EconBase
← Back to paper

Prediction in locally stationary time series

Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.

59,930 characters · 8 sections · 73 citation commands

Rendered from LaTeX for readability, not typeset faithfully. Citation keys are highlighted; maths is left as source; figures, tables and equation environments are summarised rather than reproduced; unrecognised commands are greyed out so nothing is silently dropped. Email addresses are removed.

Prediction in locally stationary time series

abstractWe develop an estimator for the high-dimensional covariance matrix of a locally stationary process with a smoothly varying trend and use this statistic to derive consistent predictors in non-stationary time series. In contrast to the currently available methods for this problem the predictor developed here does not rely on fitting an autoregressive model and does not require a vanishing trend. The finite sample properties of the new methodology are illustrated by means of a simulation study and a financial indices study.

AMS subject classification: 62M10; 62M20

Keywords and phrases: locally stationary time series, high dimensional auto-covariance, matrices, prediction, local linear regression,

Introduction

\setcounter{equation}{0}

An important problem in time series analysis is to predict or forecast future observations from a given a stretch of data, say $X_{1}, \ldots , X_{n}$, and numerous authors have worked on this problem. Meanwhile there is a well developed theory for prediction under the assumption of stationary processes [see for example brockwell2002introduction, bickel2011banded, mcmurry2015high among many others]. On the other hand, if data is obtained over a long stretch of time it may be unrealistic to assume that the stochastic structure of a time series is stable. Moreover, in many shorter time series non-stationarity can also be observed and prediction under the assumption of stationarity might be misleading.

A common approach to deal with this problem of non-stationarity is to assume a location scale model with a smoothly changing trend and variance but a stationary error process, say $X_n = \mu (n) + \sigma (n) \varepsilon_n$ [see, for example, van2004forecasting, stuaricua2005nonstationarities, zhao2008confidence, guillaumin2017analysis, das2017predictive]. In this case the trend and variance function can be estimated and prediction can be performed applying methods for stationary data to the standardized residuals. However, there appear also more sophisticated features of non-stationarity in the data, which are not captured by a a simple location scale model, such as time-changing kurtosis or skewness, and the standardized residuals obtained by this procedure may not be stationary.

To address this type of non-stationarity various mathematical concepts modeling a slowly-changing stochastic structure have been developed in the literature [see for example, priestley1988non, dahlhaus1997fitting, nason2000wavelet, zhou2009local or VO2012]. The corresponding stochastic processes are usually called {\it locally stationary} and the problem of predicting future observations in these models is a very challenging one. An early reference is fryzlewicz2003forecasting who considered centered {\it locally stationary wavelet processes}. In this model the sample covariance matrix in the prediction equation is not estimable and the authors proposed an approximation using the (uniquely defined) wavelet spectrum. van2004forecasting considered the prediction problem in a location scale model with a smoothly changing variance and stationary error process. More recent work on forecasting in centered locally stationary time series can be found in roueff2018prediction and kley2019predictive. The first named authors investigated a predictor based on auto-regression of a given order, while kley2019predictive considered predictors in stationary and locally stationary models for (possibly) non-stationary data and selected the “better” prediction among the two estimates. A common feature of most of these methods is that they are all based on auto-regressive fitting.

In the present paper we contribute to this literature and propose an alternative method for prediction in physically dependent locally stationary times series, which does not rely on auto-regressive fitting and is therefore more flexible. To be precise we consider the model

align[align omitted — 74 chars of source]

where $\mu$ is a deterministic and smooth mean or trend function on the interval $[0,1]$ and $\{ \epsilon_{i,n} : i =1, \ldots , n \}_{n\in \mathbb{N}} $ is a triangular array modelled by a locally stationary process in the sense of zhou2009local - see Section (ref) for mathematical details. We then estimate the regression function $\mu$ by local linear smoothing and define a banded estimator for the corresponding auto-covariance matrix

align[align omitted — 100 chars of source]

from the residuals of the nonparametric fit, where the width of the band increases with the sample size. Banded estimates of auto-covariance matrices have been considered by wu2009banding and mcmurry2010banded for {\bf centered} and {\bf stationary} processes using the fact that in this case the matrix $\Sigma_{n}$ in (ref) is a Toeplitz matrix. Neither of these results is applicable under the assumption of non-stationarity (even if the locally stationary process $\{X_{i,n}\}_{i=1,\ldots,n}$ in (ref) is centered).

In Section (ref) we establish consistency (with respect to the operator norm) of the new covariance operator for locally stationary processes with a time varying mean function. These results are then used in Section (ref) to develop new prediction methods, which - in contrast to the currently available literature - do not use autoregressive fitting. In Section (ref) we investigate the finite sample properties of the estimator of the covariance matrix and compare the new predictor with the currently available methodology. Finally, all proofs of our main theoretical results and technical details can be found in Section (ref).

Locally stationary processes

\setcounter{equation}{0}

Consider the time series model (ref) where $\{ \epsilon_{i,n}: i=1,\ldots,n\}_{n \in \mathbb{N}}$ is an array of centered random variables, and $\mu:[0,1]\rightarrow \mathbb R$ is a smooth mean function. More precisely we assume

description• (M1) The function $\mu$ in model (ref) has a Lipschitz continuous second order derivative on the interval $[0,1]$.

