Outlier Detection in Structural Time Series Models: The Indicator Saturation Approach

Abstract

Structural change affects the estimation of economic signals, like the underlying growth rate or the seasonally adjusted series. An important issue, which has attracted a great deal of attention also in the seasonal adjustment literature, is its detection by an expert procedure. The general-to-specific approach to the detection of structural change, currently implemented in Autometrics via indicator saturation, has proven to be both practical and effective in the context of stationary dynamic regression models and unit-root autoregressions. By focusing on impulse-

How Biased Are U.S. Government Forecasts of the Federal Debt?

“…Hendry (1999) proposed IIS as a procedure for testing parameter constancy. See Hendry, Johansen, and Santos (2008), Doornik (2009a), Nielsen (2009, 2013), Hendry and Santos (2010), Ericsson (2011aEricsson ( , 2011bEricsson ( , 2012Ericsson ( , 2016, Ericsson and Reisman (2012), Bergamelli and Urga (2014), Hendry and Pretis (2013), Hendry and Doornik (2014), Castle, Doornik, Hendry, and Pretis (2015), and Marczak and Proietti (2016) for further discussion and recent developments. Impulse indicator saturation uses the zero-one impulse indicator dummies {  } to analyze properties of a model.…”

“…Applications of IIS reflect that wide-ranging ability: see Hendry (1999) on nonconstancy, Johansen and Nielsen (2009) and Marczak and Proietti (2016) on outliers, Hendry and Doornik (2014, Chapter 15.6) on thick-tailed distributions, Hendry and Santos (2010) on heteroscedasticity and super exogeneity, Ericsson (2011b) on omitted variables and regime changes, Castle, Doornik, and Hendry (2012) on multiple breaks, Pretis, Schneider, Smerdon, and Hendry (2016) on "designer" breaks, and Ericsson (2016) on measurement errors. Gamber and Liebner (2017) underscore the benefits of IIS, stating that ".…”

A Data-Cleaning Augmented Kalman Filter for Robust Estimation of State Space Models

“…Hence, tying δ to the PESD accounts for the difficulty of detecting outliers in the case of a high overall variation. A detailed discussion of the choice of the outlier magnitude for structural time series models is provided by Marczak and Proietti (2015), who consider the same settings regarding the model for simulations (BSM) and the values for the variance parameters. In addition to the reference size δ = 7·PESD, also chosen as a reference size in Marczak and Proietti (2015), we consider δ = 14·PESD.…”

“…A detailed discussion of the choice of the outlier magnitude for structural time series models is provided by Marczak and Proietti (2015), who consider the same settings regarding the model for simulations (BSM) and the values for the variance parameters. In addition to the reference size δ = 7·PESD, also chosen as a reference size in Marczak and Proietti (2015), we consider δ = 14·PESD. Increasing the magnitude of δ is motivated by the fact that, in contrast to Marczak and Proietti (2015), the final outlier size is not given by δ but is obtained by scaling δ with a standard normal variable z t .…”

“…In addition to the reference size δ = 7·PESD, also chosen as a reference size in Marczak and Proietti (2015), we consider δ = 14·PESD. Increasing the magnitude of δ is motivated by the fact that, in contrast to Marczak and Proietti (2015), the final outlier size is not given by δ but is obtained by scaling δ with a standard normal variable z t . Since the probability that z t takes on values between −1 and 1 is 68.27%, the final outlier size is in most of the cases smaller than δ.…”