Threshold cointegration and threshold dynamics

This study utilized monthly averages of daily rates for the 10-year constant maturity Treasury note, the Ibbotson Bond Index with maturity of 20-year Treasury Index, and Moody's Aaa and Baa seasoned bond indices to investigate the threshold behavior of interest rates pairs. The data covered the period from January 1960 to December 1997 with a total of 456 observations for each variable. Three (Lo-Zivot 2001, Hansen-Seo 2002, and Enders-Siklos 2001) different non-linear, discontinuous, asymmetric time-series econometric alternatives were applied to investigate the dynamics of the four interest rates pairs. Forecasting accuracy evaluation was utilized for model evaluation by applying one-step-ahead up to six-step-ahead forecasts.;Among the findings, it was ascertained that interest spreads are stationary, yet the speeds of adjustment are asymmetric. In a bivariate setting, all of the interest rates pairs followed the threshold cointegration behavior. All the interest rates pairs were shown to be threshold cointegrated. In general, the adjustment speeds were asymmetric and, especially, the threshold estimates were asymmetric in a three-regime environment.;Long run equilibrium relationships existed between Moody's corporate bond indices and Treasury note and Ibbotson bond index. In general, for a one percent increase in Treasury rates (either Treasury note or Ibbotson index), in the long run, it will generate a more than one percent increase in corporate bond indices (Aaa or Baa). Furthermore, the Baa bond index was shown to have a greater sensitivity to interest rate changes than the Aaa bond index.;For the model evaluation side, one-step-ahead forecast to six-step-ahead forecast performance evaluations were conducted for the threshold cointegration models and the counterpart of the linear cointegration models. The results showed that no one particular threshold cointegration model dictated the overall forecasting accuracy. For different interest rates pairs under consideration, different threshold cointegration models offered a better fit. Moreover, all of the linear cointegration models performed relatively less accurate than the threshold cointegration models, which reinforce the empirical applications of the threshold cointegration models.