In order to model a local stationary error process we use a concept introduced by zhou2009local. To be precise, define for an $L^q$-integrable random variable $X$ its norm by $\| X \|_q=(\mathbb{E} [|X|^q])^{1/q} (q \geq 1)$, let $\{\varepsilon_i : i \in \mathbb{Z} \}$ denote a sequence of independent identically distributed observations and define $\mathcal{F}_i = ( \ldots ,\varepsilon_{i-2},\varepsilon_{i-1},\varepsilon_i )$. We assume that there exists a function $G : [0,1]\times \mathbb R^\mathbb{N} \rightarrow \mathbb R$ such that

align[align omitted — 66 chars of source]

is a well defined random variable. For arbitrary functions $G$ it is not guaranteed that the stochastic structure of $\{\epsilon_{i,n} \colon i \in \mathbb Z\} $ varies smoothly, but we can achieve this by the following assumptions.

description• (L1) For some $q\geq 2$ we have that $$ \sup_{t\in[0,1]}\|G(t,\mathcal{F}_0)\|_q<\infty. $$ • (L2) The function $G$ is differentiable with respect to the first coordinate and there exists a constant $M>0$ such that for all $t,s\in[0,1]$ $$ \Big \|\frac{\partial}{\partial t}G(t,\mathcal{F}_0)-\frac{\partial}{\partial t}G(s,\mathcal{F}_0) \Big \|_2\leq M | t-s | . $$

Next we quantify the dependence structure. For this purpose let $\{\varepsilon_i' : i \in \mathbb{Z} \} $ denote an independent copy of $\{\varepsilon_i : i \in \mathbb{Z}\}$, define $\mathcal{F}_i^* =(\ldots ,\varepsilon_{-2},\varepsilon_{-1},\varepsilon_0',\varepsilon_{1}, \ldots ,\varepsilon_i)$ and

align*[align* omitted — 92 chars of source]

as a measure of dependence. We assume for the same $q\geq 2$ as in assumption (L1) that

description• (L3) There exists a constant $\chi\in(0,1)$ such that $$ \delta_q(G,i)=O(\chi^{i}). $$
example{\rm A prominent example of this non-stationary model is a locally stationary $AR(p$) process where the filter in (ref) is defined by \begin{align} G(t,\mathcal{F}_i)=\sum_{s=1}^pa_s(t)G(t,\mathcal{F}_{i-s})+{\sigma}(t)\varepsilon_{i} \end{align} where $(\varepsilon_i)_{i\in \mathbb{Z}}$ is a sequence of independent identically distributed centered random variables with $\|\varepsilon_1\|_q<\infty$, and $a_1 , \ldots , a_{p}, \sigma: [0,1] \to \mathbb{R}$, are for smooth functions such that for some $\delta_0>1$ the polynomial $1-\sum_{s=1}^pa_s(t)z^s$ has {no} roots in the disc $\{ z \in \mathbb{C} \colon |z|\leq \delta_0 \} $. If the functions $a$ and {$\sigma$} have bounded derivatives, $G(t,\mathcal{F}_i)$ has a MA representation of the form $G(t,\mathcal{F}_i)={\sigma}(t)\sum_{j=0}^\infty c_j(t)\epsilon_{i-j}$, where $c_1, c_2 , \ldots $ are smooth functions with derivatives satisfying $|c'_j(t)|\leq M\chi^j$ for $j\geq 0$. Therefore assumptions (L1)-(L3) hold for model (ref). It has been shown in zhou2013inference that Model (ref) can approximate the time-varying $AR(p)$ model in dahlhaus1997fitting. }
remark{\rm Note that the definition of a locally stationary error process contains the case that each row of $\{\epsilon_{i,n} \colon i \in \mathbb Z\}_{n\in \mathbb{N}} $ is stationary, that is $G(t, \mathcal{F}_i ) = H(\mathcal{F}_i) $ for some function $H : \mathbb R^\mathbb{N} \to \mathbb{R}$. In this case the random variables $ \epsilon_{i,n}=H(\mathcal{F}_i) $ do not depend on $n$, Assumption (L2) is obviously satisfied and Assumption (L1) and (L3) reduce to \begin{description} • (S1) For some $q\geq 2$, $\|H(\mathcal{F}_0)\|_q<\infty$. • (S2) There exists a constant $\chi\in(0,1)$ such that $$ \delta_q(H,i)=\|H(\mathcal{F}_i)-H(\mathcal{F}^*_i)\|_q =O(\chi^{i})~. $$ \end{description} }

If assumption (L1) holds the covariance matrix $\Sigma_n=(\sigma_{i,j,n})_{1\leq i,j\leq n} $ in (ref) is well defined, where

align[align omitted — 124 chars of source]

Throughout this paper we do not reflect the dependence on $n$ in the notation of the entries of a matrix, whenever it is clear from the context. For example we will use $\sigma_{i,j}$ instead {of} $\sigma_{i,j,n}$ and similarly a simplified notation for corresponding estimates. We also define the (time dependent) auto-covariances

align[align omitted — 116 chars of source]

of the stationary (for fixed $t \in [0,1]$) process $\{G(t,\mathcal{F}_i)\}_{i \in \mathbb{Z}} $. To estimate the covariances in (ref) we use a local linear regression estimate of the function $\gamma_k$. In order to prove consistency of this estimator we require a smoothness condition on the auto-covariances in (ref), which is formulated as follows.

description• (A1) For any $ k \in \mathbb{Z}$ the function $\gamma_k$ in (ref) is differentiable with derivative $\dot \gamma_k(t)=\frac{\partial}{\partial t}\gamma_k(t)$. There exists constants $D_k$ such that for all $t,s\in [0,1]$ \begin{align*} \left |\dot \gamma_k(t)-\dot \gamma_k(s)\right|\leq D_k|t-s|. \end{align*}

An application of the Cauchy-Schwarz inequality and the dominated convergence theorem show that a sufficient condition for assumptions (L2) and (A1), is given by (L1) and

align*[align* omitted — 105 chars of source]

In the following section we will use the local linear estimates for the function $\gamma_k$ to define a banded estimate of the covariance matrix $\Sigma_n$ of a locally stationary process of the form (ref) and investigate its asymptotic properties for increasing sample size. We also discuss a corresponding estimator in the stationary case because usually estimators are studied under the assumption of a centered stationary process, that is $\mu \equiv 0$. In the subsequent Section (ref) we use these results for prediction in locally stationary processes with a non-vanishing trend.

