Partial Distance Correlation Screening for High Dimensional Time Series

Kashif Yousuf, Yang FengarXiv 2018 · DOI · arXiv

Abstract

High dimensional time series datasets are becoming increasingly common in various fields such as economics, finance, meteorology, and neuroscience. Given this ubiquity of time series data, it is surprising that very few works on variable screening discuss the time series setting, and even fewer works have developed methods which utilize the unique features of time series data. This paper introduces several model-free screening methods based on the partial distance correlation and developed specifically to deal with time dependent data. Methods are developed both for univariate models, such as nonlinear autoregressive models with exogenous predictors (NARX), and multivariate models such as linear or nonlinear VAR models. Sure screening properties are proved for our methods, which depend on the moment conditions, and the strength of dependence in the response and covariate processes, amongst other factors. Dependence is quantified by functional dependence measures and $\beta$-mixing coefficients, and the results rely on the use of Nagaev and Rosenthal type inequalities for dependent random variables. Finite sample performance of our methods is shown through extensive simulation studies, and we include an application to macroeconomic forecasting.

Introduction

In the era of "Big Data," researchers across diverse disciplines—from finance to neuroscience—frequently encounter the "curse of dimensionality." This challenge arises when the number of potential predictors ($p$) vastly exceeds the number of available observations ($n$). To navigate this, variable screening has become a standard preliminary step, acting as a high-speed filter to eliminate irrelevant "noise" variables before applying more computationally intensive statistical models.

However, traditional screening tools, such as Sure Independence Screening (SIS), were primarily designed for independent and identically distributed (i.i.d.) data. Time series data inherently violates these assumptions; observations are temporally dependent, and the ordered structure of lags provides critical predictive information. This paper addresses this gap by introducing a framework specifically tailored for high-dimensional time series screening using Partial Distance Correlation (PDC).

The Challenge of Temporal Dependence

Standard screening methods often fail in a time series context because they do not account for serial correlation. In many systems, the most recent lags of a response variable are its strongest predictors. If a noise variable happens to share a similar trend or seasonal pattern with the response, traditional methods may flag it as significant—a "false positive" driven by shared autocorrelation rather than a true underlying relationship. Furthermore, time series models often involve complex, nonlinear interactions that linear correlation measures (like Pearson’s) fail to capture.

Methodology: Partial Distance Correlation (PDC)

The authors propose a model-free approach leveraging Distance Correlation (DC), a metric capable of detecting both linear and nonlinear dependencies. To specifically address the time series structure, they introduce two primary algorithms:

  1. PDC-SIS: This method utilizes Partial Distance Correlation to measure the relationship between a predictor and the response while conditioning on specific confounding variables. By conditioning on the history (lags) of the response and the lower-order lags of the predictor, the algorithm ensures that a variable is only selected if it provides additional information beyond what is already contained in the time series' own history.
  2. PDC-SIS+: This is an iterative refinement of the PDC-SIS approach. It identifies the strongest signals at each lag level and adds them to the conditioning set for subsequent higher-order lags. This process uncovers "hidden" predictors that might be masked by more dominant variables in a single-pass screening.

Additionally, the framework extends to Multivariate Time Series (VAR models). This allows for the simultaneous screening of groups of variables, significantly reducing the computational overhead when analyzing hundreds of interacting series.

Theoretical Guarantees

A critical contribution of this work is the mathematical proof of the "Sure Screening" property. The authors demonstrate that as the sample size increases, the probability of the PDC-based methods retaining all truly relevant predictors approaches one. These proofs account for the complexities of temporal dependence using functional dependence measures and $\beta$-mixing coefficients, relying on Nagaev and Rosenthal type inequalities to establish bounds even under non-i.i.d. conditions.

Empirical Evaluation

Simulation Studies

The methods were tested against synthetic datasets generated from nonlinear autoregressive models (NARX) and Vector Autoregressive (VAR) models. The simulations compared PDC-SIS against standard SIS and basic Distance Correlation under various conditions, including Gaussian and heavy-tailed ($t$-distribution) errors.

In complex nonlinear scenarios, PDC-SIS demonstrated superior accuracy. For instance, in one simulation, PDC-SIS required a model size of only 61 variables to capture all true predictors, whereas standard methods required nearly 500 variables to achieve the same result. The method also proved robust to outliers, maintaining performance in heavy-tailed data environments common in finance.

Macroeconomic Forecasting

The practical utility of the framework was demonstrated using a real-world dataset of 132 monthly macroeconomic indicators (1984–2011). The task was to forecast Real Personal Income (RPI) and Employment (EMP) six months ahead. The PDC-based screening methods outperformed traditional benchmarks, such as AR(4) models and high-dimensional techniques like the Adaptive Lasso, in terms of Mean Squared Error (MSE).

Conclusion and Future Directions

This paper provides a robust, model-free framework for variable screening in high-dimensional time series. By conditioning on temporal history through Partial Distance Correlation, the proposed methods effectively "de-noise" the screening process and capture complex nonlinear dependencies that traditional tools miss.

While the current framework is highly effective, it assumes the time series is stationary. Future research may explore adaptations for non-stationary processes or structural breaks. Additionally, while the method handles lags effectively, developing data-driven techniques for selecting the maximum lag length and addressing long-range dependence remain promising avenues for further study.

Summary updated

Abstract

High dimensional time series datasets are becoming increasingly common in various fields such as economics, finance, meteorology, and neuroscience. Given this ubiquity of time series data, it is surprising that very few works on variable screening discuss the time series setting, and even fewer works have developed methods which utilize the unique features of time series data. This paper introduces several model free screening methods based on the partial distance correlation and developed specifically to deal with time dependent data. Methods are developed both for univariate models, such as nonlinear autoregressive models with exogenous predictors (NARX), and multivariate models such as linear or nonlinear VAR models. Sure screening properties are proved for our methods, which depend on the moment conditions, and the strength of dependence in the response and covariate processes, amongst other factors. Dependence is quantified by functional dependence measures (Wu [Proc. Natl. Acad. Sci. USA 102 (2005) 14150-14154]) and $β$-mixing coefficients, and the results rely on the use of Nagaev and Rosenthal type inequalities for dependent random variables. Finite sample performance of our methods is shown through extensive simulation studies, and we include an application to macroeconomic forecasting.
BibTeX
@misc{yousuf2018partial,
  author = {Kashif Yousuf and Yang Feng},
  title = {{Partial Distance Correlation Screening for High Dimensional Time Series}},
  year = {2018},
  eprint = {1802.09116},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1802.09116},
}
Figure