Analysis, modeling and estimation through a similarity based approach: an economic signal case study

A key issue in a time varying signal estimation is the problem of identification and extraction of existing underlying patterns. Pattern consistency is not a common rule in certain types of dynamics, particularly those related to economic processes. In that case, it is very difficult to provide an accurate estimation. An approach is described to analyze, capture and model those more complex structured, possibly nonlinear and nonstationary time series. Spectral analysis and digital filter theory are used to break the original series into more coherent and power compatible components. Each resulting component is then treated individually as a totally independent time series. The individual results are combined to provide future estimations for the original signal. An example of a four-week forecast of an oil price time series is presented, leading to a normalized root mean square error of 0.18