Bayesian inference on volatility in the presence of infinite jump activity and microstructure noise (Q2219235)

Advances in technology and computational power make high-frequency financial data widely available nowadays. The wide availability, in turn, implies an increasing demand for better modelling and statistical inference regarding volatility dynamics of the assets. A one-dimensional Lévy process defined on some probability space over a fixed time horizon with infinite jump activity and microstructure noise is considered one of the simplest, yet accurate enough, models for financial data at high-frequency. Utilizing this model and motivated by the approach recently presented in [\textit{R. Martin} et al., Stat. Probab. Lett. 134, 106--113 (2018; Zbl 1440.62315)], this paper proposes a `purposely misspecified' posterior of the volatility obtained by ignoring the process' jump component. By working with the ``purposely misspecified'' model, one avoids both the specification of a prior on the jump process and the necessity of obtaining samples from the full joint posterior. On the other hand, the bias and the inaccurate variance are corrected by applying a location shift and rescaling the likelihood using a Gibbs posterior. The main result of the paper establishes a Bernstein-von Mises (BvM) theorem, which states that the proposed adjusted posterior is asymptotically Gaussian, centered at a consistent estimator, and with variance equal to the inverse of the Fisher information. In this frame, the paper consists of nine sections and two appendices. In the sequel the subject of each one of them is briefly presented. Section 1 (Introduction) explains the theoretical and practical motivation of the current work and its main contribution. In Section 2 (Methodology), a detailed description of the setting and the model is provided. In Section 3 (Comparison with finite jump activity models), a motivating example using a simple finite jump activity is presented. The aim of the motivating example is to illustrate the usefulness of the approximate Bayesian inference obtained via purposeful misspecification. The differences between finite and infinite activity when deriving the ``purposely misspecified'' posterior are highlighted. In Section 4 (A semiparametric version of the misspecified BvM theorem), a modified version of the Bernstein-von Mises theorem is stated. The misspecified model is presented in Section 5 (The misspecified model). Specifically, Subsection 5.1 deals with the misspecified likelihood function and the corresponding MLE under the misspecified model proposed in the previous section, while Bayesian inference under the misspecified model is proposed in Subsection 5.2. In Section 6 (Correcting for misspecification) proposes some simple adjustments to correct for the effects of misspecification, for both the model without microstructure noise (Subsection 6.1) and the general model (Subsection 6.2). Section 7 (Extension to more general semimartingales without noise) extends the approach presented to make inference for the integrated variance of general Itô semimartingales in the absence of microstructure noise. Section 8 (Simulation) provides simulations to demonstrate the finite sample performance of the adjusted posterior, while Section 9 (Conclusion) concludes. All the proofs and technical details are given in Appendices A and B. Summarizing, this is a well-written paper which contributes to the Bayesian inference on volatility in the presence of infinite jump activity and microstructure noise.