On a simultaneous parameter inference and missing data imputation for nonstationary autoregressive models

TL;DR

This paper introduces an extended Finite Element Methodology for Vector Auto-Regressive models that simultaneously estimates parameters and reconstructs missing data in nonstationary time series, overcoming traditional assumptions.

Contribution

It extends FEM-VARX to handle missing data, enabling joint parameter inference and data imputation without relying on Gaussian or stationarity assumptions.

Findings

Outperforms standard methods on test cases

Provides a unified framework for estimation and imputation

Available as open-source software

Abstract

This work addresses the problem of missing data in time-series analysis focusing on (a) estimation of model parameters in the presence of missing data and (b) reconstruction of missing data. Standard approaches used to solve these problems like the maximum likelihood estimation or the Bayesian inference rely on a priori assumptions like the Gaussian or stationary behavior of missing data and might lead to biased results where these assumptions are unfulfilled. In order to go beyond, we extend the Finite Element Methodology (FEM) for Vector Auto-Regressive models with eXogenous factors and bounded variation of the model parameters (FEM-VARX) towards handling the missing data problem. The presented approach estimates the model parameters and reconstructs the missing data in the considered time series and in the involved exogenous factors, simultaneously. The resulting computational framework was compared to the state-of-art methodologies on a set of test-cases and is available as open-source software.