Quasi-likelihood-based EM algorithm for regime-switching SDE

Yuzhong Cheng 1 *, Hiroki Masuda 2 · 2024

Read Quasi-likelihood-based EM algorithm for regime-switching SDE by Yuzhong Cheng 1 *, Hiroki Masuda 2 (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-06305.txt.

Introduction

Yuzhong Cheng and Hiroki Masuda put forth a way to gauge regime-switching diffusion models, in a paper from 2024. In such models, the state of a market or of the wild can jump from one mode to the next. A hidden Markov chain tracks those jumps as they come. The hidden jumps make the standard likelihood tools weak for this task. That holds though the flow that can be seen runs on in plain sight.

Their method joins two parts in one loop. One part is an expectation maximization loop, which handles latent variables. The other part is a quasi-likelihood built for noise with heavy tails. The method gauges the values in the dynamics and in the noise at the same time. It works from frequent discrete observations, read off again and again. Tests with made-up data show the method still gauges well when the task is hard.

Catalog record

CREATORYuzhong Cheng 1 *, Hiroki Masuda 2
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE91 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-06305

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Quasi-likelihood-based EM algorithm for regime-switching SDE Thanks: * Corresponding author Yuzhong Cheng 1 *, Hiroki Masuda 2 Abstract . This paper considers estimating the parameters in a regime-switching stochastic differential equation(SDE) driven by Normal Inverse Gaussian(NIG) noise. The model under consideration incorporates a continuous-time finite state Markov chain to capture regime changes, enabling a more realistic representation of evolving market conditions or environmental factors. Although the continuous dynamics are typically observable, the hidden nature of the Markov chain introduces significant complexity, rendering standard likelihood-based methods less effective. To address these challenges, we propose an estimation algorithm designed for discrete, high-frequency observations, even when the Markov chain is not directly observed. Our approach integrates the Expectation-Maximization (EM) algorithm, which iteratively refines parameter estimates in the presence of latent variables, with a quasi-likelihood method adapted to NIG noise. Notably, this method can simultaneously estimate parameters within both the SDE coefficients and the driving noise. Simulation results are provided to evaluate the performance of the algorithm. These experiments demonstrate that the proposed method provides reasonable estimation under challenging conditions. 1. Introduction We consider the problem of parameter estimation for stochastic differential equations (SDEs) with Markovian switching, a class of hybrid systems that integrates continuous dynamics with discrete events. More precisely, It is a stochastic process with two components, one component X describes the evolution of the process, and another component \alpha describes the switching between different regimes. This framework is widely applicable in modeling complex systems where continuous processes interact with discrete state transitions, such as those encountered in finance, ecology, and engineering. In this paper, we focus on applying an Expectation-Maximization (EM) algorithm to estimate parameters in Markov switching equations with hidden states. Besides the EM algorithm, another widely used approach to this problem is the Markov Chain Monte Carlo (MCMC) method, with relevant literature found in [ Hahn et al., 2010 ] and [ Hibbah et al., 2020 ] . For the EM approach, a previous study by [ Chevallier and Goutte, 2017 ] introduced a two-step method to estimate parameters in Markov switching stochastic differential equations with Lévy noise. Their approach first employs an EM algorithm designed to estimate parameters in a Markov switching SDE driven by Wiener noise, followed by a second step in which the parameters in the Lévy noise are estimated by fitting the noise distribution to each regime. Their study assumes fixed observation time steps. In contrast, this paper also considers Markov switching SDEs with Lévy noise but proposes an EM algorithm that can simultaneously estimate parameters in both the drift coefficient and the Lévy noise components. Our method is based on a quasi-likelihood approach for Lévy noise under high-frequency observations, allowing for an adaptive, unfixed time step size. Let (\Omega,\mathcal{F},\{\mathcal{F}_{t}\},\mathbb{P}) be a complete filtered probability space with the filtration \{\mathcal{F}_{t}\}_{t\geq 0} admitting usual conditions (i.e., it is right continuous and \mathcal{F}_{0} contains all \mathbb{P} -null sets). The process X=(X_{t})_{t\geq 0} satisfies an \mathbb{R} -valued stochastic differential equation with switching given by the following expression (1.1) dX_{t}=\lambda(b(\alpha_{t})-X_{t})dt+dZ_{t}, while \alpha=(\alpha_{t})_{t\geq 0} be a homogeneous continuous-time Markov chain with finite state space \mathcal{S}=\{1,2,...,N\} where N is a known positive integer. We always assume the Markov chain \alpha , the process Z in ( 1.1 ) and X_{0} are mutually independent. The generator Q=(q_{ij}) of this Markov chain \alpha is a N\times N matrix whose elements are characterized by (1.2) \begin{array}[]{c}\mathbb{P}(\alpha_{t+h}=j\mid\alpha_{t}=i)=\left\{\begin{array}[]{ll}q_{ij}h+o(h),&\text{ if }i\neq j,\\ 1+q_{ii}h+o(h),&\text{ if }i=j,\end{array}\right.