Skip to Main Content
The online home for the publications of the American Statistical Association
CrossRef citations to date
Theory and Methods

Estimation of the Continuous and Discontinuous Leverage Effects

, , , &
Pages 1744-1758
Received 01 Sep 2015
Accepted 01 Sep 2016
Published online: 07 Aug 2017


This article examines the leverage effect, or the generally negative covariation between asset returns and their changes in volatility, under a general setup that allows the log-price and volatility processes to be Itô semimartingales. We decompose the leverage effect into continuous and discontinuous parts and develop statistical methods to estimate them. We establish the asymptotic properties of these estimators. We also extend our methods and results (for the continuous leverage) to the situation where there is market microstructure noise in the observed returns. We show in Monte Carlo simulations that our estimators have good finite sample performance. When applying our methods to real data, our empirical results provide convincing evidence of the presence of the two leverage effects, especially the discontinuous one. Supplementary materials for this article are available online.

Supplementary Materials

The online supplementary materials contain the appendices for the article.


Yacine Aït-Sahalia’s work was supported in part by the NSF under grant SES-0850533. Jianqing Fan’s work was supported in part by the NSF under grants DMS-1206464 and DMS-1406266. 3Roger J. A. Laeven’s work was supported in part by the NWO under grant VIDI-2009. 4Christina Dan Wang’s work was supported in part by the NIH under grant R01-GM100474 and by a grant from the Bendheim Center for Finance.


2 For example, extensive literatures in finance and econometrics are based on the model proposed by Heston (Citation1993), including (Yu Citation2005; Vetter Citation2012; Aït-Sahalia, Fan, and Li Citation2013).

3 This problem has also been studied by, for example, Fan and Wang (Citation2008) and Alvarez et al. (Citation2012). Moreover, Jacod and Todorov (Citation2010) and Li and Xiu (Citation2016) included the estimation of spot volatility as an intermediate step to their main goals. Just like the last two cited papers, the main target of this article is not the estimation of the spot volatility process either, and nontrivial results follow right after we obtain the estimator of the spot volatility process.

4 Different from CLE and DLE estimation, without correction, a bias is still present in the case of leverage parameter estimation, because the quadratic variation of estimated spot volatility involves squared estimated volatility changes.

5 More specifically, we have (δ1{|δ|κ})(μ-ν)t=0tR(δ(ω,s,x)1{|δ(ω,s,x)|κ})(μ-ν)(ds,dx) and (δ1{|δ|>κ})μt=0tR(δ(ω,s,x)1{|δ(ω,s,x)|>κ})μ(ds,dx).

6 In the definition of the co-jump leverage in Bandi and Renò (Citation2012), the jump intensity is further related to the level of the volatility (or variance), and the random jump sizes (in return and volatility series) are replaced by their correlation. As a result, the first source of randomness is restricted, although such an assumption embeds an interesting driving force in the dynamics of the jump intensity process, while the second source of randomness is eliminated.

7 In Jacod and Todorov (Citation2010), the authors studied a very general limit functional U(F)T (with minor changes in notations in what follows) given by (3.4) U(F)T=sTF(ΔXs,σs-2,σs2)1{ΔXs0},(3.4) together with its estimator (3.5) U(F,kn)T=i=kn+1T/Δn-knFΔinX,σ^i-2,σ^i+21{|ΔinX|>αn}.(3.5) Here F is a function on R×R+*×R+* where R+*=(0,). Observe that with the specification F(x, y, z) = x(zy), (Equation3.4

) and (Equation3.5) would be equivalent to (Equation2.5) and (Equation3.6) (the discontinuous leverage effect and its estimator), respectively. As we will see, however, the specific functional form of F will allow us to derive central limit results that apply under much broader conditions regarding jump stochasticity than those applying to a generic F.

8 Here, ZTnu.c.p.ZT means that the sequence of stochastic processes ZnT converges in probability, locally uniformly in time, to a limit ZT, that is, supsTZsn-ZsP0 for all finite T.

9 Refer to the decomposition of X in Equation (A.1) and note that the conditional expectation of (ΔnjX′′)σtnj2Δn is of the order Δ2n, the same as (ΔnjX′)σtnj2Δn.

10 But in Bandi and Renò (Citation2012), volatility jumps will slow down the convergence rate.

11 A possible explanation is that the Dow Jones Index is constructed from very liquid stocks. In fact, the realized variation first stays roughly the same as we increase the sampling frequency from 30-min to about 3-min, and then decreases (rather than increases) as we further increase the frequency to 5 sec. This is consistent with the finding that 5-sec returns are positively correlated, which might be explained by self- and/or mutual excitation in jumps (see, e.g., Aït-Sahalia, Cacho-Diaz, and Laeven Citation2015; Boswijk, Laeven, and Yang Citation2015). For instance, it could be the case that occurred jumps tend to excite other jumps with the same sign in the short run. If such scenario happens frequently, then the sample autocorrelations can be positive. Yet we don’t expect such same-sign-jump-exciting effect to last long, perhaps only a few minutes.

12 See also Fan and Fan (Citation2011) for an improved version of the test.

Log in via your institution

Log in to Taylor & Francis Online

Article Purchase
  • 24 hours access to article PDF & online version
  • Article PDF can be downloaded
  • Article PDF can be printed
USD 46.00 Add to cart
* Local tax will be added as applicable