Over the Stability Space of a Multivariate Time Series Journal: Econometrics and Statistics
Roberto Vásquez Martínez, Graciela González Farías, José Ulises Márquez Urbina, Francisco Corona · 2025
Read Over the Stability Space of a Multivariate Time Series Journal: Econometrics and Statistics by Roberto Vásquez Martínez, Graciela González Farías, José Ulises Márquez Urbina, Francisco Corona (2025) 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-2506-22407.txt.
Introduction
Four authors led by Roberto Vasquez Martinez tackle long data series with many linked parts in a 2025 paper. Such series often drift over time, which breaks standard tools. Older work used cointegration to find steady links between series. Here the authors define a broader idea, a stability space, meant to capture the steady parts when strict rules fail.
Two ways to estimate that space are built from common dimension cutting tools, with parts chosen to favor steady behavior. The team compares them with a standard parametric method where it can be used. Tests on made up and real data cover hard cases, such as very wide data, and sparse methods are tried as well.
Catalog record
| CREATOR | Roberto Vásquez Martínez, Graciela González Farías, José Ulises Márquez Urbina, Francisco Corona |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 74 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-22407
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Over the Stability Space of a Multivariate Time Series Journal: Econometrics and Statistics Roberto Vásquez Martínez, Graciela González Farías, José Ulises Márquez Urbina, Francisco Corona Abstract This paper jointly addresses the challenges of non-stationarity and high dimensionality in analysing multivariate time series. Building on the classical concept of cointegration, we introduce a more flexible notion, called stability space, aimed at capturing stationary components in settings where traditional assumptions may not hold. Based on the dimensionality reduction techniques of Partial Least Squares and Principal Component Analysis, we proposed two non-parametric procedures for estimating such a space and a targeted selection of components that prioritises stationarity. We compare these alternatives with the parametric Johansen procedure, when possible. Through simulations and real-data applications, we evaluated the performance of these methodologies across various scenarios, including high-dimensional configurations. Declaration of interests The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. 1 Introduction The study of phenomena or processes exhibiting temporal dependence arises naturally in various fields such as economics, environmental sciences, and epidemiology, for example, in the analysis of Gross Domestic Product per capita ( James et al. 2012 ) , global surface temperature ( Kaufmann et al. 2006 ) , or daily mortality rates ( Bhaskaran et al. 2013 ) . Although statistical methods serve various purposes in this context, many rely –at least to some extent– on the assumption that the underlying stochastic process is stationary in order to establish theoretical and practical properties. This is particularly true in time series analysis. Let \{X_{n},n\in\mathbb{Z}\} denote a time series in \mathbb{R}^{m} . Generally, each of the m components of the series X corresponds to a variable, and these variables interact with one another. Let \{X_{n},n=1,\dots,T\} represent a random sample from the series \{X_{n},n\in\mathbb{Z}\} . To illustrate the importance of the stationarity assumption, consider the h -step-ahead forecasting problem: inferring the distribution of the variables X_{T+1},\dots,X_{T+h} based on the observed data. It is well-known that both Dynamic Factor Models and Vector Autoregressive (VAR) models rely on the assumption of stationarity for forecasting purposes; see Stock and Watson 2002 and Lütkepohl 2005 , respectively. Under the stationarity assumption and employing a VAR model, it is possible to derive the asymptotic distribution of these variables, thereby enabling the construction of both point forecasts and prediction intervals. Similarly, certain time series clustering algorithms ( Maharaj 2000 ) and changepoint detection methods ( Killick et al. 2012 ) also presuppose stationarity. In the absence of stationarity, a different approach is required. In fact, numerous real-world applications involve non-stationary features, as illustrated by analysis of business cycles ( Hamilton 1989 ) , as well as studies in biomedical signal processing and seismology ( Adak 1998 ) . Non-stationarity may arise for various reasons, including deterministic trends, structural changes, seasonal behaviours, or unit roots. In this work, we address non-stationarity arising from unit roots or stochastic trends. The concepts and techniques introduced in this work offer a different approach to this longstanding problem, leveraging dimensionality reduction techniques. Another key aspect of modern data analysis is the prevalence of phenomena involving many interacting variables. In the time series context, this gives rise to the challenge of modelling high-dimensional multivariate time series. High dimensionality typically refers to situations where the number of variables or series, m , is on the order of hundreds or thousands, and often substantially exceeds the length of the series, T . For example, Peña et al. 2019 discusses practical cases in which m is ten times greater than T . Moreover, when m\gg 0 , various estimation challenges emerge. In VAR models, for instance, multicollinearity and parameter instability naturally arise. Consequently, it is of practical interest to represent and decompose the information using a smaller number of uncorrelated latent variables –in other words, to perform dimensionality reduction. A comprehensive summary of this approach in the time series context is provided by Lam et al. 2011 . The techniques introduced therein enable analysis in high-dimensional settings. To tackle the simultaneous challenges of non-stationarity and high dimensionality, we introduce a methodology building upon the concept of cointegration, as originally formulated by Engle and Granger 1987 . In this regard, the methodology for analysing cointegration in multivariate time series proposed by Johansen 1991 is particularly influential, as it allows the estimation of a basis for the so-called cointegration space. However, one of the main limitations of Johansen’s procedure is that it becomes computationally infeasible in high-dimensional systems. Moreover, it is a parametric approach that relies on a set of formal assumptions which may not be plausible in practice. This motivates the introduction of a new concept, referred to as the stability space , which we will discuss throughout this work. There are alternative approaches to Johansen’s method that also target the cointegration context and offer certain advantages, as they are based on non-parametric dimensionality reduction techniques. In this article, we examine the approaches of Muriel et al. 2012 , based on the Partial Least Squares (PLS) algorithm ( Wold 1966 ) , and Harris 1997 , based on Principal Component Analysis (PCA) ( Hotelling 1933 ) . These methods provide a useful starting point, as they have inspired a wide range of interesting extensions and modifications in the literature. Moreover, we argue that the greater flexibility of these approaches may be advantageous for estimating the less restrictive concept of stability space, which, for example, allows the construction of forecasts that are well-behaved in the long run. After applying a dimensionality reduction method, one typically obtains a smaller number of components with which to optimally represent the original information according to some criterion. For example, in PCA, it is common to retain the first two principal components. However, these first two principal components are not guaranteed to yield stationary time series –in fact, as we shall demonstrate in this article, one should often expect the opposite. We therefore propose selecting the first principal components that produce stationary time series, which is more appropriate, given that time series methods are typically applied to these latent variables. This article is organised as follows: Section 2 discusses the concept of cointegration and how it motivates the introduction of the stability space, which is addressed in Section 3 . Section 4 describes the most relevant aspects of the methodologies considered: Johansen’s approach and those based on PLS and PCA. Section 5 presents the theoretical aspects of a novel simulation study designed to evaluate these methodologies and the results of this assessment. This simulation procedure is itself of independent interest, as it enables the generation of time series with a specified cointegration space –or, more accurately, with a given stability space. Finally, in Section 6 we apply the discussed methodologies in economics practical examples, considering both low- and high-dimensional cases. The paper ends with a discussion in Section 7 . The mathematical proofs and the code associated with this work are in the Appendix. 2 Cointegration notion Consider a time series \{X_{n},n\in\mathbb{Z}\} taking values in \mathbb{R}^{m} , such that \mathbb{E}[\lVert X_{n}\rVert^{2}]<\infty . Definition 2.1 The difference operator \Delta is defined as \Delta X_{n}:=X_{n}-X_{n-1}\quad\text{for all }n\in\mathbb{Z}. The series \{X_{n},n\in\mathbb{Z}\} is said