Forecasting Macroeconomic Variables Using Several Predictors
Authors
Ypkendanz, Camilla Lund ; Hæstrup, Lærke
Term
4. term
Education
Publication year
2016
Submitted on
2016-06-10
Pages
92
Abstract
Specialet undersøger, hvordan man bedst kan forudsige centrale makroøkonomiske variabler, når der er mange potentielle prædiktorer at vælge imellem. Vi fokuserer på BNP, arbejdsløshed og inflation og sammenligner fire tilgangsvinkler til modellering og prognose: dynamiske faktormodeller med principal komponent analyse, shrinkage-metoder, VAR-modeller og Bayesian model averaging. Prognosepræcisionen vurderes ved hjælp af Diebold–Mariano-testen mod en simpel AR(p)-benchmark. I den empiriske analyse finder vi, at en factor-augmented VAR (FAVAR) giver de mest præcise prognoser for BNP og inflationsraten, mens en Lasso-VAR-model er bedst for arbejdsløsheden. Dog indikerer Diebold–Mariano-testen, at disse mere komplekse modeller ikke producerer statistisk mere præcise prognoser end den enkle AR-benchmark. Endelig undersøger vi, hvilke prædiktorer der udvælges på tværs af modellerne; ingen prædiktorer er universelle, men inden for hver variabel går visse prædiktorer igen.
This thesis investigates how to forecast key macroeconomic variables when many potential predictors are available. We focus on GDP, unemployment, and inflation, and compare four modeling and forecasting approaches: dynamic factor models with principal component analysis, shrinkage methods, vector autoregressive (VAR) models, and Bayesian model averaging. Forecast accuracy is evaluated using the Diebold–Mariano test against a simple AR(p) benchmark. Empirically, a factor-augmented VAR (FAVAR) delivers the most accurate forecasts for GDP and inflation, while a Lasso-VAR performs best for unemployment. However, the Diebold–Mariano tests indicate that these more complex models do not produce statistically more accurate forecasts than the simple AR benchmark. Finally, we examine which predictors are selected across models; none are universal, though within each target variable certain predictors consistently appear.
[This summary has been generated with the help of AI directly from the project (PDF)]
Documents