Forecasting economic and financial variables with factor models: the case of Turkey

In this thesis, industrial production growth, core inflation and change in the stock market index are forecast using a large number of domestic and international indicators. Two methods are employed to deal with the curse of dimensionality problem stemming from the availability of ever growing data sets: factor models and forecast combination. Determining the best performing models requires a comprehensive analysis of the sensitivity of the forecast performance of factor models to various modelling choice. In this respect, effects of factor extraction method, number of factors, data aggregation level and forecast equation type on the forecasting performance are analyzed. Effect of using certain data blocks such as European Union variables and interest rates on the forecasting performance is evaluated as well. Out-of-sample forecasting exercise is conducted for two consecutive periods to assess the stability of the forecasting performance. Results show that best performing specifications change with the type of the variable that one wants to forecast, with forecast horizon and with the sample that is used to evaluate the out-of-sample forecasting performance. Factor models perform better than combination of the bi-variate forecasts. Comparing models with alternative specifications shows that effect of modelling choices are not mutually independent. Hence, it is concluded that there is no "one size fits all approach" in forecasting with factor models. Thus, using a dynamic approach to continuously evaluate models from different dimensions is important in the forecasting process. Key Words: Forecasting, Factor models, Principal components.