Covariance matrix estimation

\setcounter{equation}{0}

The estimation of the covariance matrix has attracted considerable attention in the literature. We refer among many others to the work of bickel2008covariance, bickel2008regularized for high-dimensional independent identically distributed data and anderson2003introduction, wu2009banding, chen2013covariance, box2015time, and mcmurry2015high who considered this problem for time series. Most authors consider the case of a vanishing trend, i.e. $\mu \equiv 0$, and assume that the error process $\{\epsilon_{i,n} : i=1,\ldots,n \}$ is a sequence of independent identical observations or a stationary series. For example, in the case of a stationary centered process wu2009banding proposed the banded estimator

align[align omitted — 121 chars of source]

of the matrix $\Sigma_n$, where $\mathbf 1(A) $ denotes the indicator function of the set $A$ and $$ \tilde \sigma_{i,j}=\frac{1}{n-|i-j|}\sum_{s=1}^{n-|i-j|}X_{s,n}X_{s+|i-j|,n}, $$ is the sample auto-covariance of $\{X_{1,n}, \ldots , X_{n,n} \}$ at lag $|i-j|$ and $l_n \in \mathbb{N}$ denotes a tuning parameter satisfying $l_n\rightarrow \infty$, $l_n=o(n)$ as $n \to \infty$. mcmurry2010banded modified this statistic such that the new estimator leaves the band intact, and then gradually down-weighs increasingly distant off-diagonal entries instead of setting them to zero as in the banded matrix case. Both estimators use the fact that for stationary processes the matrix $\Sigma_n$ is a Toeplitz matrix.

Note that the estimator (ref) is not consistent for the auto-covariance if the mean function is not constant. As there are many applications where time series have a smoothly changing mean function we begin our discussion analyzing a mean-corrected estimator of the matrix $\Sigma_n$ for a stationary {error} process of the form (ref), which avoids this problem.

Let $\hat \mu $ be the local linear estimator defined by

align[align omitted — 214 chars of source]

where $\tau_n$ denotes the bandwidth. For the kernel $K$ we make the following assumption:

description• (K) The kernel $K$ is a symmetric, continuously differentiable, bounded density function supported on the interval $[-1, 1]$.

We consider the residuals

align[align omitted — 68 chars of source]

obtained from the local linear fit and denote by $$ \hat\sigma^\dag_{i,j} =\frac{1}{n-|i-j|}\sum_{s=1}^{n-|i-j|}\hat \epsilon_{s,n}\hat \epsilon_{s+|i-j|,n} ~~~(i,j=1 , \ldots , n) $$ the sample auto-covariance of the residuals $\{\hat \epsilon_{1,n} , \ldots , \hat \epsilon_{n,n} \}$ at lag $|i-j|$. Finally, we define for $ l_{n} \in \mathbb{N} $ the banded matrix

align[align omitted — 108 chars of source]

as an estimator of the matrix $\Sigma_n$. It will be shown below that the estimator $\hat \Sigma^\dag_{n}$ is consistent for $\Sigma_n$ in the case of a strictly stationary error process. To measure the distance between two matrices (of increasing dimension) we introduce the operator norm

align*[align* omitted — 55 chars of source]

of a matrix $A$, where $|\cdot|$ denotes the Euclidean norm (note that $\rho^2(A)$ is the largest eigenvalue of the matrix $A^\top A$).

theoremAssume that $n\tau_n^6=o(1)$, $n\tau_n^3\rightarrow \infty$, $l_n\rightarrow \infty$, $\frac{l_n^2}{n}=o(1)$. If conditions (K), (S1), (S2) and (M1) hold, then \begin{align*} \|\rho(\hat\Sigma^\dag_{n} -\Sigma_n)\|_{q/2} = O (r^\diamond_n) , \end{align*} where the sequence $r^\diamond_n$ is defined by $$ r^\diamond_n=l_n(\tau_n^2+(n\tau_n)^{-1/2})+\frac{l_n^2}{n}+\chi^{l_n}. $$

Theorem (ref) establishes consistency of the estimator of the covariance matrix in model (ref) in the operator norm under the assumption of a stationary error process. However, there also exist many time series exhibiting a non-stationary behaviour in the higher order moments and dependence structure [see stuaricua2005nonstationarities, elsner2008increasing, guillaumin2017analysis among others], and estimation under the assumption of a location model with a stationary error process might be misleading. In this case the estimator $\hat \Sigma^\dag_{n}$ in (ref) is not necessarily consistent since the unknown covariance matrix $\Sigma_n$ is not a Toeplitz matrix. To address this problem we propose an alternative approach which also yields a consistent estimator for non-stationary time series. Roughly speaking, we estimate the elements $\sigma_{i,j} $ in the matrix $\Sigma_n$ by

align[align omitted — 98 chars of source]

where $\hat \gamma_k (t) $ is a local linear estimate of the auto-covariance function (ref) of the process $\{G(t,\mathcal{F}_i)\}_{i \in \mathbb{Z}} $.

