Imputation of Multivariate Time Series Data - Performance Benchmarks for Multiple Imputation and Spectral Techniques

Missing data is a problem appearing ubiquitously across many fields and needs to be dealt with systematically. For multivariate time series data imputation can be a challenging problem. We consider the particular case of credit default swap time series, where missing data can pose a considerable problem preventing important value at risk estimates. We present a new framework to structure and classify missingness patterns, and generate suitable realistic test sets. We then benchmark the performance of a number of state-of-the-art imputation techniques, both stochastic multiple imputation (MI) approaches and deterministic spectral decomposition techniques. We demonstrate that for the missingness patterns under consideration, the MI package Amelia based on the expectation maximisation algorithm performs most robustly and reliably, however, other techniques like multiple singular spectral analysis can also perform well. Our results can serve as a valuable guideline for researchers and practicioners working with incomplete multivariate time series