UNIVERSITÀ DELLA CALABRIA
Dipartimento di Economia e Statistica
Ponte Pietro Bucci, Cubi 0/C-1/C
87036 Arcavacata di Rende (Cosenza)
Italy
http://www.ecostat.unical.it/
Working Paper n. 05 - 2009
CLASSIFICATION OF SHORT TIME SERIES
Prof. Agostino Tarsitano
Department of Economics and Statistics
University of Calabria
Ponte Pietro Bucci, Cubo 1/C
Tel.: +39 0984 492465
Fax: +39 0984 492421
e-mail: agotar@unical.it
Febbraio 2009
Pubblicazione depositata ai sensi della L. 106 del 15-4-2004 e del DPR 252 del 3-5-2006
Classification of short time series
Agostino Tarsitano
Dipartimento di Economia e Statistica
Università della Calabria
agotar@unical.it
Via Pietro Bucci, cubo 1C, Rende (CS) - Italy
Tel.: +39-0984-492465, Fax:+39-0984-492421
february 2009
Classification of short time series
Abstract: Many time series are of short duration because data acquisition has,
of necessity, proceeded for but a brief term. Such data have previously often been
analyzed by methods that either do not explicitly take into account time related
changes or that are designed for long time series. In this paper, we consider sev-
eral ways of assigning a dissimilarity between univariate time series in short term
behavior. In particular, we have defined a measure that works irrespective of diffe-
rent baselines and scaling factors and its effectiveness has been evaluated on real
and synthetic data sets.
Keywords: Time trajectories; Distances; PAM clustering; Representative trends.
1 Introduction
Short time series may be all there is available when data are acquired by an in-
frequent survey due to experimental factors or high costs. For instance, many
economic series are internally comparable for very few periods and statistical es-
timates from such short series tend to be biased. The same is true for micro array
data since technical equipment and methods of measurement change from time to
time. This type of data is obviously undersampled, and some important features
of the temporal pattern can be