To be precise, we distinguish between a lag of odd or even order and define

align[align omitted — 254 chars of source]

if the lag $k$ is of even order, where $b_n$ is a bandwidth and the residuals $\hat \epsilon_{i,n} $ are defined in (ref). In (ref) we use the notation $\hat \epsilon_{i,n}=0$ if the index $i$ satisfies $i<0$ or $i>n$. Similarly, for an odd lag $k$ we define

align[align omitted — 104 chars of source]

where

align*[align* omitted — 499 chars of source]

The estimator of the element $\sigma_{i,j}$ in $\Sigma_n$ is finally defined by (ref) and for the covariance matrix we use again a banded estimator, that is

align[align omitted — 148 chars of source]

Our next result yields the consistency of this estimator in the operator norm.

theoremAssume that $n\tau_n^3\rightarrow \infty$, $n\tau_n^6=o(1)$, $\frac{l_n^2}{n}=o(1)$, $l_nb_n^2=o(1),$ $$l_n((nb_n)^{-1/2}b_n^{-2/q}+\tau_n^2+(n\tau_n)^{-1/2})=o(1) ~ \mbox{ and } ~~~b_n^2\sum_{k=0}^{l_n}D_k=o(1) $$ If the conditions (K), (L1)--(L3), (A1) and (M1) are satisfied, then we have \begin{align*} \|\rho(\hat \Sigma_{n}-\Sigma_n)\|_{q/2}=O( r_n), \end{align*} where the sequence $r_n$ is defined by \begin{equation} r_n=l_n((nb_n)^{-1/2}b_n^{-2/q}+\tau_n^2+(n\tau_n)^{-1/2})+\frac{l_n^2}{n}+\chi^{l_n}+ b_n^2 \sum_{k=0}^{l_n}D_k = o(1) . \end{equation}
remark{\rm \begin{itemize} • In the case of a stationary and centered time series it has been demonstrated by mcmurry2015high that tapering can improve the performance of simply banded estimators of the covariance matrix and similar arguments apply to the covariance estimators (ref) and (ref) proposed in this paper for stationary times series with a time varying mean function and for locally stationary times series. To be precise consider the situation in Theorem (ref) and define the tapering function (other tapers could be used as well) by \begin{align*} \kappa(x)= (2-|x| ) \mathbf{1}(1\leq |x|\leq 2 ) + \mathbf{1}(|x|<1) \end{align*} The tapered and banded estimate of the covariance matrix $\Sigma_n$ is now defined by \begin{align*} \hat \Sigma^{tap}_{n}:=\Big (\kappa\Big (\frac{|i-j|}{l_n}\Big )\tilde \gamma_{|i-j|}(\frac{i+j}{2n})\Big)_{1\leq i,j\leq n}. \end{align*} Using the same arguments as in the proof of Theorem (ref) it can be shown that \begin{align} \|\rho(\hat \Sigma^{tap}_{n}-\Sigma_n)\|_{q/2} = O (r_n) ,\notag \end{align} where the sequence $r_n$ is defined in (ref). • It is worthwhile to mention that recently ding2018estimation proposed an alternative estimate of the the precision matrix $\Sigma_{n}^{-1}$ of a {\bf centered} locally stationary series, which is based on a Cholesky decomposition. In contrast the estimator $\hat \Sigma_{n}^{-1}$ considers the inverse of a banded estimator of the covariance matrix of a locally stationary series with a smoothly varying trend. \end{itemize} }

Prediction

\setcounter{equation}{0}

In this section we discuss some applications of the proposed estimators in the problem to perform predictions in locally stationary processes. For centered time series this problem has been recently investigated by roueff2018prediction, kley2019predictive who proposed to fit a locally stationary AR model and perform the prediction using an AR approximation. In this section, we suggest an alternative method which is not based on AR fitting. To be precise, assume that we observe a stretch of data $X_{1,n}, \ldots, X_{m,n}$ from the model (ref) and that we are interested in a prediction of the next observation $X_{m+1,n}$. To be precise, our aim is the construction of best linear predictor of $X_{m+1,n}$ based on $X_{1,n},\ldots ,X_{m,n}$. For this purpose we define

align[align omitted — 141 chars of source]

where $\mathbf X_{m,n}=(1,X_{1,n},...,X_{m,n})^\top$ and the prediction vector $\mathbf a_m=(a_{m+1,n},a_{m,n}...,a_{1,n})^\top:=(a_{m+1,n}, (\mathbf{a_m}^{*} )^{\top} )^{\top}$ is given by

align[align omitted — 199 chars of source]

In order to estimate the vector $ \mathbf a_m$ we define the local linear estimators from the sample $X_{1,n}, \ldots, X_{m,n}$ by

align[align omitted — 226 chars of source]

and denote by

equation[equation omitted — 142 chars of source]

the covariance matrix of the vector $(X_{1,n}, \ldots, X_{m,n})^T$. The residuals (ref) for estimating the auto-covariances are then replaced by residuals by $$ \hat \epsilon_{i,n}^{1:m} =X_{i,n}-\hat \mu^{(1:m)} (i/n) \quad (i =1, \ldots , m) $$ from the nonparametric fit from the data $X_{1,n}, \ldots, X_{m,n}$. Next, we define $\hat \gamma_k^{1:m}$ as the analogue of the estimator (ref) (if the lag $k$ is even) and (ref) (if the lag is odd), where the residual $\hat \epsilon_{\ell,n}$ is replaced by $\hat \epsilon_{\ell,n}^{1:m}$. We further define

align[align omitted — 159 chars of source]

as a banded estimator of the covariance matrix $\Sigma_{n,m}:=\text{Cov}(X_{i,n},X_{j,n})_{1\leq j\leq m}$ in (ref). It can be shown that, if the assumptions of Theorem (ref) are satisfied and $m\geq \lfloor cn\rfloor$ for some positive constant $c$,

align[align omitted — 79 chars of source]

where the sequence $r_n$ is defined in (ref). We shall construct a predictor based on $ \hat \Sigma_{n,m}^{-1}$ and for this purpose we show that the consistency of the estimator $\hat \Sigma_{n,m}$ in (ref) can be transferred to its inverse. \\ Throughout this paper we denote $\lambda_{min}(A)$ the minimum eigenvalue of a symmetric matrix $A$ and make the following assumption.

