Puzzling Integratedness of Interest Rates: A Case for Nonparametric Threshold Cointegration?

It is commonly found in empirical studies that nominal interest rates contain a unit root, implying that these variables have a permanent memory. One of the characteristics of a nonstationary time series is that it has no tendency to return to its mean values, meaning that the series is trending and can become arbitrarily high or even negative. With respect to interest rates this seems contrary to economic intuition and stylised facts. In this paper we question the reliability of the unit root hypothesis alone to describe the time series behaviour of interest rates and try another tack, i.e. threshold cointegration, applying nonparametric methods. Three different data sets representing two different long-run relations are examined. Two are related to the yield spread in the American and the German market while the third is akin to the relationship between U.S. and Eurodollar interest rates. Overall, our empirical evidence corroborates the existence of threshold cointegration in these data sets. Furthermore, using Monte Carlo simulations we show that undetected threshold cointegration leads to overestimate the error-correction-coefficients in error correction models. With respect to empirical modelling this implies the wide-spread view that the longer the observation period the better suited are cointegration techniques has to be, at least partly, revised for variables such as interest rates or inflation rates.