Abstract
Several recent studies have documented structural shifts in the stochastic process of the ex post (or realized) real interest rate (see for example Huizinga and Mishkin (1986), Antoncic (1986), Garcia and Perron (1994), and Evans and Lewis (1995)). This paper investigates sources of shifts in real rates by incorporating systematic time variation in the parameters and variance shifts in the equation specified by Huizinga and Mishkin (1986) to predict the ex ante real rate. Huizinga and Mishkin (1986) use a general specification in which the ex ante real rate depends on the nominal interest rate, the inflation rate and a supply shock variable. This approach permits us to simultaneously model two types of shifts in the stochastic process of real rates: 1) shifts in the coefficients of the relationship between the ex ante real rate and the variables used in the prediction equation and 2) unconditional shifts in the variance of the stochastic process. The model is estimated using Kim's (1993,1994) methodology combining dynamic linear models with Markov switching heteroscedasticity.
Using monthly data for the period between January 1961 and January 1991, I find that the random walk model used in Antoncic (1986) is rejected in favor of the general time-varying parameter model that includes the explanatory variables. The results are broadly consistent with the Garcia and Perron (1994) study in which the ex ante real rate is characterized as a three-state Markov switching model. However, the results from our longer sample indicate that the mean and variability of the ex ante real rate change again after 1986 where the Garcia and Perron (1994) sample ends. This highlights the importance of modeling continual change in the ex ante real rate in terms of other economic variables rather than use a statistical characterization that only permits a limited number of discrete jumps in the mean of the process. I examine the contribution of each variable to the mean of the ex ante real rate using the estimated coefficients at each time period. Consistent with the Fisher hypothesis, I find that the nominal interest rate comprises the largest component of the mean of the ex ante real rate over most of the sample period and that this relationship becomes stronger after 1981. The overall predictive ability of the variables is examined using a period-by-period Bayes factor that compares the general model to the random walk model. Interestingly, I find that the predictability of the ex ante real rate diminishes considerably after 1986. Finally, the estimates of the model are used to infer a time-series for expected inflation. The estimates show that although expectations of inflation decreased in 1980, they did not decrease sufficiently which kept nominal rates unusually high during this period.