預測市場是近年來新發展出的預測方法,許多實證研究均證明預測 市場能有效整合資訊並提出準確的預測。然而在大多數預測市場研 究中, 研究者只能由過去的歷史準確率來衡量市場預測的可靠性, 無法針對單一市場合約預測的正確與否進行事前的評估。本文提出 一個植基於市場交易特徵的合併鑑別方法, 藉由整合迴歸模型、多 變量分析、決策樹、及支持向量機等四種模型來擷取與市場預測準 確率有關的潛在資訊。本文使用未來事件交易所自2006年至2011 年共650個選舉合約作為資料, 經實證分析後驗證合併鑑別模型可 以非常準確地於事前對任一合約預測的正確與否提出評斷。本文所 提出的合併鑑別方法不但比單一鑑別模型更為可靠,而且可依決策 者不同的目標函數提出不同的評斷以進行風險控管。

In prediction markets (PM) which are being used widely in many fields, contemporary researchers and practitioners have to rely on historical accuracy to evaluate the plausibility of current events. Based on the empirical and theoretic findings concerning the accuracy of prediction markets, this paper proposes a combined identification method which can evaluate the accuracy of PM events in advance. The proposed method not only takes a variety of market features into account, but also combines the forecasts of different statistical and machine learning techniques to fully capture the patterns underneath. We test the proposed method with transaction data from 2006 to 2011. This study proves that it is possible to evaluate the accuracy of any PM event in advance with high accuracy. We also show that the combined modeling is a superior method in the sense that it not only can provide higher identification accuracy, but is also flexible enough to incorporate decision makers’ goals and preferences into the identification process.