Exact Modeling of a Finite Time Series

From MaRDI portal





The problem of exact data modelling is considered. The central issue is the exact representation of a finite time series by means of a (deterministic) autoregressive system. The modelling objectives are to infer from the data as many laws as possible, and to accept laws only if they are sufficiently supported by the data. These intuitively formulated objectives are formalised by defining the mathematical concepts of falsifiability (or simplicity) and corroboration. The main result of the paper is the introduction of an identification procedure which takes these two concepts into account. The structure of the paper is as follows. A class of (linear, shift invariant) finite time systems is defined and analysed. For this class of systems the concepts of simplicity and corroboration are defined. Various properties (monotonicity, shift invariance, linearity, truthfulness and prudence) of deterministic identification procedures are introduced. These properties concern the effect of an increase in the amount of information (length of the observed time series or additional a priori information) and the question whether systems are identified which give a faithful description of the data generating system. The case of univariate time series is investigated in detail. It is shown that the well-known procedure of partial realization theory has undesirable properties, as it takes into account only the concept of simplicity and not that of corroboration. By investigating and formalizing the concepts of remarkability and compatibility of unfalsified laws, a procedure is defined which has desirable properties. Roughly stated this procedure identifies a system with the following properties: (i) it is unfalsified by the data; (ii) it is corroborated by the data; (iii) it is the simplest (i.e., most easily falsifiable) system satisfying (i) and (ii). This procedure also gives a new solution to the partial realization problem.











This page was built for publication: Exact Modeling of a Finite Time Series

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3783890)