description• (E1) There exists a constant $c>0$ such that $$ \eta= \liminf_{n\to \infty }\inf_{\lfloor cn\rfloor\leq m\leq n}\lambda_{min}(\Sigma_{n,m}) >0. $$
corollaryAssume that the conditions of Theorem (ref) and condition (E1) are satisfied. If $n\to \infty $, $\lfloor cn\rfloor\leq m\leq n$ we have \begin{align} \rho(\hat \Sigma_{n,m}^{-1}-\Sigma_{n,m}^{-1})= O_{\mathbb{P}}(r_n) \end{align}

We can now define an estimate $\hat{ \mathbf {a}}_m=(\hat a_{m+1,n},(\hat {\mathbf{a}}_m^*)^\top)^\top$ of the vector $\mathbf a_m$ in (ref) by $$ \hat a_{m+1,n} =\hat \mu^{1:m}(m/n)-\sum_{s=1}^m\hat a_{m+1-s,n}\hat \mu^{1:m}(s/n),\label{hatam}\\ $$ and

eqnarray[eqnarray omitted — 160 chars of source]

where

align[align omitted — 309 chars of source]

The final predictor of $X_{m+1,n}$ is defined by

align[align omitted — 111 chars of source]
theoremAssume that the conditions of Theorem (ref) and assumption (E1) are satisfied, $\liminf_{n\rightarrow 0} \frac{l_n}{\log n}\geq \eta>0$ and assume that there exists a constant $c\in(0,1)$ such that for $m\geq cn$, $m\geq \lfloor nb_n\rfloor$. (a) The vector $\hat {\mathbf{a}}_m=(\hat a_{m+1,n}, (\hat {\mathbf a}_m^*)^\top)^\top$ is a consistent estimator of the coefficient vector $\mathbf a_m$ of the best linear predictor defined in (ref), i.e., \begin{align*} |\hat {\mathbf a}^*_m-\mathbf a^*_m|=O_{\mathbb{P}}(r_n), \quad\hat a_{m+1,n}- a_{m+1,n}=O_{\mathbb{P}}(r^\circ_n) \end{align*} where $r_n$ is defined in (ref), and \begin{align} r_n^\circ =(l_n^{1/2}\log^{1/2} n)r_n+\sqrt n\chi^{l_n}. \end{align} (b) Assume that $r_n^\circ=o(1)$. If the error $\epsilon_{i,n}$ is a locally stationary AR($p$) process as defined in Example (ref) and \begin{description} • (P1) $n^{\frac{1}{q}}r_n=o(1)$. • (P2) $\delta_q(\dot G, i)=O(\chi^i)$, • (P3) $\sup_{t\in [0,1]}\|\dot G(t,\mathcal{F}_i)\|_q<\infty$, \end{description} where $\dot G(t,\mathcal{F}_i)=\frac{\partial}{\partial t}G(t,\mathcal{F}_i)$ denotes the derivative of the filter $G$, we have \begin{equation} \frac{X_{m+1,n}-\hat X^{\rm Pred}_{m+1,n}}{\sigma(\frac{m+1}{n})}\Rightarrow \varepsilon_{1} \end{equation} where $\Rightarrow $ denotes the convergence in distribution and $\varepsilon_1$ denotes the error in model (ref) .

The rate $r_n^\circ$ in (ref) results from convergence rate of the nonparmetric estimate of the time-varying mean and does not appear if the trend is not estimated because it is known to be $0$. Conditions (P2) and (P3) can be verified by checking the coefficients of the MA representation of the locally stationary AR process (ref). They assure that for any $i,j$, the process $\{\mathbb{E}(G(t,\mathcal{F}_i)G(s,\mathcal{F}_j))\}_{t,s\in [0,1]}$ is sufficiently smooth on $[0,1]\times [0,1]$.

remark{\rm Similar arguments as given in the proof of Theorem (ref) show that the estimator $\hat \Sigma_{n,m}$ is positive definite if the sample size is sufficiently large. However, for finite sample sizes the matrix $\hat \Sigma_{n,m}$ can be singular. As the prediction in (ref) requires a non-singular sample covariance matrix we propose in applications to replace the estimator $\hat \Sigma_{n,m}$ by a a positive definite estimator, say $\hat \Sigma^{pd}_{n,m}$, which is defined as follows. If $\hat \Sigma_{n,m}=U_{n,m}V_{n,m}U_{n,m}^{\top}$ is the spectral decomposition of $\hat \Sigma_{n,m}$ and $V_{n,m} = \mbox{diag} (v_1,\ldots, v_{m})$ is the diagonal matrix containing the corresponding eigenvalues, we define \begin{equation} \hat \Sigma^{pd}_{n,m}:=U_{n,m}V^{+}_{n,m}U_{n,m}^{\top} \end{equation} where $V^{+}_{n,m}$ is a diagonal matrix with its $i$th diagonal element given by $$ v_{i}^{+} = \max \Big \{v_i, \frac{10 \int_{0}^{\frac{m}{n}} \hat \gamma^{1:m}_0(t)dt}{m^\beta}\Big \}~,~i=1,\ldots. m $$ for some $\beta>0$. As a rule of thumb, we choose $\beta=0.5$ because for this choice $\rho(\hat \Sigma_{n,m}^{pd}-\hat \Sigma_{n,m})=O(n^{-\beta}) =O(r_n)$. This type of modification has been also advocated by mcmurry2010banded and mcmurry2015high for stationary time series. Using similar argument as in the proof of Theorem (ref) of this paper and in the proof of Theorem 3 of mcmurry2010banded, it can be shown that $\|\hat \Sigma^{pd}_{n,m}-\Sigma_{n,m}\|_{q/2}=O(r_n)$. Now the arguments given in the proof of Corollary 1 of wu2009banding yield an analogue of Corollary (ref), that is $$ \rho((\hat \Sigma^{pd}_{n,m})^{-1}-\Sigma_{n,m}^{-1})= O_{\mathbb{P}}(r_n). $$ A careful inspection of the proof of Theorem (ref) finally shows that its assertion remains valid, if $\hat \Sigma_{n,m}$ in (ref) is replaced by $\hat \Sigma^{pd}_{n,m}$. }

