Expected correlation in time-series analysis.

Abstract

Time-series analysis often involves the characterization of order or predictability, qualities that are related to internal structure and autocorrelation. Investigating a recently proposed algorithm for solving a density prediction task, we demonstrate that if the same system can be viewed on multiple time scales, there is an inevitable degree of expected order and predictability that increases as the system size grows. In particular, we introduce bounds on the expected second-order structure function and autocorrelation function of a time series where multiple observation scales are available, and conclude with a lower bound on the expected correlation time. Such a lower bound shows that there is an inevitable degree of correlation induced when time-series data is aggregated, quantifying a previously overlooked source of bias towards high correlations.