ETF 定價效率之決定因素：費用率異質性效果之跨市場實證研究

Abstract

指數股票型基金（Exchange-Traded Fund, ETF）在過去十年間經歷前所未有之規模擴張，截至 2025 年底，全球 ETF 資產管理規模已突破 19.85 兆美元，美國市場單一年度淨流入即達 1.49 兆美元之歷史新高。在台灣，ETF 資產規模亦於 2025 年攀升至 7.59 兆新台幣，受益人數突破 1,500 萬人次。然而，ETF 之定價效率，即其市場價格能否忠實反映標的資產價值，始終為學術研究與實務投資關注之核心議題。

本研究以美國與台灣市場之 ETF 為研究對象，涵蓋 2024 年 10 月至 2025 年 10 月之橫斷面資料，分別以追蹤誤差絕對值（Absolute Tracking Error）與折溢價偏離絕對值（Absolute Premium/Discount）作為定價效率之代理變數。研究採用遞進式（incremental Specification）普通最小平方法（OLS）迴歸並搭配異質變異穩健標準誤（HC3），逐步納入基金基本特徵、交易與流動性指標、投資策略與市場分類、ETF 結構類別，以及費用率（Expense Ratio）與各分類變數之交乘項，系統性地檢驗各因素對定價效率之影響，並額外以美國與台灣子樣本分別進行迴歸，以檢驗結果之跨市場穩健性。

實證結果顯示，對數淨資產價值（ln NAV）、持股數量（Holdings）與一年期報酬率（1Y Return）為跨模型一致顯著之核心變數。費用率（Expense Ratio）本身在控制其他變數後不具獨立顯著性，惟透過交乘項分析發現其效果具有明顯異質性：以台灣為參照組，費用率對美國市場 ETF 之定價效率拖累顯著受到緩衝，且在主題型 ETF 中之放大效果亦顯著高於市值加權型 ETF。子樣本分析進一步驗證此結果之穩健性。兩項定價效率指標之關鍵差異在於買賣價差（Bid-Ask Spread）的角色，其在追蹤誤差模型中始終不顯著，卻在折溢價模型中高度顯著，反映兩項指標本質上的驅動機制差異：前者主要源自基金內部複製操作，後者則高度依賴外部次級市場之即時定價效率。

本研究之學術貢獻在於：首先，同時採用追蹤誤差與折價/溢價雙指標進行跨市場比較，補充現有文獻偏重單一市場或單一指標之不足；其次，透過遞進式模型設計與費用率交乘項，揭示費用率影響之非線性與情境依賴特質；最後，實證結果為投資人之 ETF 選擇、基金公司之產品設計及監管機構之市場發展策略提供具體參考依據。

Exchange-traded funds (ETFs) have experienced unprecedented growth over the past decade, with global ETF assets surpassing USD 19.85 trillion by year-end 2025 and the U.S. market alone attracting a record USD 1.49 trillion in net inflows. In Taiwan, ETF assets reached TWD 7.59 trillion with over 15 million beneficiary accounts. Despite this expansion, the pricing efficiency of ETFs, whether their market prices faithfully reflect underlying asset values, remains a central concern for both academic research and investment practice.

This study examines ETFs listed in the U.S. and Taiwan markets using cross-sectional data spanning October 2024 to October 2025. Two complementary measures of pricing efficiency are employed as dependent variables: the absolute value of tracking error (|TE|) and the absolute value of premium/discount deviation (|PD|). Through a set of incrementally specified OLS regression models (M1 through M5) with HC3 heteroskedasticity-robust standard errors, this study systematically investigates the effects of fund structural characteristics, trading and liquidity variables, investment strategy classifications, and ETF construction attributes on pricing efficiency, with particular focus on the heterogeneous effects of expense ratios across different ETF types, market environments, and replication methods.

Empirical results reveal that net asset value (NAV), number of holdings, and one-year return are consistently significant across all model specifications in both experiments, with NAV exhibiting a negative effect and the latter two showing positive effects. The expense ratio itself is not significant in baseline models but becomes prominent once interaction terms are introduced: its positive drag on pricing efficiency is significantly larger in Taiwan than in the U.S. and more pronounced for thematic ETFs than for cap-weighted funds. A critical distinction between the two experiments lies in the role of the bid-ask spread, which is consistently insignificant in the tracking error model yet highly significant in the premium/discount model, reflecting the fundamentally different driving mechanisms: the former primarily originates from internal fund replication processes, whereas the latter depends heavily on the real-time pricing efficiency of the secondary market.

The academic contributions of this study are threefold. First, by simultaneously employing both tracking error and premium/discount as dual pricing efficiency metrics in a cross-market comparison framework, this study addresses the predominant focus on single markets or single indicators in the existing literature. Second, through the incremental model design incorporating expense ratio interaction terms, this study reveals the nonlinear and context-dependent nature of the expense ratio's impact on pricing efficiency. Third, the empirical findings provide concrete reference points for investors in ETF selection, fund companies in product design, and regulators in market development strategies.