預測市場準確度之實證研究：準確度、變數與預測誤差

Abstract

預測市場準確度之實證研究：準確度、變數與預測誤差 在國家科學委員會贊助下，本人與三位共同主持人合作建構的「未來事件交易所」 （電子交易市場）在2006 年7 月上線且公開營運，已經累積大量的預測市場交易資料。 在過去三年累積的大量交易資料基礎上，本計畫將進一步設計相關的預測合約，以便 在「未來事件交易所」累積更多的交易資料，以便對預測市場準確度進行實證研究。 實證研究分成三個部份： （1）、分析預測市場的準確度：包括是否預測型合約價格與機率的卡方檢定、是否 預測型合約價格與選舉結果的Logit 模型分析、比較預測市場與民調的準確度（包括正 確率、精準確、命中率、假警報率、貴氏比率差、得票率預測的準確度）及各種穩健 性的測試。 （2）、評估影響預測市場準確度的因素：包括是否預測型合約準確度的Logit 模型 分析、鑑別分析、決策樹歸納法；落點預測型合約準確度的迴歸分析、主成分分析。 同時，本研究將以2010‐2013 年的合約結果進行樣本外測試各種模型的預測能力。 （3）、預測市場的預測標準誤差：針對落點預測型合約，本研究以時間序列模型估 計預測市場某個時間點的預測誤差，並且考慮選舉合約的價格相互影響。再者，結合 是否預測型合約與落點預測型合約，本研究以隱含波動率模型推估得票率的區間範圍。 Empirical Studies on the Accuracy of Prediction Markets: Accuracy, Variables, and Forecast Error With sponsorship of the National Science Council, the Exchange of Future Events (an electronic exchange market) constructed by I and three other co‐director of the projects has been online and operate in public, accumulating enormous trading data of prediction markets. Based on the cumulative trading data over the last three years, this project will further design related prediction contracts in order to accumulate more trading data and thus conduct empirical studies on the accuracy of prediction markets. The empirical studies are divided into three parts: (1) Analysis on the accuracy of prediction markets, including Chi‐square test on the prices and probabilities of “yes or no contracts”, Logit model analysis on the prices and election results of “yes or no contracts”, comparison of accuracies for prediction markets and opinion polls (including correctness rate, precision rate, hit rate, false alarm rate, Kuipers score, and accuracy of vote share prediction) and various robustness tests. (2) Evaluation of variables affecting the accuracy of prediction markets, including Logit model analysis, discriminant analysis, decision tree induction on the accuracies of “yes or no contracts”; regression analysis and principal components analysis on the accuracies of “point‐estimation contracts”. (3) Forecast standard errors of prediction markets. Regarding “point‐estimation contracts”, this project uses time series models to estimate forecast standard errors of prediction markets at particular timing and take into consideration the interaction for prices of election contracts. In addition, combining both “yes or no contracts” and “point‐estimation contracts”, this project uses the implied volatility model to estimate the ranges of real vote shares for candidates of elections.