b,kalnina:11a}. This implies that $\sigma_{t}$ is about 20% on an average, annualized basis, but the configuration of the model adopted here can generate a substantial degree of intraday variation in volatility via $\xi$. An example simulation is provided in Figure (ref).
An autocorrelated and heteroscedastic noise term is added to $X$. First, we create a set of serially dependent standard normal random variables via an MA(1) filter: $u_{i/n} = u_{i/n}' + \zeta u_{(i-1)/n}'$, where $\zeta$ is a parameter and $\displaystyle u_{i/n}' \overset{ \text{i.i.d.}} \sim \text{N} \bigg(0, \frac{1}{1 + \zeta^{2}} \bigg)$, independent of $X$. Hence, $u_{i/n} \sim \text{N} \big(0, 1 \big)$ and $\text{cov}(u_{i/n}, u_{(i-1)/n}) = \displaystyle \frac{ \zeta}{1+\zeta^{2}}$, so that $\zeta$ controls the degree of first-order serial correlation in the noise. In our simulations, we take $\zeta = -0.4$. Second, as in \citet*{ait-sahalia-jacod-li:12a}, we set $\displaystyle \epsilon_{i/n} = \gamma \frac{\sigma_{i/n}}{\sqrt{n}} u_{i/n}$, where $\gamma$ is the noise-ratio parameter \citep*[e.g.,][]{oomen:06a}. This formulation implies that, conditional on $\sigma$, $\displaystyle \omega_{i/n} = \gamma \frac{ \sigma_{i/n}}{ \sqrt{n}}$ and ensures microstructure noise variation is conditionally heteroscedastic and proportional to the spot volatility of the efficient price. We assume that $\gamma = 0.50$, which is a realistic choice for more liquid assets, see, e.g., \citet*[][]{christensen-oomen-podolskij:14a}.\footnote{We also experimented with a much larger noise-ratio of $\gamma = 2$, as done by \citet*{ait-sahalia-jacod-li:12a}. The results were almost identical, albeit slightly worse, than those we report in the main text and, hence, are omitted to conserve space.} Although this noise setting is not formally covered by our theoretical frame, we include it here as a robustness check. To alleviate the impact of noise, we pre-average using the bandwidth $k_{n} = [ \theta \sqrt{n} ]$, and we experiment with two choices of the tuning parameter $\theta = 1/3$ and $1$. We set the weight function $w(x) = \min(x, 1-x)$, which has been shown to deliver nearly efficient estimates of the integrated variance, when further parametric assumptions are imposed \citep*[see, e.g.,][]{podolskij-vetter:09a}.
\@startsection {subsection}{1} {0mm} {-\baselineskip}{0.2\baselineskip} {\bf}{A preliminary analysis}
We begin by verifying the conjecture made in the introduction of this paper, namely that the finite sample properties of existing estimators of $\Sigma^{*}$ are poor, which renders a significant fraction of such estimates either nonpositive definite or at least ill-conditioned. To do this, we implement the estimator suggested by \citet*{podolskij-vetter:09a}. It is constructed by first defining
and setting
for any $0 \leq m \leq n - 4k_{n} + 1$ and $0 \leq l \leq 2k_{n}$. Then,
and it follows that $\tilde{\Sigma}_{n}^{*} = \bigl( \tilde{ \Sigma}_{ij,n}^{*} \bigr) \overset{p}{\to} \Sigma^{*}$.
{0.30cm}
\afterpage
Table (ref) shows several diagnostics that highlight the properties of $\tilde{ \Sigma}_{n}^{*}$ based on $q = (2,1)'$ and $r = (0,1)'$, i.e. pre-averaged realized variance and $(1,1)$-bipower variation. In the table, we report the outcome from the general noise model, but we also include a comparable i.i.d. noise environment, which we base on setting $\zeta = 0$ in the above and replacing $\sigma_{i/n}$ by $\sqrt{ \int_{0}^{1} \sigma_{s}^{2} \text{d}s}$ in the noise variance, while noting that \citet*{podolskij-vetter:09a} operate under the latter conditions. In addition, the results are also obtained for a scaled Brownian motion (BM), in which volatility has been fixed at its steady-state value of $\sigma^{2}$. The column SV is for the Heston stochastic volatility model, while SPY represents some real high-frequency data that are further commented on in Section (ref).
