Portal de Periódicos da CAPES

Reliability of the Automatic Identification of ARIMA Models in Program TRAMO

2014; Springer Nature (Netherlands); Linguagem: Inglês

10.1007/978-3-319-03122-4_7

2214-7977

Agustı́n Maravall, Roberto López-Pavón, Domingo Pérez-Cañete,

Statistical and numerical algorithms

In so far that—as Hawking and Mlodinow state—"there can be no model-independent test of reality," time series analysis applied to large sets of series needs an automatic model identification procedure, and seasonal adjustment should not be an exception. In fact, the so-called ARIMA model-based seasonal adjustment method (as enforced in programs TRAMO and SEATS) is at present widely used throughout the world by data producers and analysts. The paper analyzes the results of the automatic identification of ARIMA models of program TRAMO. Specifically, the question addressed is the following. Given that many ARIMA models are possible, how likely is it that (default) use of TRAMO yields a satisfactory result? Important requirements are proper detection of seasonality, of non-stationarity (i.e., of the proper combination of unit autoregressive roots), and of the stationary ARMA structure, and eventual identification of either the correct model, or a relatively close one that provides zero-mean normally identically independently distributed residuals and good out-of-sample forecasts. A comparison with the default AMI procedure in the present X12-ARIMA and DEMETRA+ programs (based on older versions of TRAMO) is made.The simulation exercise shows a satisfactory performance of the default automatic TRAMO procedure applied to very large sets of series; certainly, it can also provide good benchmark or starting point when a careful manual identification is intended.