Title: Real-time signal extraction of vector time series via multivariate direct filter analysis
Authors: Tucker McElroy - Census Bureau (United States) [presenting]
Marc Wildi - Zurich University (Switzerland)
Abstract: Real-time signal extraction for multivariate time series attempts to utilize information from related series in order to extract trends, cycles, or other patterns of interest using only present and past observations. When signal content is obfuscated by noise in the series of interest, but is more salient in related time series, a substantial improvement in extraction results can be expected. Multivariate direct filter analysis (MDFA) avoids using model-based concurrent filters, being predicated instead on using a frequency domain characterization of multivariate time series. We explicitly show how nonstationary effects, level constraints, and time shift constraints can be accounted for in the MDFA filter. While model-based frameworks can be replicated by the MDFA methodology, more complicated structures (e.g., nonlinear models) can also be entertained, thereby illustrating the flexibility of these nonparametric techniques. We demonstrate the power of the new methods on construction and employment data.