In Panel A, we report the fraction of the computed $\tilde{ \Sigma}_{n}^{*}$, which fail to be positive definite, i.e. which have a minimum eigenvalue $\min( \lambda_{i}) \leq 0$. It suggests that for a small sample of $n = 2,340$ and depending on $\theta$, between 18% -- 35% of $\tilde{ \Sigma}_{n}^{*}$ are nonpositive definite. As expected, these numbers decrease as the sample size increases, but even with a fairly large sample of $n = 23,400$ the failure rate is far from negligible. Moreover, it increases if a longer pre-averaging horizon is employed. As such, this issue therefore has substantial bite in practice, because larger values of $\theta$ are typically preferred, when the noise is suspected to violate the i.i.d. assumption \citep*[e.g.,][]{christensen-oomen-podolskij:14a,hautsch-podolskij:13a}. Note that going from constant to stochastic volatility changes the numbers only slightly, so allowing volatility to be time-varying has no discernable impact on the failure rate.
Turning next to Panel B, we investigate how often the linear combination $\omega' V^{*}(q,r)^{n}$ with $\omega = (1, -\mu_{1}^{-2})'$ results in a negative variance estimate $\omega' \tilde{ \Sigma}_{n}^{*} \omega \leq 0$. The difference $V^{*}(2,0)^{n} - \mu_{1}^{-2}V^{*}(1,1)^{n}$ is often used in applied work, as it provides information about presence of jumps in the price process and permits a statistical test of this hypothesis. Even when the covariance matrix estimate is not positive definite, it could still result in a positive variance estimate $\omega' \tilde{ \Sigma}_{n}^{*} \omega > 0$, thereby allowing the t-statistic to be computed. While the numbers in Panel B are lower compared to Panel A, they are still high.
In Panel C, we look at those $\tilde{ \Sigma}_{n}^{*}$ estimates that are positive definite. We compute the percentage of these, which return a condition number $\text{cond}( \tilde{ \Sigma}_{n}^{*}) \geq 20$. \footnote{The condition number of an invertible matrix $\text{A}$ is defined as $\text{cond}(\text{A}) = ||\text{A}|| \cdot ||\text{A}^{-1}||$, where $|| \cdot ||$ is the L$_{2}$ matrix norm. $\text{cond}(\text{A})$ can be shown to be the ratio between the largest and smallest singular value of $\text{A}$. It can be interpreted as saying how much small changes in the input matrix get amplified under matrix inversion.} The condition number measures the numerical accuracy of a matrix, and a value above $20 = 10 \times \text{dim}( \tilde{ \Sigma}_{n}^{*})$ is generally taken as a sign of an ill-conditioned and nearly singular matrix \citep*[e.g.,][]{greene:11a,hautsch-kyj-oomen:12a}. As readily seen, we find that about 5% -- 10% of the $\tilde{ \Sigma}_{n}^{*}$ that are deemed ok by a definiteness criteria show signs of being badly scaled.
Lastly, in Panel D we attempt to estimate the joint asymptotic covariance matrix of a 4-dimensional parameter by using $q = (2,1,4,2)'$ and $r = (0,1,0,2)'$. We then report the percentage of the $\tilde{ \Sigma}_{n}^{*}$ estimates, which are not positive definite, i.e. the numbers can be compared to Panel A. Not surprisingly, increasing the complexity of the problem has a detrimental impact on the estimation errors, and up to 80% of the estimates are now nonpositive definite, making this a devastating issue for inference \citep*[e.g.,][employ a 3-dimensional statistic to test for the parametric form of volatility in diffusion models (both with or without noise)]{ dette-podolskij-vetter:06a,dette-podolskij:08a,vetter-dette:12a}.
