A justification of conditional confidence intervals

DOI10.1214/21-EJS1833MaRDI QIDQ2044389zbMATH OpenOpenAlexFDO

Authors Alexander Heinemann, Stephan Smeekes, Eric Beutner

Publication date 9 August 2021

Published in Electronic Journal of Statistics (Search for Journal in Brave)

Full work available at URL https://arxiv.org/abs/1710.00643

conditional confidence intervals merging parameter uncertainty prediction sample-splitting

Stationary stochastic processes (60G10) Prediction theory (aspects of stochastic processes) (60G25) Nonparametric tolerance and confidence regions (62G15) Time series, auto-correlation, regression, etc. in statistics (GARCH) (62M10) Inference from stochastic processes and prediction (62M20)

Abstract: To quantify uncertainty around point estimates of conditional objects such as conditional means or variances, parameter uncertainty has to be taken into account. Attempts to incorporate parameter uncertainty are typically based on the unrealistic assumption of observing two independent processes, where one is used for parameter estimation, and the other for conditioning upon. Such unrealistic foundation raises the question whether these intervals are theoretically justified in a realistic setting. This paper presents an asymptotic justification for this type of intervals that does not require such an unrealistic assumption, but relies on a sample-split approach instead. By showing that our sample-split intervals coincide asymptotically with the standard intervals, we provide a novel, and realistic, justification for confidence intervals of conditional objects. The analysis is carried out for a rich class of time series models.

