Previsão de variáveis macroeconômicas utilizando modelos fatoriais

• Autor(a) principal: Silva, Thiberio Mota da
• Data de Publicação: 2017
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Repositório Institucional da Universidade Federal do Ceará (UFC)
• Texto Completo: http://www.repositorio.ufc.br/handle/riufc/22952
• topic: Macroeconomia; Variáveis macroeconômicas

The first article analyzes the performance of several high quality models to predict nine Brazilian macroeconomic variables, including unemployment rate, industrial production index, six price indices and exchange rate. The factors are extracted from a set of data composed of 117 macroeconomic variables. Reduction methods and combinations of forecasts are used to select or combine the best predictor factorial model. Regressions of minimum angle and Elastic Net, Bagging, Non-Negative Garrote. In turn, the methods of prediction combinations applied were based on the criterion of Mallows, weighted Bayesian methods, cross-validation, Leave-h-Out and Jacknife weighting model. All predictions were evaluated using the Model Condition Set (MCS) that establish the best forecast models that satisfy a confidence interval for the forecast error. The results suggest that, in general, the factorial models present a mean square error of prediction (MSFE) lower than the AR benchmark (4). The best models of factors to predict Brazilian macroeconomic variables. In some positions of the Non-Negative Garrote and Bayesian weighting models presented satisfactory predictions. The second article analyzes the performance of supervised versus unsupervised factorial models to predict four Brazilian macroeconomic variables, including industrial production index, broad consumer price index, national consumer price index and Long-term interest. The factors are extracted from a set of data composed of 117 macroeconomic variables. The models were extracted by means of combination of factors, no case, supervised models, and there are no cases of unsupervised models, was used by the method of combined forecasts, or even used by Tu and Lee (2012). The Autoregressive Models Augmented (FAAR), Weighted Bayesian Models (BMA), weighted Mallows model (MMA), weighted jacknife model (JMA), cross validation (SMA). The best model is the one that presents a smaller mean square error quadrant (REQM). The results show that the mobile window model was more capable and predictive model that obtained better performance for the BMA weighted model, for both factorial models, supervised or not, in addition, supervised models are more efficient to perform non-average predictions Three-year-old, among the four, target variables, presenting a minor (REQM). The third article proposes a method of dynamic data weighting (DMA) applied to large databases. How variables contained in the database have a size reduced to a number of factors that are dynamically combined through forgetting factors. The extraction of factors also considers an exponential window of exhaustion that aims to reduce the impact of very old observations. It is shown that model, called FDMA, converge asymptotically to a dynamic combination of observed factors when the number of variables or the size of the sample from outside grows. In addition, the FDMA model is applied in two different empirical exercises. 1. The empirical results show that the FDMA and its variations are considered as an alternative to the forecast.