This study tests whether changes in the short-term interest rate can best be modelled in a non-linear fashion. We argue that there are good theoretical and empirical reasons for adopting this strategy. Using monthly data from several industrialized countries, namely Canada, Germany, Sweden, Switzerland, UK, and US, we show that the short-term interest rate movements are better explained, usually via the exponential smooth transition autoregression (ESTR). Unlike the existing literature on non-linear estimation, we consider a number of candidates for the transition variable. These include: an error correction term, estimated from an underlying cointegrating relationship predicted by the expectations hypothesis, the US term spread, the domestic spread, inflation and output growth forecasts, and deviations from an inflation target in the case of Canada, the UK and Sweden. The sample spans the period from 1960-1998. We reject linearity in the behavior of short-term interest rate changes and instead find support for a non-linear model with the (lagged) domestic spread as the transition variable. However, other more economically meaningful alternatives perform just as well. For example, in the case of the inflation targeting countries in our sample, the most appropriate transition variable can be the deviation from the publicly announced inflation target. We supplement estimates with extensive diagnostic testing of the non-linear model to ensure that we can reject the linear alternative with reasonable confidence. We believe that changes in central bank policies, and in the reaction of market participants over time to such changes, argue in favor of the non-linear estimation approach. We also argue that any model of the term structure estimated over a fairly long span of time necessitates resort to non-linear estimation methods.
