A non-parametric Bayesian approach to decompounding from high frequency data (Q1744221): Difference between revisions

Let \(N_t\), \(t \geq 0\) be a Poisson process with constant intensity \(\lambda\) and let \(Y_1,Y_2,\dots\) be a sequence of i.i.d. random variables with probability density \(f(x)\) independent from \(N_t\). A compound Poisson process is defined as \(X_t=\sum_{j=1}^J Y_j\) with \(X_0=0\) by definition. It is assumed that \(f(x)=\sum_{j=1}^J \rho_j \phi(x;\mu_j,\sigma)\) where \(\phi(x;\mu_j,\sigma)\) denotes the density of normal distribution with mean \(\mu_j\) and variance \(\sigma^2\). The prior distribution for \(\{\rho_j, \mu_j\}\) is assigned by a Dirichlet process. An independent prior for \(\lambda\) is assumed to be compactly supported. To estimate the posterior distribution for \(\lambda\) and \(f(x)\) one observes the discrete time sample \(X_\Delta,X_{2\Delta},\dots,X_{n\Delta}\). High frequency data are defined by requirement \(\Delta_n\to 0\) as \(n \to \infty\), but the results hold true for constant \(\Delta>0\) as well. It is also assumed that \(n \Delta_n \to \infty\) as \(n \to \infty\). In the article, a contraction rate of the posterior distribution to its true parameters \((\lambda_0, f_0)\) is established to be, up to a logarithmic factor, \((n \Delta)^{-1/2}\). Markov chain Monte Carlo simulation algorithm and simulation results are presented to confirm the theoretical estimates.