Higher-order small time asymptotic expansion of Itô semimartingale characteristic function with application to estimation of leverage from options (Q2239273): Difference between revisions

This is a highly technical paper aimed to study the asymptotic behavior over small time scales of a typical Itô semimartingale \[ x_t= x_0+\int_0^t\alpha_sds+\int_0^t\sigma_sdW_s+\sum_{s\le t}\Delta x_s \] widely used in finance for modeling financial asset prices \(x_t\) (semimartingale model). Since ``it is well known that over short time intervals, the diffusion component of \(x\) determines its behavior'', in several previous papers the author and other researchers have studied the asymptotic expansion of the (scaled) characteristic function \(\mathbb E_0^{\mathbb Q}(e^{iux_T/\sqrt T})\) of \(x_T\) for \(T\downarrow 0\) (see equation (1.4)) and they have found that the leading term is Gaussian (with the initial value \(\sigma_0\) of the diffusion coefficient \(\sigma_t\)), while the residual term ``due to the jumps in \(x\) and the variation in the characteristics of \(x\) over the time interval'' is of the order of \(\sqrt T\). ``For many applications of interest, however, the variation in \(\sigma\) can be rather nontrivial and therefore the expansion in (1.4), which ignores this variation, can provide an inaccurate approximation to \(\mathbb E_0^{\mathbb Q}(e^{iux_T/\sqrt T})\) when applied with small but finite \(T\). The goal of this paper, therefore, is to generalize the result in (1.4) by expanding \(\mathbb E_0^{\mathbb Q}(e^{iux_T/\sqrt T})\), for \(T \downarrow 0\), by explicitly accounting for the variation in \(\sigma\), in particular ... More specifically [the author derives] a higher-order expansion result with a residual term of asymptotic order of only \(O_p(T^{3/2-r/2}vT^{3/2}| \log(T )|)\), where \(r \in [0, 1]\) captures the so-called degree of activity of the jumps in \(x\) ... (in particular, \(r = 0\) corresponds to jumps of finite activity often used in applications)'' The diffusion coefficient \(\sigma_t\) of \(x\) being moreover another Itô semimartingale \[ \sigma_t= \sigma_0+\int_0^tb_sds+\int_0^t\eta_sdW_s+\int_0^t\widetilde\eta_sd\widetilde W_s+\sum_{s\le t}\Delta \sigma_s \] with \(W\) and \(\widetilde W\) independent Brownian motions, the previous asymptotic expansions are then applied ``to propose a (nearly) rate-efficient estimator of the coefficient \(\eta\), entering the dynamics of \(\sigma\) ... This coefficient, together with \(\sigma\), determines the continuous part of the quadraticccovariation between \(x\) and \(\sigma\), which is referred to as leverage effect''. This method sensibly improves the estimates proposed by other authors in previous papers. In particular the new expansion ''suggests that an easy estimator of \(\eta_0\) can be formed from the principal argument of \(\mathbb E_0^{\mathbb Q}(e^{iux_T/\sqrt T})\) together with an estimator of the spot volatility \(\sigma_0^2\)''. The resulting estimate of \(\eta_0\), however, will have a rather non trivial bias when \(r\) approaches 1. To improve then on the above estimator two characteristic functions \(\mathbb E_0^{\mathbb Q}(e^{iux_{T_1}/\sqrt {T_1}})\) and \(\mathbb E_0^{\mathbb Q}(e^{iux_{T_2}/\sqrt {T_2}})\) are used with two different \(T_1\) and \(T_2\), which are both approaching asymptotically zero. ''An appropriate combination of these (for different values of u) can annihilate the leading terms in them that is due to the jumps in \(x\), ... and lead to an estimate of \(\eta_0\) which has an asymptotic bias of order only \(O_p(\sqrt T_1)\). This bias is of even smaller asymptotic order in cases when \(x\) and its volatility \(\sigma\) do not jump together or if the jump compensator is smooth.'' The author moreover derives an associated Central Limit Theorem (CLT) which in turn allows for quantifying the precision in estimating \(\eta_0\) from the available options, and he further shows that his estimator of \(\eta_0\) is nearly efficient in a minimax sense. In the Section 2 the assumptions of the model are clearly stated: ``Assumption A1-r is a moment condition for the various processes appearing in the dynamics of x. Assumption A2-r imposes smoothness in expectation.'' The higher-order asymptotic expansion of the characteristic functions of Itô semimartingale increments are then derived in the Section 3 and summarized in the Theorem 1. It is also dicussed here how the Theorem 1 can be used to make inference for \(\eta_0\), and how the error in recovering \(\eta_0\) is reduced nontrivially when using two characteristic functions in the estimation. The Section 4 is then devoted to discuss in detail the estimation of the leverage effect \(\eta\): three further assumptions are introduced (``Assumption A3 is a moment condition on the increments of \(x\) and processes that show in its dynamics ... Assumption A4 is about the strike grid and A5 is about the option observation error''), the estimators are the presented in the equations (4.20)--(4.22) and their behavior is discussed in the Corollary 1 on the basis of the CLT of Theorem 2. The Theorem 3 in the Section 5 next gives a lower bound for the minimax risk for estimating the leverage coefficient from noisy short-dated option data; and finally in the Section 6 the performance of the leverage coefficient estimator \(\widehat\eta_{0,T_1,T_2}(u,v)\) is evaluated on the base of Monte Carlo simulated data. Lengthy proofs are wisely collected in the final Section 7 in order not to overload an already dense text.