本篇論文運用Stock and Watson (2002)的動態因子模型(dynamic factor model)去預測7個台灣重要的經濟變數。我們利用184個台灣、美國和日本的季資料,資料期間從1981年第一季到2006年第四季,將資料分成為只有包含95個台灣資料與包含全部184個台灣、美國和日本的資料,然後比較其預測的結果。並且我們利用Diebold and Mariano (1995) 所發展的檢定與Clark and West (2007)檢定,作為模型準確性的衡量準則。在研究結果中顯示在一到四季的模型外預測,利用變數本身幾期的落後與少數的因子能夠改善傳統的總體預測模型,並且當我們分析全部184個資料時,預測的結果與只分析95個台灣的資料時並沒有顯著的不同。此外利用動態因子模型所產生的十二個因子當中,我們發現第一個與第二個因子可以表示為實質指標,而第三個因子可以表示為物價指標。
This article extends the dynamic factor model of Stock and Watson (2002) to forecast seven important Taiwan variables. The data contain 184 quarterly time series for Taiwan, the United States, and Japan from 1981Q1 to 2006Q4. We divide the data into two parts; the first part only contains Taiwan 95 quarterly data. The second part contains all 184 quarterly data as a compare with the first part. We use Diebold and Mariano (1995) criterion and Clark and West (2007) criterion to be the forecasting accuracy test. We find a way to improve the traditional macroeconomic forecasting model which is using lags of itself and few factors for 1-, 2-, 3- and 4-quarter-ahead forecasts. The performance of forecasting has no significant difference when we add US and Japan series. Moreover, we find the first and second factors are the real indictors and the third factor is the price indictor.