\end{array} for h\downarrow 0 . Here we have q_{ij}>0 for i\neq j and q_{ii}=-\sum_{i\neq j}q_{ij} for all i . The driving process Z is a Normal Inverse Gaussian(NIG) Lévy process on \mathbb{R} with \mathcal{L}(Z_{1})\sim\operatorname{NIG}(a,0,\delta,0) . The density function of this NIG distribution is given by (1.3) f(z)=\frac{a\delta}{\pi}\exp\left\{\delta a\right\}\frac{K_{1}\left(a\sqrt{\delta^{2}+z^{2}}\right)}{\sqrt{\delta^{2}+z^{2}}}, where \delta>0 , a>0 , K_{1} is the modified Bessel function of third order with index 1. It is well known that NIG distribution is closed in convolution operation. Therefore we have Z_{t}\sim\operatorname{NIG}(a,0,\delta t,0) . For more properties of NIG distribution, see [ Barndorff-Nielsen, 1997 ] . In the equation ( 1.1 ), \lambda\in\mathbb{R}_{+} , and b(\alpha_{t})_{t\geq 0} is a continuous-time Markov chain with state space \{b(1),b(2),...,b(N)\} , b(i)\neq b(j) for i\neq j . The existence and uniqueness of a strong solution of equation ( 1.1 ) is guaranteed by Theorem 2.1 in [ Xi and Zhu, 2017 ] . In our case, X is observed and \alpha is latent (unobserved). Let h>0 , t_{j}=jh for j=0,1,...,n and T_{n}=nh . The observation is the vector (X_{t_{0}},...,X_{t_{n}}) . We should let T_{n} be large enough to ensure that the solution of ( 1.1 ) switches through all states in the state space \mathcal{S} . We assume the following high-frequency and long-term sampling conditions: T_{n}:=nh\to\infty and h\to 0 as n\to\infty . Let \theta=(b(1),b(2),...,b(N),\lambda,\delta), and \theta\in\Theta where \Theta=\prod_{i=1}^{N}\Theta_{b(i)}\times\Theta_{\lambda}\times\Theta_{\delta} , and N\geq 2 be given. All \Theta_{\lambda}\subset\mathbb{R}_{+},\Theta_{\delta}\subset\mathbb{R}_{+},\Theta_{b(i)}\subset\mathbb{R} for i=1,...,N are compact convex sets. Our objective is to estimate the parameter \theta from (X_{t_{j}})_{j=0}^{n} . We should mention here that the parameter a is a nuisance parameter, which may be unknown. To simplify the computational complexity of the estimation procedure, we assume that Q is known and fixed. The primary focus of this study is on the estimation of \theta . When the generator Q is unknown, refer to Remark 3.2 . The letter C denotes a positive constant which may change from location to location. We use p_{\theta} as a generic symbol for densities and distributions parameterized by \theta . We will write \mathbb{X}_{1,n}:=(X_{t_{1}},...,X_{t_{n}}) , \mathbb{A}_{1,n}:=(\alpha_{t_{1}},...,\alpha_{t_{n}}) . The paper is organized as follows: In Section 2 , we introduce the Cauchy quasi-likelihood approach to the NIG distribution. Section 3 begins with a review of the standard EM algorithm, followed by our main contribution, proposing a numerical method based on an approximated EM algorithm, including methods for calculating the probability values required in the algorithm. In Section 4 , we present numerical experiments to evaluate the performance of our method. Finally, an Appendix provides relevant background information. 2. Quasi likelihood approach for NIG Lévy noise In this section, we introduce a Cauchy quasi-likelihood function for the equation ( 1.1 ), where the driving noise Z is modeled as a NIG Lévy process. Quasi-likelihood functions are well established in the literature and are an effective approach for approximating the true likelihood function. Gaussian quasi-likelihood functions are commonly used for models driven by a Wiener process, while for more general driving noises, such as jump processes, non-Gaussian quasi-likelihood functions may also be employed. This has been demonstrated to yield satisfactory statistical properties, as discussed in [ Masuda, 2019 ] and [ Clément and Gloter, 2020a ] . When Z is a NIG Lévy process where Z_{1} follows \operatorname{NIG}(a,0,\delta,0) where the density is given in ( 1.3 ), we have that Z_{h} and \Delta_{j}Z follow the same distribution, and thus Z_{h} follows an \operatorname{NIG}(a,0,\delta h,0) distribution. Therefore, by Euler approximation, we have the approximate expression for \Delta_{j}Z : X_{t_{j}}-\left(X_{t_{j-1}}+\lambda(b(\alpha_{t_{j-1}})-X_{t_{j-1}})h\right)\approx\Delta_{j}Z\sim\operatorname{NIG}(a,0,\delta h,0). Also it is known that for any c,d\in\mathbb{R} Y=cX+d\sim\operatorname{NIG}\left(\frac{a}{|c|},0,c\delta,d\right), when X\sim\operatorname{NIG}(a,0,\delta,0) , see [ Chevallier and Goutte, 2017 ] . From this, we obtain that \frac{\Delta_{j}Z}{\delta h}\sim\operatorname{NIG}(ah|\delta|,0,1,0). Substituting the approximate expression for \Delta_{j}Z into the above relation, we have \frac{X_{t_{j}}-\mu_{j-1}(\lambda)}{\delta h}\approx\frac{\Delta_{j}Z}{\delta h}\sim\operatorname{NIG}(ah|\delta|,0,1,0), where \mu_{j-1}(\lambda):=X_{t_{j-1}}+\lambda(b(\alpha_{t_{j-1}})-X_{t_{j-1}})h. As noted in [ Kawai and Masuda, 2013 ]

Read the complete text