USING VARIATIONS TO THE TARGETED PREDICTORS APPROACH FOR FORECASTING CANADIAN GDP WITH AND WITHOUT US INDICATORS.

Abstract

In recent years, Data-driven forecasting models are becoming increasingly prominent in macroeconomic forecasting. This thesis investigates the usefulness of di_erent variations to the Targeted Predictors approach for forecasting Canadian GDP and compares their forecasting performance with conventional dimension reduction and shrinkage methods. In addition, this thesis looks at the potential improvement in forecasting accuracy when US indicators are added to the national dataset. Compared to traditional benchmarks, factor models and penalized regressions, the Targeted Predictors approach yields comparable, if not, substantially lower RMSE for 1-month, 3-months and 6-months forecasts. Furthermore, I _nd that the US dataset contains substantial information content concerning Canada's business cycle. Even without preselection, forecasting using US macroeconomic indicators alongside the national dataset generally improve the forecasting performance of factor-based models for 1-month, 3-months and 6-months forecasts. The accuracy gain is particularly large when penalized regression is utilized to screen out uninformative predictors prior to estimating factors.