Browsing by Author "Raikamo, Joackim"

Producing timely information regarding the current and future state of the economy is important for the practice of economic policy: the delay between the implementation of policy measures and the emergence of their effects is typically considerable, which creates a need to anticipate developments in macroeconomic variables. The producer price index is one such variable: producer price indices are used to track changes in the general price level of goods produced within an economy from the point-of-view of producers, which makes them prominent indicators of inflationary pressures and business cycle conditions. The principal objective of this thesis is to investigate whether the Finnish Producer Price Index for Manufactured Goods could be reliably forecasted in the short run using large sets of external predictors. Increasing the number of predictors exposes standard forecasting methods to inaccuracies and makes their application outright infeasible once the number of variables exceeds the number of observations available for the estimation of the forecasting model. Various alternative methods have been proposed to counter this issue. This thesis provides a broad overview of these methods as well as other relevant issues pertaining to the forecasting macroeconomic variables. Given that no single framework has proven to dominate others in practical applications, a selection of methods has been chosen for the empirical section of this thesis. These methods represent two different approaches to high-dimensional forecasting: dynamic factor models and penalized regressions. The effectiveness of dynamic factor models is based on the assumption that relevant information contained in high-dimensional data can be summarized using only relatively few underlying factors, the estimates of which can, in turn, be used for forecasting. The solution offered by penalized regressions, on the other hand, is based on striking a balance between the bias and variance of the forecasts. Out of the broader class of penalized methods, four different variations will be utilized in this thesis: the Ridge, Lasso, Elastic Net, and Adaptive Lasso. The empirical performance of the methods will be assessed by conducting a simulated out-of-sample forecasting experiment, in which a series of consecutive forecasts are estimated for the target variable using historical data. These forecasts are, in turn, compared to their realized counterparts. The objective of the experimental arrangement is to produce representative information regarding the empirical accuracy of the respective forecasting models by emulating circumstances faced in real-time forecasting: only information that would have been available at the time is used to produce each forecast. The set of predictors used in the experiment is composed of monthly economic time series collected from a variety of sources. Based on the forecasting experiment, the benefit of the high-dimensional models in terms of average forecasting accuracy turns out to be only marginal in comparison to a univariate autoregressive benchmark at the one-, two-, and three-month horizons. Moreover, the differences among the respective high-dimensional methods are found to be insignificant. On the other hand, more favorable results are achieved by using relatively timely market-based variables to predict the concurrent rather than strictly future values of the index. In this case, the penalized models perform particularly well. The results indicate that leveraging the advantage in publication lag enjoyed by external predictors for the purpose of contemporaneous prediction, or nowcasting, could represent the most potential for predicting the producer price index.