\@startsection {subsection}{1} {0mm} {-\baselineskip}{0.2\baselineskip} {\bf}{Implementation of the subsampler}
We now turn to the subsampler, where we again base our investigation on $V^{*}(q,r)^{n}$ using the parameters $q = (2,1)'$ and $r = (0,1)'$. As $\hat{ \Sigma}_{n}^{*}$ depends on two tuning parameters, $p$ and $L$, we compute it by varying these across a broad range of values in order to gauge the sensitivity of our estimator to specific choices. We set $p = 3, 5$ and $10$, so that the block length of noisy returns before pre-averaging goes from three to ten times the pre-averaging horizon $k_{n}$. Moreover, we slice the sample into $L = 5, 10$ and $15$ subsamples, yielding a total of nine combinations of $p$ and $L$.\footnote{This implies that for the sample size $n = 2,340$, there are some combinations of $p$ and $L$, for which there is not enough data to compute the subsampler. Therefore, we restrict attention to $n = 23,400$ in the following. The results for $n = 2,340$, when attainable, are not materially worse, reflecting the slow rates of convergence, and all are available by request.}
Our initial set of simulations suggested that, for small $p$ and $L$, the raw estimator defined by Eq. (ref) is downward biased, thereby leading to a systematic underestimation of $\Sigma^{*}$. We therefore start by briefly outlining a few corrections that are important in finite samples.
First, in order to center the $V_{l}^{*}(q,r)^{n}$, we should in theory use the unobserved $V^{*}(q,r)$, which we are forced to replace by a feasible, consistent estimator, i.e. $V^{*}(q,r)^{n}$. While this has no impact asymptotically, because $V^{*}(q,r)^{n}$ converges much faster than $V_{l}^{*}(q,r)^{n}$, a closer inspection of $\hat{ \Sigma}_{n}^{*}$ shows that the substitution does entail a standard small sample correction. This implies that $\hat{ \Sigma}_{n}^{*}$ should be divided by $1 - 1/L$, i.e. the “right” normalization in Eq. (ref) is $L - 1$ and not $L$.
Second, there is a HAC error associated with $p$, which---in contrast to the $L$ correction---is more subtle to deal with. The problem originates from the estimation of the autocovariances of the pre-averaged returns, $|n^{1/4} \Delta \bar{Y}_{i}^{n}|^{q_{k}} | n^{1/4} \Delta \bar{Y}_{i+k_{n}}^{n}|^{r_{k}}$, as exemplified by Eq. (ref). As such, it depends on the covariance structure of this series, which, in turn, is a function of several parameters and variables, including spot volatility, the weight function and the variance of the noise process \citep*[see, e.g., the proof of Theorem (ref) in the Appendix around Eq. (ref), or][]{podolskij-vetter:09a}. If we approximate this function, e.g., by assuming that volatility is constant, a detailed calculation (which is omitted here, but available upon request) shows that for the pre-averaged bipower variation estimator $V^{*}(q_{k}, r_{k})^{n}$ defined by Eq. (ref), we can roughly correct for the $p$ error by dividing $\hat{ \Sigma}_{n}^{*}$ with $1 - 1/p$.
It turns out, however, that this is too much, if either $q_{k}$ or $r_{k}$ is zero, as it happens for the pre-averaged realized variance, $V^{*}(2,0)^{n}$. This is because the summands in Eq. (ref) are, asymptotically, $2k_{n}$-dependent for non-zero values of both $q_{k}$ and $r_{k}$, while they are only $k_{n}$-dependent, if either is zero. Thus, the bias induced by $p$ in the latter is, loosely speaking, half that of the former. This indicates that we ought to divide the elements of $\hat{ \Sigma}_{n}^{*}$ involving the variance of $V^{*}(2,0)^{n}$ and its covariance with $V^{*}(1,1)^{n}$ only by $1 - 0.5/p$. This is less appealing, because it would break the positive semi-definiteness property of $\hat{ \Sigma}_{n}^{*}$. We therefore proceed by using a constant scaling for the entire matrix, and---to strike a balance between the two alternatives---we propose to meet in the middle and rescale $\hat{ \Sigma}_{n}^{*}$ by $(1 - 0.75/p)$. This choice produces excellent results for the values of $p$ considered in this paper, as corroborated by our numerical experiments below, while for $p \geq 10$ the correction is of limited importance and can be ignored.
