探索預測台灣通膨隱而未現的重要因子──監督式降維模型的實證

Disentangling Latent Factors for Inflation Forecasting in Taiwan via Supervised Dimension-Reduction Approach

研究目的：根據傳統經濟理論認定與通膨相關的變數，侷限了研究者發現其他重要變數的可能性。本文探討高維度資料集於通膨預測之應用，嘗試由下而上利用改良後之擴散指數預測法(diffusion index forecasting)預測台灣通膨。
研究設計／方法：本研究以2000年至2021年間，近百個對於台灣通膨具有潛在影響力變數，採納Stock and Watson (2002b)之建議，事先將變數分為11大類後再進行預測，探討不同降維方法所萃取之潛在因子(latent factor)對模型預測力的影響。
研究結果：本文發現使用監督式的降維方法有助於提升模型整體預測能力。
研究限制／啟發：發現在分類前預測力最好的偏分量迴歸(PQR)於分類後建模之模型預測力有了更進一步提升。
理論／實務／社會意涵：分析架構中可以探討預測通膨的關鍵變數，在不同的時空背景下11大類別相對重要性之消長與經濟意義，解構其中與後續政策干預攸關的變數說明。
創見／價值：本研究也建構溫和與極端通膨(縮)預警模型，做為台灣央行制定貨幣政策時的參考依據。

Purpose–Past literature on Taiwan’s inflation forecasting mostly confines to only few theory-specific variables, which limits the possibility of other potential important variables. In view of the superior forecasts from the diffusion index method via incorporating large dimension information.
Design/methodology/approach–We generalize the framework to allow for a richer spectrum that encompass and compare various linear/nonlinear, supervised/unsupervised dimensionality reduction methods. We collected nearly 100 potential variables, from the period of 2000 to 2021, in order to extract the hidden common factors for inflation forecasting.
Findings–Among the examined 4 approaches, our results indicate that the supervised partial quantile regression (PQR) dominate the other 3 approaches in anticipating inflation.
Research limitations/implications–Once we further divide variables into 11 categories and extract category-specific factors for the subsequent forecasting as in Stock and Watson (2002b), we found that the predictability of PQR became even better.
Practical implications/Social implications–We can not only visualize the importance of each category in 1-step ahead inflation projection across time.
Originality/value–We can establish an early warning model for monitoring the arrival of radical inflation/deflation and promptly adjusting for policy interventions.