Implementation and numerical results

\setcounter{equation}{0}

To implement our method we need to choose several tuning parameters: the bandwidths $\tau_n$ and $b_n$ for the local linear estimators of the trend $\mu$ and auto-covariance function $\gamma_k$ and the width $l_n$ of the banded estimator of the covariance matrix $\Sigma_n$. For choosing $\tau_n$, we recommend the Generalized Cross Validation (GCV) method proposed in zhou2010simultaneous.

More precisely, let $\hat \mu^{1:m}(\cdot,\tau), 1\leq i\leq m$ be the local linear estimate of the mean trend defined in (ref) using bandwidth $\tau$, then we choose $\tau_n$ as

align*[align* omitted — 164 chars of source]

where $T^{1:m}_{\tau,ii}$ is the $i_{th}$ diagonal entry of the matrix $$ J^{1:m}_0 \big ( (X^{1:m}(i/n))^\top W^{1:m}_\tau(i/n)X^{1:m}(i/n) \big )^{-1}(X^{1:m} (i/n))^\top W^{1:m}_\tau(i/n), $$ $J^{1:m}_0$ and $X^{1:m}(i/n)$ are $m\times 2$ matrices defined by

eqnarray*[eqnarray* omitted — 273 chars of source]

respectively, and $W_\tau(x)$ is an $m \times m$ diagonal matrix with elements $\big\{ K \big (\frac{x-s/n}{\tau} \big ) \big \}_{s=1, \ldots m}$. The bandwidth $b_n$ for the estimation of the auto-covariance function $\gamma_k$ in (ref) is defined similarly. For example, if $k$ is even, we choose $b_n$ as

align[align omitted — 223 chars of source]

where $\hat {\gamma}_k^{1:m}(i/n,c)$ is the local linear estimator with bandwidth $c$ defined as in (ref) using $m$ observations and $T^{1:m}_{c,ii}$ is defined as in the previous paragraph.

To motivate the choice of the width $l_n$ in the banded estimator of the covariance matrix, note that

align[align omitted — 193 chars of source]

[see Section 4.3 in zhang2012inference], where $\tilde \sigma_k^2=\int_0^{\frac{m}{n}\wedge \frac{n-k}{n}}g^2(t)dt$, and the function $g^2$ is the long-run variance of the locally stationary process $\{\epsilon_{i,n}\epsilon_{i+k,n}\}_{i=1}^{n-k}$. For its estimation we use a statistic proposed by dette2018change, which is defined as follows. Consider the partial sum of lag $k$ $$ ^kS^{1:m}_{r_0,r_1}=\sum_{i=r_0}^{r_1}\hat \epsilon^{1:m}_{i,n}\hat \epsilon^{1:m}_{i+k,n}, $$ where we use the notation $\hat \epsilon^{1:m}_{i,n}=0$ if the index $i$ satisfies $i<1$ or $i>m$. For an integer $b\geq 2$ we introduce the quantities $$ ^k\Delta^{1:m}_{j,b}=\frac{^k S^{1:m}_{j-b+1,j}- {^k S^{1:m}_{j+1,j+b}}}{b}. $$ Finally, we define for $t\in[b/n,(m-b)/n]$

align*[align* omitted — 89 chars of source]

where $$ \omega(t,i)=K\Big(\frac{i/n-t}{b_n}\Big){\Big /}\sum_{ {i}=1}^nK\Big(\frac{i/n-t}{b_n}\Big) $$ and the bandwidth $b_n$ is given by (ref) with $\hat \epsilon^{1:m}_{i-k/2,n}\hat \epsilon^{1:m}_{i+k/2,n}$ there replaced by $\hat \epsilon^{1:m}_{i,n}\hat \epsilon^{1:m}_{i+k,n}$. For $t\in[0,b/n)$ and $t\in((m-b)/n,m/n]$ we define $\hat g^2 (t)=\hat g^2(b/n)$ and $\hat g^2 (t)=\hat g^2 ((m-b)/n)$, respectively. Finally, we propose

align[align omitted — 203 chars of source]

as a data-driven choice of the width $l_n$, where $ \kappa(\alpha)$ is the $\frac{1+(1-\alpha)^{1/(l_1-l_0+1)}}{2}$-quantile of the standard normal distribution and $l_{0}$ and $l_1$ are constants (if the set $\big\{ n^{-1/2}|\sum_{i=1}^n\hat \epsilon^{1:m}_{i,n}\hat \epsilon^{1:m}_{i+l,n}|\geq \kappa({0.01}) \hat \sigma_l, 1_0\leq l\leq l_1\big\}$ is empty we define $l_n=l_0-1$).

