利用高維數據探討臺灣國際貿易預測的改進：本地與全球變數的應用

Enhancing Trade Forecasting Accuracy in Taiwan: A High-Dimensional Data Approach with Local and Global Variables

本研究旨在進一步探討如何透過靈活的數據轉換與適用於高維度數據的預測方法來提升臺灣國際貿易預測的準確性。除了臺灣自身相關的總體變數我們亦考慮臺灣五十大貿易夥伴的數據，並運用降維與變數選擇方法（如least absolute shrinkage and selection operator，Lasso）來提升預測表現。儘管高維度數據與變數是否為定態確實對傳統計量推論帶來許多的挑戰，但本研究整合了因子分析、多種高維度處理的方式以及不同的數據轉換方式，並利用全球與本地因子的正交特性來探討預測能力的改善。實證結果顯示，臺灣貿易預測並不存在通用的最佳策略;此外，若變數的集合選擇合適，水準數據亦可能更適用於提升預測能力。

This paper bridges the gap in forecasting Taiwan’s international trade by exploring flexible data transformations and estimation strategies suitable for high-dimensional datasets. It focuses on local variables and data from Taiwan’s fifty major trade partners, employing dimension reduction and variable selection methods like Lasso to enhance forecasting accuracy. Despite the challenges associated with managing high-dimensional data and ensuring data stationarity, our approach integrates factor extraction with regularization. This takes advantage of the orthogonal properties of global and local factors to improve forecasting capabilities. Empirical results indicate the absence of a universally robust forecasting strategy. They also suggest that level data could be viable for forecasting when the data pool is appropriately selected.