使用標的對應訊息的新經濟預測方法

Abstract

近年來，由於大量時間序列資料的建立，如何有效地從其中萃取有用的訊息並應用於經濟預測上已經吸引眾多學者的目光。在此計畫中，針對不同預測標的變數，我提出了一個相當可信且具有可行性的方法將有效的訊息從大量時間序列資料中擷取出來。我結合了主成分分析 (principal component analysis, PCA) 及最小絕對值縮減及選擇運算法 (least absolute shrinkage and selection operator, LASSO)兩種方法的優點，建立了一個兩階段的預測模式。在第一階段中，我利用 PCA 將大量時間序列資料中的訊息重新組織成具有正交化 (orthogonality) 特性的各個主成分。在第二階段中的線性預測模型中，我再利用 LASSO 來選取對於預測標的變數具有良好解釋預測能力的訊息(主成分)並同時估計其相對應的模型係數。這樣根據預測標的變數而對應出的訊息排序方式將和一般直接利用 PCA 的排序方式大不相同。再者，因為這些主成分都具有正交的特性，因此 LASSO 具有良好的性質且其對應的模型係數估計具有明確的形式。這些方法和特質都和文獻上目前盛行採用 Stock and Watson(2002）的因子模型及其相關研究大不相同。我在計畫中也探討了該方法相對應的理論性質，並進行了模擬及實證研究。結果顯示，此計畫提出的方法應可為研究者在進行預測時的另一種可行的選擇。

Recently,because lots of time series data are available,how to efficiently extract the useful information from these time series for macroecominic forecasting has been a heavily researched topic in the literature.In this project,I propose a promising approach to forecast target variable based on extracting useful information from large dimensional variables.I take both the advantages of the principal component analysis (PCA)and the least absolute shrinkage and selection operator (LASSO) by a two-step approach.In the first step,I re-organize the information containing in the original predictors by using PCA. All these orthogonally principal components are then introduced in forecasting target variable. In the second step, the targeted principal components and the estimates of the linear forecasting model will be simultaneously determined by implementing the LASSO.Unlike the ""natural ranking‘‘ of PCA, it gives the ""targeted ranking‘‘ instead.Moreover,because the principal components are orthogonal,the LASSO has good properties and it also gives the analytic form of the estimates.The proposed approach differs greatly from the factor approach framework of Stock and Watson(2002)and other related in the literature. I discuss the properties of the proposed and present a simulation and an empirical study in this report.Based on these results, I think the proposed can be a candidate for economic forecasting.