臺灣景氣狀態之預測

FORECASTING BUSINESS CYCLES IN TAIWAN

國發會認定並發布景氣循環峰谷時點存在一定時間的落後， 使施政單 位不易即時掌握景氣狀態變化， 鑑於近年臺灣經濟成長變動劇烈， 本文搜 集大量國內、外總體經濟變數， 預測 2001 年以來臺灣景氣狀態變化。除採 logistic 模型逐一檢視各變數對當期或未來臺灣景氣衰退的預測表現外， 亦 運用 4 種能考量大量變數資訊的預測模型:以向前選取法（forward stepwise approach）篩選變數組成多變量 logistic 模型，運用主成分分析從大量總體變 數萃取主要因子，組成 logistic 因子模型，以及 random forest 與 boosting 兩 種機器學習演算法， 進而比較不同預測模型的樣本外預測表現。根據樣本外 預測結果，在向前零期預測（zero-step ahead forecast）方面，採用實質海關出 口值單一變數即能有效捕捉當月景氣狀態，而對未來 1-6 個月的短中期預測 時， 考量大量變數資訊的機器學習法預測表現普遍較佳， 而因子模型對未來 1-3 個月的預測亦有不錯的表現。

Announcements of the dates of peaks and troughs in business cycles are made with such a considerable lag that policy makers can't know the current business cycle regime early. In this paper, we collect a wide range of domestic and foreign business indicators to assess their abilities to forecast Taiwan's recessions since 2001. We adopt logistic models to examine the usefulness of each indicator and consider four different methods to extract relevant information from all indicators: a forward stepwise approach, a logistic factor model, and two machine learning approaches. The analysis focuses on outof-sample performance from the current month (zero-step-ahead forecast) to six months ahead. Our empirical results show that the ''real exports of goods'' has the best predictive power for zero-step-ahead forecast. However, for onemonth- to six-month-ahead forecasts, machine learning techniques show better out-of-sample performance than other forecast models.