Third, our theoretical results hinge on $n$ being a multiple of $Lpk_{n}$. In practice, where $n$ varies randomly over time and can be very odd, this is an unrealistic assumption, which is almost never satisfied. Instead, for a given choice of $k_{n}$, $p$ and $L$, the maximum number of blocks of length $pk_{n}$ that can be assigned to each of the $L$ subsample estimates $V_{l}^{*}(q_{k},r_{k})^{n}$ is:
The effective amount of data used to construct the subsampler is therefore often less than the total sample size, i.e. $n_{ \text{block}}Lpk_{n} \leq n$.
We compute $\hat{ \Sigma}_{n}^{*}$ from the data that fall within the window $[0, n_{ \text{block}}Lpk_{n}/n]$ and subsequently inflate this estimate to cover the whole unit interval. While this entails some loss of information about the underlying variation of the process towards the end of the sample, in our experience this has a very limited influence on the results, unless the data, from which $\hat{ \Sigma}_{n}^{*}$ is computed, is not representative of the overall level of volatility. In practice, one can minimize this effect by choosing the parameters, such that $n_{ \text{block}}Lpk_{n}$ is close to $n$.
\@startsection {subsection}{1} {0mm} {-\baselineskip}{0.2\baselineskip} {\bf}{Results}
In Figure (ref), we plot some kernel smoothed density estimates of the standardized pre-averaged bipower variation, i.e. $n^{1/4} \bigl( V^{*}(q_{k},r_{k})^{n} - V^{*}(q_{k},r_{k}) \bigr) / \sqrt{ \hat{ \Sigma}_{kk,n}^{*}}$, where $\hat{ \Sigma}_{kk,n}^{*}$ is the $k$th diagonal element of our subsampling covariance matrix estimate $\hat{ \Sigma}_{n}^{*}$. Here, we use $\theta = 1$. The results in Panels A -- B are for $V^{*}(1,1)^{n}$, while Panels C -- D are for $V^{*}(2,0)^{n}$. In addition, the left-hand portion of the figure is for $L = 15$ and $p$ changing, while the right-hand part is based on $p = 10$ and for different $L$. The infeasible result for $V^{*}(2,0)^{n}$ replaces the subsampler with the true variance, which is known here (cf. footnote (ref)).
As the figure shows, the studentized pre-averaging estimators tend to track the asymptotic normal approximation closely across combinations of $p$ and $L$. The sole exception, appearing in Panel A, is for $V^{*}(1,1)^{n}$, when $\hat{ \Sigma}_{n}^{*}$ is implemented using $p = 3$. Note that if $p = 3$, the effective sample size within a block of noisy high-frequency data is $k_{n}+2$ after pre-averaging, whereas the summands of $V^{*}(1,1)^{n}$ are $2k_{n}$-dependent. With such a small value of $p$, the block length is therefore inadequate to permit estimation of all the required autocovariances. As $|n^{1/4} \Delta \bar{Y}_{i}^{n}|^{q_{k}} | n^{1/4} \Delta \bar{Y}_{i+k_{n}}^{n}|^{r_{k}}$ is strongly positively autocorrelated, this leads to a severe underestimation of the true variation of $V^{*}(1,1)^{n}$ and, hence, a pronounced overdispersion of the estimated density. As a practical guide, one should therefore avoid computing $\hat{ \Sigma}_{n}^{*}$ with $p = 3$, if both $q_{k}$ and $r_{k}$ are different from zero, for any $i = 1, \dots, m$. In comparison, the corresponding graph for $V^{*}(2,0)^{n}$ in Panel C is much better scaled, reflecting the lesser dependence inherent in this estimator. Apart from that, the fit tends to improve for larger values of $p$ and $L$, as expected.\footnote{Of course, with a fixed sample size $n$ and pre-averaging window $k_{n}$, the parameters $p, L$, and $n_{ \text{block}}$ are not free. Thus, everything else is not “fixed”, because $n_{ \text{block}}$ is decreasing for larger values of either $p$ or $L$ (while holding the other fixed).} Lastly, comparing with the infeasible result in Panels C -- D, we note that the estimated densities appear slightly negatively skewed, owing to a modest, positive correlation between $V^{*}(q_{k},r_{k})^{n}$ and $\hat{ \Sigma}_{kk,n}^{*}$.
