本論文旨在探討債券超額報酬之可預測性(predictability)，並首度將工具變數主成分分析(instrumented principal component analysis, IPCA)模型應用於固定收益市場。相較於傳統主成分分析(principal component analysis, PCA)其聚焦於變數間共變異的特性，IPCA模型結合資產特徵與潛在因子結構，能同時捕捉橫斷面與時間序列之變動，有助於提升預測能力。本文使用1972年至2024年之美國零息公債殖利率資料，結合遠期利率與總體經濟變數，進行樣本內與樣本外的遞迴(recursive)預測分析，並針對不同債券年期及景氣階段比較模型的預測表現。實證結果顯示，傳統PCA模型雖於樣本內預測表現較佳，但其樣本外預測能力顯著變差；反觀IPCA模型展現出穩定的樣本外預測效果。另外，經過PCA方法進行降維處理的總體經濟變數，能有效過濾雜訊並提升模型的預測效能。此外，IPCA模型所估計的因子具備一定的經濟解釋力，其在不同年期債券間呈現出預測能力差異，並受到景氣循環影響，於景氣擴張期間對短天期債券表現較佳，衰退期間則是對長天期債券之預測表現較佳。本研究驗證IPCA模型於固定收益資產上的應用潛力，亦強調總體經濟變數之結構與景氣循環週期對於預測債券超額報酬的重要性。;This thesis studies the predictability of bond excess returns and is the first to apply the instrumented principal component analysis (IPCA) model to the fixed income market. Compared to the traditional principal component analysis (PCA), which focuses on the co-movement between variables, the IPCA model links asset characteristics to hidden factor structures, capturing both cross-sectional and time-series variations to improve prediction accuracy. Using U.S. zero-coupon government bond yield data from 1972 to 2024, this study combines forward rates and macroeconomic variables to conduct in-sample and out-of-sample recursive forecasting, and compares models’ forecasting performances across different bond maturities and phases of business cycle. The empirical results show that while the PCA model performs better in in-sample forecasting, its out-of-sample predictive power declines significantly. In contrast, the IPCA model provides more stable performances in out-of-sample forecasts. Moreover, applying PCA to reduce the dimensionality of macroeconomic variables helps filter out noises and improve the model’s forecasting performance. The factors estimated by the IPCA model also have economic implications, and the model′s predictive power varies across bond maturities and phases of business cycles. The IPCA model performs better for short-term bonds during economic expansions and for long-term bonds during recessions. This study confirms the potential of the IPCA model for fixed income applications and highlights the importance of macroeconomic variables and business cycles in predicting bond excess returns.
