Testing for ARCH in the presence of additive outliers

Abstract

In this paper we investigate the properties of the Lagrange Multiplier (LM) test for autoregressive conditional heteroskedasticity ( A R CH) and generalized ARCH (GARCH) in the presence of additive outliers (AO's). We show analytically that boththe asymptotic size and power are adversely a ected if AO's are neglected: the test rejects the null hypothesis of homoskedasticity too often when it is in fact true, while the test has di culty detecting genuine GARCH e ects. Several Monte Carlo experiments show that these phenomena occur in small samples as well. We design and implement a robust test, which has better size and power properties than the conventional test in the presence of AO's. Applications to the French industrial production series and weekly returns of the Spanish peseta/US dollar exchange rate reveal that, sometimes, apparent GARCH e ects may be due to only a small number of outliers and, conversely, t h a t g e n uine GARCH e ects can be masked by outliers.

Product Innovations, Advertising, and Stock Returns

“…We tested for autoregressive conditional heteroskedasticity (ARCH) in the residuals using Engle's Lagrange Multiplier ARCH test (Engle 1982;van Dijk, Franses and Lucas 1999) and fail to reject the null hypotheses of no ARCH (p<.01).…”

Output Growth and its Volatility: The Gold Standard through the Great Moderation

“… analyze 15 post–WWII U.S. macroeconomic time series using the outlier identification procedure based on and find that outliers may prove important for U.S. macroeconomic data and that such aberrant observations may lead to large ARCH test statistics. demonstrate that neglecting additive outliers frequently leads to a rejection of the null hypothesis of homoskedasticity, when it is in fact true. and , however, show another possibility.…”

The Effects of Additive Outliers and Measurement Errors when Testing for Structural Breaks in Variance*

“…Condition A3 ′ may be sufficient when the series are uncorrelated but not independent, as is mostly the case in financial time series. Finally, assumption A4 is a maintained assumption in related studies, such as Franses and Haldrup (1994) and van Dijk et al . (1999).…”