We turn next to Figure (ref), where we explore how sensitive our findings are to the choice of $\theta$. In this figure, and throughout the remainder of this section, we fix the parameters of $\hat{ \Sigma}_{n}^{*}$ to $p = 10$ and $L = 15$. As apparent from both panels, the relatively large change in $\theta$ has only a minuscule effect on the shape of the estimated density for $V^{*}(1,1)^{n}$ and $V^{*}(2,0)^{n}$.
In Figure (ref), using $\theta = 1$, we look at an application that requires one to use information about the full covariance matrix estimate by reporting some results for the linear combination $\omega' V^{*}(q,r)^{n}$ with $\omega = (1, -\mu_{1}^{-2})'$, i.e. $V^{*}(2,0)^{n} - \mu_{1}^{-2} V^{*}(1,1)^{n}$. In Table (ref), we noted it was problematic to standardize this difference with the covariance matrix estimator put forth by \citet*{podolskij-vetter:09a}, which was often found to be nonpositive definite. The subsampler does not suffer from this issue. In Panel A of Figure (ref), we therefore plot the time series of the studentized statistic across the simulations runs, using the delta method to conclude that $n^{1/4} \omega' \bigl( V^{*}(q,r)^{n} - V^{*}(q,r) \bigr) / \sqrt{ \omega' \hat{ \Sigma}_{n}^{*} \omega} \overset{d}{ \to} \text{N}(0,1)$. Panel B inspects the kernel smoothed density estimate of the t-statistic. As evident, the asymptotic distribution theory is a decent description of the actual finite sample variation, although the fit is not perfect. As explained above, $V^{*}(2,0)^{n} - \mu_{1}^{-2} V^{*}(1,1)^{n}$ provides information about the presence of jumps in asset prices, and significant positive values would lend support to this hypothesis. Here, the shape of the estimated density implies that the one-sided coverage probabilities of the t-statistic are slightly too large in the right tail. This would render such a hypothesis test mildly conservative, which is preferable in practice (e.g., using the 95% quantile from the standard normal distribution gives a coverage rate of 96.7% in the above figure). We note that it requires additional properties of $\hat{ \Sigma}_{n}^{*}$ for such a test to retain any power under the alternative. That is, $\hat{ \Sigma}_{n}^{*}$ is not jump-robust in its current form, so it needs to be accompanied by truncation, as shown in the noiseless setting in Theorem (ref). We shall explore the joint impact of noise and price jumps on the subsampler in a companion paper, and the results could of course change with a different, albeit related, estimator of $\Sigma^{*}$.
At last, we compare our subsampler $\hat{ \Sigma}_{n}^{*}$ to an alternative, nonparametric estimator of $\Sigma^{*}$, namely the observed asymptotic variance (AVAR) of \citet*{mykland-zhang:17a}. As in our setting, the observed AVAR is based on squared increments (or outer products) of the original statistic(s) computed on smaller streches of high-frequency data, but there are several differences between the construction of the subsampler and observed AVAR. Moreover, there is little guidance on how to select tuning parameters for the latter. We therefore proceed as follows. The sampling grid consists (using their notation) of $B = L$ blocks at the outset. This helps to ensure comparability with $\hat{ \Sigma}_{n}^{*}$. We then compute the observed AVAR using a two-scale approach, as a linear combination of the $K$-averaged apparent quadratic covariation with $K_{1} = 1$ and $K_{2} = 2$; see Eq. (24) in \citet*{mykland-zhang:17a}.\footnote{The observed AVAR has a bias, which---although asymptotically negligible---could impair its accuracy in finite samples. The virtue of the two-scale construction, as advocated by \citet*{mykland-zhang:17a}, is that the bias term cancels out.} A forward half-interval approach is adopted to reduce the impact of edge effects induced by pre-averaging. The outcome is reported in Figure (ref), where we plot the standardized pre-averaged bipower variation $V^{*}(2,0)^{n}$ and $V^{*}(1,1)^{n}$ using both $\hat{ \Sigma}_{n}^{*}$ and the observed AVAR. As apparent, standardization with the subsampler tracks the standard normal curve closer compared to the observed AVAR. Indeed, the standard error of the studentized pre-averaged bipower variation is about 1.05 using $\hat{ \Sigma}_{n}^{*}$, while it is about 1.20 using the observed AVAR.
