預測市場資料庫之建構 (II)

Abstract

「預測市場」資料庫之建立是基於人類對於「預測」迫切的需求。「預測市場」已經被廣泛應用到預測選舉結果、電影票房、產品銷售業績、專案進度、總體經濟指標、甚至國際政經風險變化。本資料庫採用「預測市場」機制為研究方法，透過國立政治大學預測市場研究中心、御言堂公司及中央研究院資訊科學研究所合作設立與管理的「未來事件交易所」（http://xfuture.org），提供以「未來事件」為買賣標的之交易平台。所有交易之形成皆以參與者對時事未來發展之預測為基礎，類似一般金融交易市場。隨著資料庫的持續推廣，上線人數不斷增加。迄2011 年4 月為止，「未來事件交易所」吸引了來自121 個國家、4,842 個城市的訪客。其中，台灣所有的主要城市都有參與者，中國地區的參與者來自858 個城市以上，而美國參與者則來自1,921 個城市以上。而交易者針對曾經發行和正在進行中的2,290個合約組（16,645 個合約），累積產生了超過2 億7 千萬口的交易量。本研究根據未來事件交易所的交易資料和本研究彙整的民調資料，分析預測市場對台灣2009 年縣市長選舉預測結果，並對比預測市場和民調機構對於此次選舉的預測。本研究發現：對當選人預測合約，預測市場的加權平均價格對當選比率在統計上有顯著正向影響，並且在統計上相當程度可表示為候選人當選之機率。再者，根據正確率、精準率、命中率、假警報率與貴氏比率差等五項指標，預測市場對當選人預測的能力均高於民調機構。對得票率的預測，預測市場的預測能力在選前20 天以後便高過民調機構，而且預測市場的預測準確度會隨著合約到期日的接近而逐漸增加。不過，本研究也認為民意調查的優點在於可以協助研究者進行變項的相關分析，所以可和預測市場同時運用，相互增強。本研究利用「未來事件交易所」的交易資料，分析預測市場「是否預測型」合約能否準確預測選舉結果，合約包括2006 年的北高市長選舉、2008 年的立委選舉及總統大選。本文分成四個部份分析「是否預測型」合約的準確度：加權平均價格與事件發生機率；加權平均價格與該合約結果的迴歸分析；五項比率（正確率、精準率、命中率、假警報率與貴氏比率差）分析；三項穩健性測試。本文發現：1、選舉當選預測合約的加權平均價格能充分反映候選人當選機率；2、加權平均價格是影響當選與否的主要因素；3、預測市場的預測準確度高於民調機構；4、預測市場的準確度隨時間接近選舉而逐步提升。

The construction of the prediction markets databases is based upon the aspiration of the human kind on prediction. The “prediction markets” have been extensively applied to predict election, movie box, product sales, progress of special projects, macroeconomic index, and even international political economic risk. The databases adopt the methodology of “prediction markets” and provide a platform of to trade future events through “the Exchange of Future Events”（http://xfuture.org）, an electronic trade market jointly established and managed by the Center for Prediction Market, National Chengchi University, Swarchy, Inc., and Institute of Information Science, Academia Sinica. All trades are conducted in accordance with the prediction of future events by participants, like a regular financial market. Along with the continuous expansion of the databases, online participants have been increasing very rapidly. By the end of April 2011, the Exchange of Future Events attracts participants from 121 countries and 4,842 cities around the world. Among them, participants come from all cities in Taiwan and more than 858 cities in China, more than 1,921 cities in the United States. Traders have accumulated more than 2.7 billion trades for 2,290 sets of contracts (16,645 contracts). Using trading data of the Exchange of Future Events, this paper analyzes whether yes-or-no contracts of prediction markets can accurately predict the election results, including Taipei and Kaohsiung mayoral election in 2006, legislator election and presidential election in 2008. This paper analyzes the accuracy of the yes-or-no contracts in four parts: volume-weighted average price and the probability of event occurrence, logit model analysis on the volume-weighted average prices and the results of these contracts, analysis of five rates (correctness rate, precision rate, hit rate, false alarm rate, Kuipers score), and three robustness tests. This paper finds that: 1. volume-weighted average price can reflect the probability of candidates winning elections; 2. volume-weighted average price is the main factor to predict the events; 3. the prediction accuracy of prediction markets is higher that of polling institutions; 4. the accuracy of prediction markets increases as the election day approaches. According to trading data of the Exchange of Future Events and opinion polls collected by this study, this paper analyzes the prediction results of the 2009 magistrate and mayoral election in Taiwan, and compares the prediction accuracies on this election between prediction markets and poll institutions. What this paper finds are: for prediction contracts on election winners, the weighted average prices of prediction markets are positive and statistically significant on the ratio of winning elections and can be regarded as the candidates’ probability of winning elections. In addition, based upon five indicators of correctness rate, precision rate, hit rate, false alarm rate and Kuipers score, predictive power of prediction markets on election winners is obviously higher than that of poll institutions. For prediction on vote shares, predictive power of prediction markets is higher than that of poll institutions within 20 days before the election, and prediction accuracy of prediction markets is getting higher along with approaching the expiration of the contracts. Nevertheless, we also agree that opinion survey can help researchers conduct covariance analysis, which can be used together with prediction market to reinforce the findings of each other.