建構景氣指標方法之研析

A Study on the Method of Compiling Business Indicators

本研究參考Stock & Watson（1998 、2002） 利用主成分分析法 （Principal Component Regression, PCR）建立可預測景氣變化的擴散指標做法，嘗試由另一種方法建構景氣指標。進行實證時，首先將本會的領先指標（CLI）構成項目以主成分分析法建立新的綜合指標，觀察新指標預測景氣波峰谷底之表現，並建立靜態預測迴歸模型。此外，本研究另以LARS統計方法篩選出具領先性質的數列，同樣以PCR合成新指標，並建立動態預測迴歸模型。實證結果顯示：1.以主成分分析法將領先指標構成項目建立新綜合指標時，PCR與CLI在歷次景氣循環的領先期數均相同，但PCR的靜態迴歸模型預測績效優於CLI。2.以PCR結合LARS建立綜合指標時，在1982~2011.06樣本區間中，PCR與CLI在高峰谷底的平均領先期數均相同。在1995~2011.06及1998~2011.06樣本區間中，CLI在高峰領先期數優於PCR，而在谷底部分，PCR表現則優於CLI。3.若將PCR結合LARS合成之綜合指標與CLI分別建立動態迴歸預測模型時，兩者均以3個月前的預測表現較佳。三個樣本區間中，CLI的預測績效表現優於PCR。

This study tests the application of the diffusion index method of Stock and Watson (1998, 2002), who employed principal component regression (PCR) for macroeconomic forecasting, as an alternative method of constructing Taiwan’s business index. In the first part of the study, PCR is applied to construct a new composite index from components of the CEPD’s Composite Leading Index (CLI). This is followed by observation of the new index’s performance in predicting business cycle turning points, and construction of a static regression forecasting model. Also in the study, the least angle regression (LARS) method is applied to select series with leading characteristics, which are likewise compiled into a new index using PCR, with the construction of a dynamic regression forecasting model. The empirical results are as follows: 1. When the PCR method was used to construct a new composite index from the components of the CLI, it was found that it led the turning points of all of Taiwan’s business cycles by the same number of months as the CLI, but that its static regression model forecasted more effectively than the CLI’s. 2. When PCR was combined with LARS to construct a new composite index, it led the turning points of business cycles by the same number of months as the CLI during the sample period from 1982 to June 2011; but CLI performed better for the peaks during 1995 to June 2011 and 1998 to June 2011, whereas PCR performed better for the troughs in those two sample periods. 3. When dynamic regression forecasting models were constructed respectively for the LARS-PCR index and the CLI, both had relatively good forecasting performances three months ahead. In the three sample periods, the CLI had the better forecasting performance.