Overall, the simulation results suggest that inference based on $\hat{ \Sigma}_{n}^{*}$ is fairly robust and delivers excellent outcomes, even for modest values of its tuning parameters.
\@startsection {section}{1} {0mm} {-\baselineskip}{0.2\baselineskip} {\bf}{Empirical work}
Here, we provide a brief illustration of the subsample estimator in the context of some real financial high-frequency data. We analyze tick-data from the Standard and Poor's depository receipts, which is an exchange-traded fund that tracks the performance of the S&P 500 stock index. The shares are listed on several U.S. stock exchanges and trade under the ticker symbol SPY. It is a highly liquid security and provides a good starting point for the subsampler, which is data-intensive. We extracted a transaction price series of the SPY from the TAQ database. The data are recorded at milli-second precision and our complete sample covers the time period from January, 2007 to March, 2011; a total of 1,169 business days. The raw data was filtered for outliers using the recommendations of \citet*{christensen-oomen-podolskij:14a}.\footnote{The \citet*{christensen-oomen-podolskij:14a} filter is to a large extent based on the cleaning routines of \citet*{barndorff-nielsen-hansen-lunde-shephard:09a}. The former use a backward-forward matching algorithm to compare a trade to the quote conditions prevailing in the market around the time of the transaction. The latter evaluate each trade against a single preceding bid-ask quote, which may lead to excessive removal of data in fast-moving markets. Apart from that, the filters are identical.} We also restrict attention to the NYSE trading session, which runs from 9:30am till 4:00pm Eastern Time. Table (ref) provides a few descriptive statistics that summarize key features of the dataset.
{0.45cm}
In Panel A of Figure (ref), we plot the sample autocorrelation function (acf) of the noisy return series, $\Delta_{i}^{n} Y$, up to lag 15. There is a pronounced, significantly negative first-order autocorrelation of about -0.35, which is consistent with a bid-ask bounce interpretation of microstructure noise. The acf then increases and turns positive at lag three. The fourth and fifth autocorrelation actually fall outside the 95% confidence bands based on a white noise null hypothesis. Together with the subsequent monotonic decay of the acf, this indicates that noise operating at the tick-level is not i.i.d., as consistent with prior work \citep*[e.g.,][]{hansen-lunde:06b}. We therefore proceed using a pre-averaging window based on $\theta = 1$, which should be a robust choice in light of the empirical evidence. The acf of the corresponding pre-averaged returns, $\Delta \bar{Y}_{i}^{n}$, is presented in Panel B of the figure. As expected, there is a strong dependence in this series up to lag $k_{n}$.
We report the resulting time series of $V^{*}(2,0)^{n}$ and $\mu_{1}^{-2} V^{*}(1,1)^{n}$ in Panel A of Figure (ref). Note that the graph for $\mu_{1}^{-2} V^{*}(1,1)^{n}$ has been reflected in the $x$-axis. The statistics are first computed day-by-day across the whole sample and subsequently updated using Eq. (ref) to provide annualized measures of the integrated variance (assuming 250 trading days p.a., on average), i.e.:
with an identical transformation of $\mu_{1}^{-2} V^{*}(1,1)^{n}$.