Review text

This paper deals with conditional confidence intervals. Its authors present a new method they call the ``sample-split approach. Next, let's see a brief introduction to the paper's mathematics. Let $(\Omega,F,P)$ be a probability space; $X_{t}:\Omega\to\mathbb{R}$ is a random variable, $X=(X_{t})_{t\in\mathbb{Z}}$ is the observed process for which one wishes to estimate a parameter, $\theta\in\Theta$, characterizing the stochastic structure of $X$, and the so-called prediction function $\Psi:\mathbb{R}^{\mathbb{Z}}\times\Theta\to\mathbb{R}$. For each $T\in\mathbb{Z}$ let $\Psi_{T+1}:\Omega\times\Theta\to\mathbb{R}$ be a function defined as \[ \Psi_{T+1}(\omega,\theta)=\Psi((X_{T}(\omega),X_{T-1}(\omega),\dots),\theta),\quad (\omega,\theta)\in\Omega\times\Theta. \] Some examples of interest discussed in the paper are: Example 1: $X=(X_{t})_{t\in\mathbb{Z}}$ is an AR(1) process defined by \[ X_{t}=\beta X_{t-1}+\varepsilon_{t} \] where $|\beta|<1$, $\varepsilon=(\varepsilon_{t})_{t\in\mathbb{Z}}$ are independent and identically distributed (i.i.d.) random variables with $E(\varepsilon_{t})=0$. If the conditional mean of $X_{T+1}$ given $X_{T}$ is to be estimated, the prediction function is \[ \Psi((X_{T},X_{T-1},\dots),\beta)=\beta X_{T}. \] Example 2: $X=(X_{t})_{t\in\mathbb{Z}}$ is a GARCH(1,1) process defined as \[ X_{t} = \sigma_{t}\varepsilon_{t}\sigma_{t}^2= w+\alpha X_{t-1}^2+\beta\sigma_{t-1}^2 \] where $w>0$, $\alpha\geq 0$, $1>\beta\geq 0$, $\varepsilon=(\varepsilon_{t})_{t\in\mathbb{Z}}$ is like before with $E(\varepsilon_{t}^2)=1$. It can be seen that \[ \sigma_{T+1}^2 = \frac{w}{1-\beta}+\alpha\sum_{k=0}^\infty\beta^{k}X_{T-k}^2. \] If the conditional mean of $\sigma_{T+1}^2$ given $X_{T},X_{T-1},\dots$ is to be estimated, the prediction function is \[ \Psi((X_{T},X_{T-1},\dots),\theta) = \frac{w}{1-\beta}+\alpha\sum_{k=0}^\infty\beta^{k}X_{T-k}^2, \] with $\theta=(w,\alpha,\beta)'\in(0,\infty)\times[0,\infty)\times[0,1)$. Let $s=(s_0,s_{-1},\dots)$ be a sequence of real numbers. An approximate prediction function is defined by \[ \Psi^{s}((X_{T},\dots,X_1),\theta)=\Psi((X_{T},\dots,X_1,s_0,s_{-1},\dots),\theta). \] It is shown that under certain not very restrictive assumptions and valid in general for the examples above, it is possible to define $s$ sequences that do not affect the estimates of $\theta$ and of the prediction function. Since the parameter $\theta$ is unknown, it must be estimated. Let $\theta(X_1,\dots,X_{T})$ be an estimator of $\theta$ based on $X_1,\dots,X_{T}$. Let us assume that the observed values of $X_1,\dots,X_{T}$ are $x_1,\dots,x_{T}$. The standard estimator of the prediction function based on $x_1,\dots,x_{T}$ is \[ \Psi_{T+1}^{STA}:=\Psi^{s}((x_{T},\dots,x_1),\theta(x_1,\dots,x_{T})). \] A confidence interval is constructed in practice by using the approximate quantiles of the ``random variable. \[ \Psi_{T+1}^{STA*}:=\Psi^{s}((x_{T},\dots,x_1),\theta(X_1,\dots,X_{T})). \] This is not correct because $X_1,\dots,X_{T}$ is considered to be both random and non-random simultaneously. The authors propose a new method to correct this flaw, which they call ``Sample-split estimation based on ``An intuitive motivation for the sample-split approach is the successive decline of the influence of past observations present in a substantial class of time series models. This property permits to split our sample into two (asymptotically) independent subsamples. The new estimator is then defined as follows: with $T$ greater than 1 being $1\leq T_{E}<T_{P}\leq T$, $\theta(X_1,\dots,X_{T_{E}})$ is an estimator of $\theta$. Then the prediction function estimator conditional on observing $X_{T_{P}}=x_{T_{P}},\dots,X_{T}=x_{T}$ is given by \[ \Psi_{T+1}^{SPL}:=\Psi^{s}((x_{T},\dots,x_{T_{P}},c_{T_{P}-1},\dots,c_1),\theta(X_1,\dots,X_{T_{E}})), \] where $c_{T_{P}-1},\dots,c_1$ are conveniently chosen initial constants. A conditional confidence interval can be defined by means of the conditional probability of $(X_1,\dots,X_{T_{E}})$ given $X_{T_{P}}=x_{T_{P}},\dots,X_{T}=x_{T}$. The problem that arises when using the standard technique to estimate the parameter and the prediction function is detailed with examples. However, it is shown that the proposed confidence intervals coincide asymptotically with the standard intervals, but the proposed technique is based on sound theoretical arguments as opposed to the assumptions on which the standard technique is based. One of the most interesting issues of the paper is perhaps the one discussed in Section 3 where it is shown that the concept of weak convergence is not adequate to analyze the asymptotic proximity of two sequences of conditional distributions. This concept is replaced by the concepts of ``merging and ``merging in probability of sequences of random elements, appeared in the last decades. Several assumptions are justified by the authors for the results about these concepts. Through a simulation study, the paper shows the superiority of the proposed method over the standard method from both the asymptotic and the finite sample point of view. The cases considered are very common in time series applications. In two appendices, the main results of the paper and the auxiliary ones necessary to prove them are rigorously and in detail proved. Extensive and updated reference list concludes the paper, which allows extending the subject under analysis in several areas of research.

Recommendations

Cites work

Cited in

(11)

This page was built for publication: A justification of conditional confidence intervals

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q2044389)