Covariance estimation

In this section we investigate the finite sample properties of the estimators (ref) and (ref) for the covariance matrix $\Sigma_n$ of a locally stationary process, where we consider

eqnarray[eqnarray omitted — 134 chars of source]

as mean functions. Recalling the notation $\mathcal{F}_i = ( \ldots , \varepsilon_{i-1},\varepsilon_i ) $ we investigate four different distributions for the errors in model (ref):

description• (a) $\{ \epsilon_{i,n}: i=1,\ldots,n\}$ is a stationary $AR(0.3)$ process with independent standard normal distributed innovations. • (b) $\epsilon_{i,n}=0.8 G(i/n,\mathcal{F}_i)$ where $$ G(t,\mathcal{F}_i)=0.7\sin(2\pi t) G(t,\mathcal{F}_i)+\varepsilon_i $$ and $\{\varepsilon_i\}_{i\in \mathbb Z}$ is a sequence of independent, standardized ($\mathbb E [\varepsilon_i] =0$, Var($\varepsilon_i)=1$) $t$-distributed random variables with six degrees of freedom. • (c) $\epsilon_{i,n}= G(i/n,\mathcal{F}_i)$ where $$ G(t,\mathcal{F}_i)=\frac{1}{6}(\exp(4(t-0.5)^2)+1)\varepsilon_i+0.6(|\varepsilon_{i-1}|-\mathbb{E}(|\varepsilon_{i-1}|)) $$ and $\{\varepsilon_i\}_{i\in \mathbb Z}$ is a sequence of independent standard normal distributed random variables. • (d) $\epsilon_{i,n}= G(i/n,\mathcal{F}_i)$ where $$ G(t,\mathcal{F}_i)=\frac{1}{4}(\cos(\pi t)+2)(\varepsilon_i+0.9\varepsilon_{i-1}-0.6\varepsilon_{i-2}) $$ and $\{\varepsilon_i\}_{i\in \mathbb Z}$ is a sequence of standardized $(\mathbb{E}[\varepsilon_i]=0$, Var$(\varepsilon_i)=1$) independent chi-square distributed random variables with five degrees of freedom.

Note that model (a) defines a stationary process and model (b) defines a locally stationary AR(1) process. Model (c) defines a nonlinear $tvMA(1)$ process. Since the innovations $\varepsilon_i$ in model (c) have a symmetric distribution, the covariance matrix of model (c) is diagonal. Model (d) defines a $tvMA(2)$ process, where only the entries in the diagonal and the first two off diagonals of the covariance matrix do not vanish.

table[table omitted — 1,442 chars of source]

We examine the estimator for covariance matrix $\Sigma_n$ for sample sizes $n=250$, $500$ and $1000$ using $1000$ simulation runs. For the estimation of the width $\l_{n}$ of the band in (ref) we use (ref) with $l_0=1$, $l_1=6$. In each simulation run the tuning parameters ($\tau_n$, $b_n$) are determined as described at the beginning of this section. In Table (ref) and (ref) we display the simulated mean squared error of the spectral loss $\rho (\hat \Sigma_n - \Sigma_n)$ for different estimators $\hat \Sigma_n$, where different mean functions and error processes in model (ref) are considered. In particular we compare the mean corrected estimator (ref) for non-stationary error processes with the mean corrected estimator (ref) which assumes a stationary error process. The numbers in brackets show the standard error of the estimates. We observe that in the stationary model (a) the accuracy of both estimators improve with increasing sample size. Moreover, the estimator (ref) outperforms (ref) because this estimator is constructed for stationary processes. On the other hand, for the dependence structures (b) - (d) corresponding to locally stationary processes the stationary method in (ref) is not consistent and the estimator (ref) shows a substantially superior behaviour.

table[table omitted — 1,431 chars of source]

Prediction

To illustrate the finite sample properties of the estimator proposed in Section (ref) for prediction we examine the mean trend (ref). As error process we consider a locally stationary AR(6) model defined by

equation[equation omitted — 113 chars of source]

where the functions $a_1(t), \ldots , a_6(t)$ are given by

align*[align* omitted — 188 chars of source]

$ \sigma(t)=(1+0.5\sin 2\pi t)^{0.5}$ and $\mathcal B$ is the lag operator on the filter $\mathcal{F}_i$, i.e., $\mathcal B G(t,\mathcal{F}_i)=G(t,\mathcal{F}_{i-1}).$ We consider a standard normal as well as a $\chi^2(6)$ distribution for the errors $\varepsilon_i$ (centered and standardized such that $\mathbb{E}[\varepsilon_i]=0$ Var$(\varepsilon_i)=1$) and examine the mean squared error of the prediction for sample sizes $n=250, n=500,n=1000$. We also compare the new predictor with the methods in roueff2018prediction, kley2019predictive and giraud2015aggregation which were theoretically investigated for centered data. In a first step we used these methods with the residuals $\hat \epsilon^{1:m}_{i,n}$ to obtain a prediction for the de-trended series. In a second step we add to this estimate the value $\hat \mu^{{1:m} }(m/n)$ to obtain the final prediction of $X_{m+1,n}$. Notice that these authors use time-varying AR$(d)$ processes to approximate the time series for prediction without knowing $d$. Since the error process (ref) is a locally AR$(6)$ process, we investigate the performance of the methods proposed by roueff2018prediction, kley2019predictive and giraud2015aggregation for $d=3$, $d=6$ and $d=9$ (note that in the predictor of kley2019predictive $d$ denotes the maximum lag that their algorithm allows). These cases represent the situation of underestimation, correct-estimation and overestimation of $d$. Note that in the cited references there are no rules how to select $d$. Moreover, for the method proposed by kley2019predictive we choose the parameter $\delta$ in their procedure as $0.05$, as a small parameter $\delta$ prefers the choices of a time-varying model to a stationary model.