The term $\displaystyle \frac{\psi_{1}^{k_{n}} \hat{ \omega}^{2}}{\theta \psi_{2}^{k_{n}}}$ is a small bias correction that compensates the pre-averaged bipower variation for the residual effect of microstructure noise.\footnote{The bias correction in Eq. (ref) is only correct, when the noise is i.i.d. Meanwhile, the estimator of $\omega^{2}$ we propose in Eq. (ref) is robust to the presence of a heteroscedastic noise process, but it is generally not consistent for $\rho^{2}$ from Section (ref), if the noise is autocorrelated. As the current application is merely illustrative, we ignore that issue here.} $\hat{ \omega}^{2}$ is an estimate of the noise variance, $\omega^{2}$. There are several estimators, which can serve the role of $\hat{ \omega}^{2}$ \citep*[see, e.g.,][]{gatheral-oomen:10a}. Among these, we adopt the one from \citet*{oomen:06a}, which relies on the first-order autocorrelation of the noisy returns:
We find a high degree of time-variation and persistence in the $\widehat{\text{IV}}$ series. The onset of the financial crisis and---in particular---the unprecedented volatility surrounding the collapse of Lehman Brothers in 2008 stands out visibly. To attach a measure of uncertainty to these estimates, Panel B charts the associated standard error estimate, $\sqrt{ \hat{ \Sigma}_{11,n}^{*}} / \theta \psi_{2}^{k_{n}}$ and $\sqrt{ \hat{ \Sigma}_{22,n}^{*}} / \theta \psi_{2}^{k_{n}}$, where the latter are based on the subsampler with $L = 15$ and $p = 10$. As expected, high levels of volatility spill over into the standard errors and tend to decrease estimation accuracy. The apparent outliers showing up in the standard error series in Q3, 2007 and Q2, 2010 correspond to single days with unusual market activity. The first is September 18, 2007, where the Federal Open Market Committee (FOMC) announced an unexpected reduction of its target for the federal funds rate by 50 basis points, while the second is May 6, 2010; the day of the S&P 500 Flash Crash.
Turn next to Figure (ref), where we conduct inference about $V^{*}(2,0)^{n} - \mu_{1}^{-2} V^{*}(1,1)^{n}$. In Panel A, we compute the difference in the logarithms of these numbers, i.e. $\ln \bigl( V^{*}(2,0)^{n} \bigr) - \ln \bigl( \mu_{1}^{-2} V^{*}(1,1)^{n} \bigr)$, which tends to be less volatile compared to the raw statistic. As shown, the majority of the point estimates hover around zero, which is the theoretical limit in diffusion models. There are some notable exceptions though, and in Panel B we examine one of these by zooming in on the month of December, 2007. Alongside the statistic, we here report a two-sided 95% confidence interval. Standard errors were found by applying the delta method (for the function $f(x,y) = \ln(y) - \ln(\mu_{1}^{-2}x)$) to the joint asymptotic distribution in Eq. (ref) and then replacing the asymptotic variance of the difference by a feasible estimate. In particular, we compare a set of intervals based on the subsampler, $\hat{ \Sigma}_{n}^{*}$, with those computed from the observed AVAR of \citet*{mykland-zhang:17a}, which is again computed as explained in the simulation section, and to the \citet*{podolskij-vetter:09a} estimator, $\tilde{ \Sigma}_{n}^{*}$. If the latter leads to a negative variance estimate, it is excluded. As consistent with Table (ref), this is a recurrent problem. Moreover, if all three estimates are well-defined, they are often closely aligned, but both the subsampler and observed AVAR appear less erratic, while $\tilde{ \Sigma}_{n}^{*}$ is often very narrow or wide. This is most visible from the big discrepancy on December 11, 2007, marking a day with yet another rate cut by the Fed. On this day, the condition number of $\tilde{ \Sigma}_{n}^{*}$ is $\text{cond}( \tilde{ \Sigma}_{n}^{*}) = 452.36$, which suggest that the underlying covariance matrix estimate is very fragile. The corresponding figure for the observed asymptotic variance is $64.68$, which is again rather high, and indeed it also leads to a very large confidence interval here. Meanwhile, the condition number of the subsampler is more modest at $\text{cond}( \tilde{ \Sigma}_{n}^{*}) = 15.16$, and it generally appears to be the most stable over time.
