社會科學院: 經濟學研究所指導教授: 管中閔陳映君Chen, Ying-ChinYing-ChinChen2017-03-032018-06-282017-03-032018-06-282014http://ntur.lib.ntu.edu.tw//handle/246246/275344本論文討論三種考慮應變數之因素模型(supervised factor model): 偏最小平方回歸模型(Partial Least Square,PLS)、主共變量回歸模型(Principal Covariate Regression,PCovR)及組合預測之主成分分析模型(Combining Forecasts Principal Component Analysi,CFPC)。我們將上述三種考慮應變數之因素模型應用在台灣經濟成長率之預測，並以均方根預測誤差(RMSFE)及平均絕對預測誤差(MAFE)衡量其預測之優劣。我們發現，考慮應變數之因素模型通常較不考慮應變數的因素模型(Principal Component Analysis, PCA)在預測上有更小的預測誤差，其中又以CFPC表現最好。另外，我們也比較考慮應變數之因素模型與主計總處對經濟成長之預測。我們發現CFPC之預測與主計總處之預測能力不相上下，因此我們認為CFPC是一個能夠避免模型錯誤設定的簡化模型(reduce form)。This thesis discusses three supervised factor models: Partial Least Square (PLS), Principal Covariate Regression (PCovR) and Combining Forecasts Principal Component Analysis (CFPC). We apply the supervised and unsupervised factor models to forecast Taiwan’s economic growth rates with 77 macroeconomic variables. We evaluate the performance of different models by comparing their RMSFE and MAFE. We found that the supervised factor models usually outperform unsupervised factor model (Principal Component Analysis, PCA) and that CFPC performs the best among the three supervised factor models. Besides, the forecasts of CFPC and Directorate General of Budget, Accounting and Statistics (DGBAS) have similar performance based on the Diebold-Mariano (DM) test, so CFPC may be a good alternative when we want to avoid ad hoc models. PCovR and PLS also have smaller RMSFE and MAFE than PCA, but they are not statistically significantly better than PCA.4247557 bytesapplication/pdf論文公開時間: 2015/8/11論文使用權限: 同意有償授權(權利金給回饋本人)因素模型主成份分析偏最小平方回歸模型主共變量回歸模型組合 預測之主成分分析預測Factor AnalysisPrincipal Component AnalysisPartial Least SquarePrincipal Covariate RegressionCombining Forecast Principal Component AnalysisForecast[SDGs]SDG8thesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/275344/1/ntu-103-R01323010-1.pdf