應用機器學習方法評估台灣通膨風險

NCU Institutional Repository > 管理學院 > 經濟研究所 > 博碩士論文 > Item 987654321/98144

請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/98144

題名:	應用機器學習方法評估台灣通膨風險
作者:	邱宇潔;Chiu, Yu-Chieh
貢獻者:	經濟學系
關鍵詞:	通膨風險;機器學習;特徵篩選;分量迴歸;偏斜t分配;Inflation Risk;Machine Learning;Quantile Regression;Skewed-t Distribution;Feature Selection;Inflation-at-Risk
日期:	2025-06-25
上傳時間:	2025-10-17 12:25:01 (UTC+8)
出版者:	國立中央大學
摘要:	本研究旨在建立台灣的通膨風險評估架構，透過結合機器學習與分量迴歸方法，分析未來不同期別下消費者物價指數年增率的極端變動風險。本研究以2000 年1月至2024年10月之台灣總體經濟與金融變數月資料為樣本，使用LASSO、Elastic Net、Adaptive LASSO、PCA、隨機森林、XGBoost、LSTM 七種特徵選取法，從大量變數中篩選重要因子，並透過分量迴歸估計各分位數下的通膨風險，結合t-skew分布進行尾部風險調整，建立ES、EL兩指標來選擇模型，以其篩選出來的變數做各分位係數分析，最後以Q95與Q5的差距為衡量，探討重大政策調整是否有助於穩定風險。實證結果顯示，不同模型在上行與下行風險的預測上具異質性，Elastic Net在短期EL評估上表現較優，PCA與LSTM在中長期EL評估上均表現優異。而不同變數對通膨風險的影響在各分位間呈現分位數不對稱性，部分變數甚至在低分位與高分位方向相反，顯示傳統線性預測方法或許低估極端價格變動風險。本研究建議政策制定者應採用多模型架構與分量風險評估方法，不同觀察期採以不同模型分析，以強化通膨風險管理之全面性。 ;This study aims to establish an inflation risk assessment framework for Taiwan by integrating machine learning techniques with quantile regression to analyze the extreme fluctuation risks of the Consumer Price Index (CPI) annual growth rate across different forecast horizons. Using macroeconomic and financial data from Taiwan spanning January 2000 to October 2024, the study applies seven feature selection methods including LASSO, Elastic Net, Adaptive LASSO, PCA, Random Forest, XGBoost, and LSTM to extract key predictors from a high-dimensional dataset. These selected features are then incorporated into quantile regression models to estimate Inflation-at-Risk (IaR) at various quantiles. To better capture tail risk dynamics, the t-skew distribution is employed for smoothing, enabling the computation of Expected Shortfall (ES) and Expected Longrise (EL). Models with superior performance are identified, and the corresponding features are visualized using quantile coefficient plots. Finally, the study uses the gap between Q95 and Q5 as a measure to evaluate whether major policy adjustments help stabilize inflation risk. Empirical results reveal heterogeneity in model performance regarding upward and downward risk prediction. Elastic Net performs better in short-term EL estimation, while PCA and LSTM outperform in medium to long-term EL forecasting. The impact of individual variables on inflation risk demonstrates nonlinearity and quantile-specific asymmetry, with some variables exhibiting opposite effects across low and high quantiles. These findings suggest that traditional linear forecasting models may underestimate extreme inflation risks. The study recommends that policymakers adopt a multi-model architecture and quantile-based risk assessment approach, applying different models across different forecast horizons to enhance the accuracy and robustness of inflation risk management.
顯示於類別:	[經濟研究所 ] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	18	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....