To end the paper, we provide an alternative application, where the subsampler is used to draw inference about the amount of heteroscedasticity in noisy high-frequency data. To do this, we start by computing the statistics $V^{*}(2,0)^{n}$ and $V^{*}(4,0)^{n}$, i.e. the pre-averaged bipower variation based on the parameter $q = (4,2)'$ (and $r = (0,0)'$). Taking these as input, we appeal to Eq. (ref) by forming the estimate:
which converges to the so-called integrated quarticity. We then exploit that $\sqrt{ \widehat{\text{IQ}}} / \widehat{\text{IV}} \overset{p}{ \to} \sqrt{ \int_{0}^{1} \sigma_{s}^{4} \text{d}s} / \int_{0}^{1} \sigma_{s}^{2} \text{d}s \geq 1$, with equality if and only if $\sigma$ is constant. Thus, an estimated ratio far above one suggests there is significant variation in volatility within the day, while a ratio close to one means $\sigma$ can be regarded, as if it was approximately constant. This type of statistic has been exploited in earlier work to test for the parametric form of volatility \citep*[e.g.,][]{dette-podolskij-vetter:06a,vetter-dette:12a}, and it is also finds use in the jump-testing literature \citep*[e.g.,][]{barndorff-nielsen-shephard:06a,kolokolov-reno:16a}.
The outcome of this exercise is collected in Figure (ref). In Panel A, we plot the time series of the estimated log-ratio, i.e. $\ln \Bigl( \sqrt{ \widehat{\text{IQ}}} / \widehat{ \text{IV}} \Bigr)$. We again use a log-transformation in order to improve the scaling of the results and facilitate interpretation of the graphs. The log-ratio should cluster around zero, if volatility is constant. We observe an extreme degree of fluctuation in this statistic over time, and, as anticipated, there are many days, where the log-ratio is large. Still, we also find a decent portion of estimates, which are close to zero. Small negative numbers can be observed as a result of sampling variation. In Panel B, we complement the analysis by looking at the month of December, 2007. We add a left-sided 95% confidence interval for $\ln \Bigl( \sqrt{ \widehat{\text{IQ}}} / \widehat{ \text{IV}} \Bigr)$, where standard errors are again retrieved via the delta method and three estimates of the asymptotic covariance matrix. The interpretation is that on some days, such as the day of the FOMC meeting, volatility is changing a lot, while on others it is not moving much, which is consistent with the findings of \citet*[][Figure 1]{mykland-zhang:17a}. Of course, the latter finding can also arise, if the high-frequency data is not informative enough to discriminate random sampling errors from genuine parameter variation in $\sigma$, which could be difficult in times of severe stress in financial markets. In this respect, it is important to acknowledge the limitations of the subsampler, which, albeit consistent, is itself subject to a substantial degree of sampling uncertainty in practice.
\@startsection {section}{1} {0mm} {-\baselineskip}{0.2\baselineskip} {\bf}{Conclusion}
In this paper, we propose a subsample estimator of the asymptotic conditional covariance matrix of bipower variation. The theory is developed for diffusion models both with and without microstructure noise. We show our estimator is consistent and, under suitable conditions, we find an error decomposition of our statistic, from which we are able to derive its rate of convergence. To complement the theory, we conduct a Monte Carlo study, which documents how the subsampler can be used to draw feasible inference about pre-averaged bipower variation in the presence of noise. The results are compelling and show that the subsampler delivers accurate inference and that it is robust to the choice of its tuning parameters, i.e. the number of subsamples and the block length. We concluded with an empirical analysis that provides an illustration of a few of the directions, in which the subsampler can be applied to real high-frequency data.
In future research, several extensions of the current paper are possible. First, as we noted here, we can use the subsampler as an ingredient to test for the presence of jumps in noisy high-frequency data. This requires the subsampler to be robust against such jumps, while also being resistant to the influence of microstructure noise, which the current implementation is not. This is a line of research that we are currently looking into ourselves in a companion paper. Second, there is some work to be done on the optimal selection of tuning parameters. Finally, as we stressed in the main text, pre-averaging can not only be used to consistently estimate integrated variance and quarticity, but also more general functionals of volatility. Meanwhile, subsampling facilitates estimation of the asymptotic covariance matrix of such pre-averaged high-frequency statistics under mild conditions on the noise process. The combination is therefore potent, and we envision that a lot of empirical papers can benefit from and follow in the aftermath of our work.