table[table omitted — 2,021 chars of source]
table[table omitted — 2,074 chars of source]

In Table (ref) and (ref) we present the simulated mean squared error $$ \mathbb{E}[(\hat X^{pred}_{m+1,n} - X_{m+1,n})^2] $$ for the four different prediction methods and different distributions of the innovations. The columns denoted by $t_{pred}=0.5$ and $t_{pred}=1$ correspond to a prediction of $X_{\lfloor n/2 \rfloor +1}$ from on $X_{1,1}, \ldots, X_{\lfloor n/2 \rfloor }$ and a prediction of $X_{n,n}$ from $X_{1,1},\ldots, X_{n-1,n}$, respectively, where {we use $l_0= \lceil \log (m) \rceil $ and $l_1=5+\lceil \log (m) \rceil$ in (ref).} The first row shows the simulated mean squared error of the prediction (ref). With increasing sample size this mean squared error approximates $1$. This corresponds to our theoretical result in Theorem (ref), because we have for the model under consideration $\sigma(0.5)=\sigma(1)=1$. The rows denoted by R-S, G-R-S and K-P-F show the simulated mean squared error for predictors proposed by roueff2018prediction, giraud2015aggregation and kley2019predictive, respectively, with different time lags $d=3,6,9$. In general, the non-stationary predictor (ref) performs better or similar as the alternative methods with different time lag $d$ in all scenarios. Our simulation results also demonstrate that the performance of R-S, G-R-S and K-P-F predictors depend sensitively on the choice of $d$. Finally, the large numbers in R-S predictor is due to the singularity of estimated local covariance matrix. We expect that this can be corrected by using an eigenvalue corrected positive definite covariance matrix estimator similar to (ref).

We also examine the distribution of the prediction error as investigated in Theorem (ref). For this purpose we show in Figure (ref) the QQ plot of prediction errors of the predictors (ref) for standard normal distributed errors and centered and standardized $\chi^2(6)$-distributed errors {in} model (ref), respectively. The model is given by (ref) and the sample sizes is $n=1000$. These results confirm the theoretical findings in Theorem (ref).

figure[figure omitted — 346 chars of source]

Finally, we compare the new predictor (ref) with the methods proposed by roueff2018prediction, giraud2015aggregation and kley2019predictive in a locally stationary MA(6) model defined by

align[align omitted — 111 chars of source]

where the time varying coefficients $a_1, \ldots a_{6}$ and the function $\sigma$ are the same as those defined in the locally stationary AR$(6)$ model (ref), the mean function is given by (ref) and the random variables $\varepsilon_{i}$ are independent standard normal distributed. The results are presented in Table (ref) and we observe similar properties as in the locally stationary AR$(6)$ model (ref). A detailed discussion is omitted for the sake of brevity.

table[table omitted — 2,026 chars of source]

Market indices analysis

In this section we apply our method to predict market indices. Let $p_t$ be the adjusted daily closing value at day $t$, then the log return $r_t$ is defined as

align*[align* omitted — 40 chars of source]

As pointed out by stuaricua2005nonstationarities, the sign of $r_t$ is unpredictable. As a result, these authors proposed to model $r_t$ as

align[align omitted — 62 chars of source]

where $\mu $ and $\sigma $ are time varying functions and $\epsilon_t$ denotes a zero-mean noise process. stuaricua2005nonstationarities used model (ref) to study the non-stationarity of stock returns. In this section we apply the new method to predict $y_t:=\log(|r_t|)$ for the SP500, NASDAQ and Dow Jones Index. We consider data from Dec. $19$, $2016$ to Dec. $17$, $2019$. For SP500, NASDAQ and Dow Jones Index, we delete the log return of Jan. 10, 2017, Nov. 13, 2018 and Nov. 12, 2019 respectively due to their negative infinity values. Therefore the lengths of the series are $752$. We use the new method to predict the market indices at trading days between April. 8, 2019 and Dec. 17, 2019 for SP500 and NASDAQ and at trading days between April. 5, 2019 and Dec. 17, 2019 for Dow Jones Series, respectively, and calculate the empirical mean squared error for these predictions. For the sake of comparison we also apply the methods of roueff2018prediction (R-S), giraud2015aggregation (G-R-S) and kley2019predictive (K-P-F) to the same series. As in the simulation, for fair comparison we perform those algorithms on non-parametrically de-trended data and use the outcome plus $\hat \mu((T-1)/T)$ as the prediction of indices at day $T$. The corresponding results are listed in Table (ref), where we use the different lags $3,6,9$ in the procedures based on autoregressive fitting. We observe that the new prediction method (ref) shows the best performance for all three market indices. For NASDAQ index the method proposed by \ kley2019predictive with $d=9$ shows a similar performance. In general the parameter $d$ for the prediction method proposed by roueff2018prediction, giraud2015aggregation and kley2019predictive is difficult to select, while it has a complicated impact on the predictions when applying those approaches. In Figure (ref) we also plot the prediction error of the different methods for the three market indices. The left panels display $\log|r_t|$, while the right panels show absolute prediction errors of the prediction (ref) and of the predictors proposed by roueff2018prediction (R-S), giraud2015aggregation (G-R-S) and kley2019predictive (K-P-F) for the corresponding parameter $d \in \{ 3,6,9\} $, which achieves the smallest mean squared error.

table[table omitted — 935 chars of source]
figure[figure omitted — 577 chars of source]