A two-state regime switching autoregressive model with an application to river flow analysis

We propose a regime switching autoregressive model and apply it to analyze daily water discharge series of River Tisza in Hungary. The dynamics is governed by two regimes, along which both the autoregressive coefficients and the innovation distributions are altering, moreover, the hidden regime indicator process is allowed to be non-Markovian. After examining stationarity and basic properties of the model, we turn to its estimation by Markov Chain Monte Carlo (MCMC) methods and propose two algorithms. The values of the latent process serve as auxiliary parameters in the first one, while the change points of the regimes do the same in the second one in a reversible jump MCMC setting. After comparing the mixing performance of the two methods, the model is fitted to the water discharge data. Simulations show that it reproduces the important features of the water discharge series such as the highly skewed marginal distribution and the asymmetric shape of the hydrograph.     

Periodograms asymptotic distributions in periodically correlated processes and multivariate stationary processes: An alternative approach

Discrete time periodically correlated (PC) processes are viewed as the processes with time-dependent spectra. This, together with an auxiliary operator which is defined here is employed to apply classical results on the asymptotic distribution of the periodogram of the univariate white noise (innovations) to derive the asymptotic distributions of the periodograms for the PC processes and also for the multivariate stationary processes. We assume only the continuity and positive definiteness of the spectral densities together with the independence of the innovations.     

The Lindsay transform of natural exponential families

Let μ be an infinitely divisible positive measure on R. If the measure ρ_μ is such that x^-2[ρ_μ(dx)—ρ_μ({0})δ₀(dx)] is the Lévy measure associated with μ and is infinitely divisible, we consider for all positive reals α and β the measure T_α,β(μ) which is the convolution of μ*^α and ρ_μ*β. For example, if μ is the inverse Gaussian law, then ρ_μ is a gamma law with paramter 3/2. Then T_α,β(μ) is an extension of the Lindsay transform of the first order, restricted to the distributions which are infinitely divisible. The main aim of this paper is to point out that it is possible to apply this transformation to all natural exponential families (NEF) with strictly cubic variance functions P. We then obtain NEF with variance functions of the form □ΔP(□Δ), where A is an affine function of the mean of the NEF. Some of these latter types appear scattered in the literature.     

Computationally efficient classes of higher-order kernel functions

Classes of higher-order kernels for estimation of a probability density are constructed by iterating the twicing procedure. Given a kernel K of order l, we build a family of kernels K_m of orders l(m + 1) with the attractive property that their Fourier transforms are simply 1 — {1 —$(.)}^m+1, where ? is the Fourier transform of K. These families of higher-order kernels are well suited when the fast Fourier transform is used to speed up the calculation of the kernel estimate or the least-squares cross-validation procedure for selection of the window width. We also compare the theoretical performance of the optimal polynomial-based kernels with that of the iterative twicing kernels constructed from some popular second-order kernels.     

New estimation and feature selection methods in mixture‐of‐experts models

We study estimation and feature selection problems in mixture‐of‐experts models. An $l_2$ ‐penalized maximum likelihood estimator is proposed as an alternative to the ordinary maximum likelihood estimator. The estimator is particularly advantageous when fitting a mixture‐of‐experts model to data with many correlated features. It is shown that the proposed estimator is root‐$n$ consistent, and simulations show its superior finite sample behaviour compared to that of the maximum likelihood estimator. For feature selection, two extra penalty functions are applied to the $l_2$ ‐penalized log‐likelihood function. The proposed feature selection method is computationally much more efficient than the popular all‐subset selection methods. Theoretically it is shown that the method is consistent in feature selection, and simulations support our theoretical results. A real‐data example is presented to demonstrate the method. The Canadian Journal of Statistics 38: 519–539; 2010 © 2010 Statistical Society of Canada     

On Existence of Maximum Likelihood Estimators in Exponential Families

We propose a simple necessary and sufficient condition for existence of maximum likelihood estimators in a large class of canonical exponential families. We give an application to log-spline families.