預測臺灣股市風險溢酬分位數 - 以偏分量迴歸整合高維資訊;Predicting Quantiles of Taiwan Equity Risk Premium - Aggregating Information via Partial Quantile Regression

NCUIR > School of Management at National Central University > Graduate Institute of Finance > Electronic Thesis & Dissertation > Item 987654321/89544

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/89544

Title:	預測臺灣股市風險溢酬分位數 - 以偏分量迴歸整合高維資訊;Predicting Quantiles of Taiwan Equity Risk Premium - Aggregating Information via Partial Quantile Regression
Authors:	霜皓帆;Shuang, Hao-Fan
Contributors:	財務金融學系
Keywords:	臺灣股價指數;機器學習;降維;分位數;投資策略;樣本外預測;Return premium predictability;Machine learning;Dimension reduction;Quantiles;Asset allocation;Out-of-sample forecasting
Date:	2022-09-12
Issue Date:	2022-10-04 11:44:38 (UTC+8)
Publisher:	國立中央大學
Abstract:	本文以臺灣加權股價指數之風險溢酬第 5(崩盤)、20(大跌)、50(平盤)、80(大漲)、 95 (過熱) 之條件分量為預測目標。樣本為 2000 年 1 月至 2021 年 12 月的月頻率資料。本文收集包括總體經濟指標、基本價值變數、交易活動指標、價格趨勢、貨幣及信用指標共 34 個預測變數，採用包括主成份分析、偏最小平方迴歸法，以及偏分量迴歸共計 3 種降維模型萃取有用資訊估計共同因子，進而預測風險溢酬之條件分量。透過全樣本與樣本外遞迴式預測實證分析中比較各模型的優劣，我們發現主成份分析法之預測績效最差;偏最小平方迴歸法則受制於其共同因子無法因應因各目標分量的需求，只在預測崩盤、大跌具有預測能力;偏分量迴歸則得益於其共同因子會隨目標分量改變而不同，使其能夠穩定地預測 5 個不同分量且具有明顯優勢。此外，透過檢驗偏分量迴歸之共同因子變數組成成份，我們發現價格趨勢變數在各個條件分量下都具有預測能力;基本價值變數在預測平盤時較具優勢;交易活動變數在預測崩盤與過熱時較能體現其重要性;總體經濟指標適用於大漲及過熱;貨幣及信用變數則是以平盤以下較具有預測力。最後，本文嘗試以偏分量迴歸之預測值建構交易策略，在考慮了交易成本、參考夏普比率、確定等價收益率 (CER) 等績效衡量指標下，我們所建構的交易策略明顯優於買進並持有 (Buy and Hold)，可見預測條件分量具有相當的經濟價值。;We conduct a predictive performance analysis of the 5 conditional quantiles of the TAIEX equity risk premium using extracted information from high dimensional data from different aspects of risk premium: price trend, fundamental, trading activity, macroeconomic, monetary variables. We pick the 5th, 20th, 50th, 80th, 95th quantiles to represent the 5 market states of the risk premium as “collapse”, “plummet”, “stable”, “soar”, “boom”, respectively. We perform a comparative analysis among dimension reduction models, including Principal Component Analysis (PCA), Partial Least Squared (PLS), Partial Quantile Regression (PQR). The empirical results show that (1) PCA performs the worst in forecasting; (2) As the PLS factors fail to characterize the spectrum of each target quantile, there is an obvious "asymmetric" phenomenon in that PLS predicts well only in the “collapse” and “plummet” states; (3) PQR can predict 5 different quantiles consistently and have obvious advantages over the other two methods. Finally, in order to confirm the economic benefits of predicting conditional quantiles, we construct a trading strategy using the predicted values from the PQR. Taking into account transaction costs, and performance measures such as the Sharpe ratio, and the certainty equivalent rate of return, the constructed trading strategy delivers significantly economic values and outperforms the commonly used Buy and Hold strategy.
Appears in Collections:	[財務金融研究所] 博碩士論文

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	68	View/Open

社群 sharing

Loading...