Economic activity and recession probabilities: information content and predictive power of the term spread in Italy

Abstract

The aim of the present article is to examine the information content of the Italian term spread as for real economic growth rates and recession probabilities and to test its predictive power in forecasting regime probabilities. To this end the relationship between the term spread and economic growth rates is modelled as a nonlinear one and specifically the Logistic Smooth Transition model is used, while a probit model is implemented to forecast recession probabilities. Specific to this article is the use of the OECD business cycle chronology, which was never used before to this end for the Italian case. Overall evidence supports the informative content of the spread in Italy over the whole period (1984–2005) although results are more satisfactory as from 1992. In particular, recession forecasts are generally better than those obtained with other chronologies previously adopted for the Italian case (ISAE and ECRI).

Notes

¹For example, in Pederzoli and Torricelli (2005) regime predictions are used to estimate default probabilities and then, based on a forward-looking approach, capital requirements are calculated within the Basel II framework.

²See e.g. Stock and Watson (2003) for an extensive survey of this literature.

³The same Markov-Switching framework is used in many other papers but with different aims: e.g. Vazquez (2004) to investigate the relationship between the term spread and the short rate changes, Kim and Nelson (1999) to predict business cycle turning points of US business cycle.

⁴The smooth transition (STR) models were first used by Terasvirta in seminal works, basically aimed to find the best specification for nonlinear time-series. As an example, in Terasvirta and Anderson (1992) smooth transition autoregressive (STAR) models are used to describe various time-series representing business cycles, such as production and unemployment. Similarly, Terasvirta (1995) compares the fit of the annual per capita GNP to the logistic and the exponential STAR model.

⁵Bec et al. (2002) find that the empirical description of monetary policy by linear Taylor rules sensibly improves using a STR form.

⁶EH can be tested in different ways ranging from simple regressions to cointegration tests (e.g. see Campbell and Shiller 1991; Boero and Torricelli 2002; Sarno et al., 2005; Kalev and Inder, 2006). Here, a Johansen's procedure has been implemented on interest rates prior to all other analyses. Evidence of cointegration and thus of the EH validity in Italy was found. Detailed results for this analysis are available upon request.

⁷See among others Galbraith and Tkacz (2000) and Venetis et al. (2003).

⁸Along with the spread, Venetis et al. (2003) consider several other variables as potential transition variables, such as past growth rates in aggregate economic activity, quarterly output-gap and time. However, as the null of linearity is rejected using all the variables and ‘the strongest rejections correspond to the spread […]’, they ‘finally retain the lagged spread as the transition variable’.

⁹A logit model could alternatively be used (as in Sensier et al., 2004). In this article a logit model was estimated on the same data set with similar results and hence it is not presented.

¹⁰Data source: Datastream.

¹¹Business cycle dating is not the aim of this article, but it is a very important issue which has fostered a specific literature also for the case of Italy: see Otranto (2005) and Bruno and Otranto (2004).

¹²Different chronologies may be associated to different business cycle dynamics in terms of possible asymmetries. An investigation of the symmetric vs. asymmetric nature of the business cycle goes beyond the scope of this article, but a renewed interest in the issue is present in the literature (e.g. Stanca, 1999; Andreano and Savio, 2002; Peirò, 2004).

¹³Detailed results are available upon request.

¹⁴This procedure is in line with Venetis et al. (2003) and could in principle lead to inconsistent estimates; however, provided that  γ is sufficiently large, the bias is practically negligible.

¹⁵Detailed results are available upon request.

¹⁶See www.oecd.org for additional information.

¹⁷See for instance Granger and Teräsvirta (1993) and Clements et al. (2004).

¹⁸Recall that the OECD chronology reports also minor cycles and thus in our sample it turns out that 136 periods (out of 260) are classified as recessions.