Are levels effects important in out-of-sample performance of short rate models? Sandy Suardi ⁎ Department of Economics and Finance, La Trobe University, Bundoora, VIC 3086, Australia Received 13 March 2005; received in revised form 24 April 2007; accepted 21 June 2007 Available online 30 June 2007 Abstract This paper derives short-term interest rate volatility forecasts from various interest rate models. While models that specify both GARCH and levels effects are superior in their forecasts accuracy, they systematically under predict interest rate volatility more frequently than simple short rate models. © 2007 Elsevier B.V. All rights reserved. Keywords: Level effects; Volatility; Forecast evaluation; Interest rates JEL classification: C53; E47; E43 1. Introduction Theoretical and empirical short rate models commonly parameterise levels effect where short-rate volatility peaks (diminishes) when the short-rate level rises (falls), see Brennan and Schwartz (1979), Dothan (1978),Cox et al. (1985), and Chan et al. (1992) (hereafter, CKLS). In addition, short rate volatility exhibits serial dependence that is modeled with a generalised autoregressive conditional heteroskedastic (GARCH) specification, see Longstaff and Schwartz (1992), Brenner et al. (1996) (hereafter, BHK) and Koedijk et al. (1997). While useful insights into the volatility of short rate are drawn from model fitting, few papers have examined the importance of capturing the ‘stylised facts’ of short rates for forecasting. The aim of this paper is to examine the importance of levels effect in improving the short rate model out-of- sample variance forecasts. To this end, we estimate different classes of conditional volatility models that capture the salient features of short rates and compare their out-of-sample variance forecast accuracy across different forecast horizon. Both symmetric and asymmetric error statistics are employed Available online at www.sciencedirect.com Economics Letters 99 (2008) 181 – 184 www.elsevier.com/locate/econbase ⁎ Tel.: +613 9479 2754; fax: +613 9479 1654. E-mail address: s.suardi@latrobe.edu.au. Table 1 Summary statistics and diagnostic tests of Dr t Countries US AUS Mean - 0.0014 - 0.0022 Variance 0.0665 0.0930 Skewness - 0.5434 - 0.1327 Kurtosis 20.9800 15.8088 Jarque–Berra χ 2 (2) 27,390.01 15,624.30 [0.0000] [0.0000] Diagnostic tests ARCH(5) 61.0037 50.8880 [0.0000] [0.0000] Level effect test LM(δ ⁎ ) δ ⁎ = 0.0 14.6480 6.4610 [0.0001] [0.0110] δ ⁎ = 0.5 17.0109 7.4235 [0.0002] [0.0244] δ ⁎ = 1.0 17.1914 7.1278 [0.0001] [0.0283] δ ⁎ = 1.5 18.6506 6.9125 [0.0001] [0.0315] Notes: Figures reported in [.] are p-values. The level effect test statistic tests for the null of no levels effect. 0165-1765/$ - see front matter © 2007 Elsevier B.V. All rights reserved. doi:10.1016/j.econlet.2007.06.023