Abstract:
Inflation data may exhibit structural instability that might have been engineered by informative macroeconomic variables. Failure to include this relevant economic information in the statistical modelling procedure may produce disingenuous information leading to wrong conclusion. In view of this, this thesis proposed Bayesian-Gaussian process regression, GPR methods based on compound covariance function for modelling inflation in Ghana. The approach model inflation as a mean zero Gaussian process in terms of the observation time with a compound covariance function designed to account for the short-, medium-, and long-term structural characteristics of the inflation process. Macroeconomic variables that drive inflation are incorporated into the model via the covariance using moment-based statistics as alternative macroeconomic predictors. The moment-based macroeconomic predictors serve as transformed predictors and were built based on the existing interrelationships among the variables such that they allow automatic control of the interrelationships, autocorrelation and outliers. This allows Bayesian GPR to be applied to macroeconomic data in which there exist interrelationships. MCMC inference methods were built for the developed GPR models and experimented using real macroeconomic data from Bank of Ghana (BoG) website. Results show that Bayesian GPR models with moment-based macroeconomic predictors outperform their original data predictors in fitting inflation on food and non- food data.
