Forecasting inflation using the term structure and MARS

Abstract

This study revisits the issue of forecasting changes in inflation using non-linear non-parametric methods. The results indicate the presence of threshold effects in the relationship between the information in the term structure and changes in the rate of inflation.

Notes

This section draws heavily on work presented by Sephton (2001).

The M-period inflation rate is defined as log (P _{t  + 1}/P _t )*(1200/M) where M takes on values 3, 6, 9, 12, 36, 60, and 120. The yield spreads and inflation changes always involve the difference between the long run and the short run. Hence (

) denotes the difference between the yields on 10 year government bonds and three month treasury bills.

The ACE approach to modelling finds the non-linear transformation of the predictors which maximizes the correlation between the dependent variable and the transformed predictors. A plot of the transformed series against the dependent variable is sometimes helpful in identifying a functional form to be used in parametric modelling. Hallman (1990) and Granger and Hallman (1991) employed ACE to examine non-linear cointegration.

Since there is only one predictor in this model there is no question of which variables to allow to interact. Subsequent work incorporating additional factors into an inflation forecasting equation may demonstrate the sensitivity of the MARS algorithm to the degree of variable interaction allowed.