Inflation prediction in USA using VAR models

This thesis focuses on forecasting inflation in the US. The aim of the thesis is to predict short-term inflation using VAR(p) models. Models with different numbers of lags are tested and compared using multiple parameter estimation methods, including classical OLS, LAD, and machine learning methods, specifically ridge regression. A quarterly point prediction is generated based on the model with the lowest MSE across eleven shifting datasets. For this purpose, a forward validation resampling technique is employed, which estimates the mean out-of-sample error. Apart from MSE, the models are also evaluated based on MAE and information criteria (AIC and BIC), but this time only using a single test and training set. In the setting of one training dataset, the parameters of the models are also estimated by a Bayesian regression employing the Minnesota prior distribution. The theoretical section is devoted to VAR models and parameter estimation methods. Among other things, the equivalence between Bayesian regression and ridge regression under specific conditions is highlighted. Besides VAR models and parameter estimation, attention is paid to economic theory, inflation, and the interrelationships among indicators (variables of the models). Alongside quarterly inflation, the selected time series are quarterly changes in energy prices, quarterly changes in GDP, and the differenced one-year real interest rates. Ultimately, the VAR(1) model estimated by LAD is identified as the optimal model, nowcasting a 0.78% quarterly inflation. The results of the tested VAR(p) models suggest that inflation is less predictable during economic recessions, regardless of the estimation method or lag applied.