共識預測的實證研究

Abstract

本研究採用平均數、中位數、切尾均值及Winsor化均值來探討哪一個方法適合用於共識預測，因此用這四個方法去研究1995年至2008年台灣、中國、日本、美國、英國、法國和德國這七個國家每月實質國內生產毛額、消費者物價指數及失業率預測機構的預測誤差，接著計算預測機構預測的準確性，最後比較在經濟正長期和衰退期做預測時，預測的方法是否有改變。實證結果發現：1. 整體而言切尾均值及Winsor化均值這二個穩健估計(Robust statistics)的方法確實會使預測誤差變小。2. 無論是中位數、切尾均值及Winsor化均值在做估計時會較平均數準確。3. 在衰退期時，用切尾均值和Winsor 化均值這二個方法並無顯著的不同。

We use the mean, median, trimmed mean and Winsorized mean in this study to investigate which method is easy to use and understand in consensus forecast. The data which we employ from Consensus Economics, Inc. are monthly real GDP, inflation and unemployment rate of seven countries (Taiwan, China, Japan, the United States of America, United Kingdom, France and Germany). We use them to state the forecast bias and accuracy during the period 1995 to 2008 and then to compare which method is better in recession and non-recession years. We find that: First of all, two simple robust methods, trimmed and Winsorized means, are slightly more accurate than other methods. Secondly, no matter what the median, trimmed mean and Winsorized mean are more accurate than the value of mean. Thirdly, in the recession years, the methods of trimmed and the Winsorized means are not different significantly.