群眾募資專案之動態預測：台灣群眾募資平台flyingV之實證研究

Abstract

隨著群眾募資的興起，越來越多學者對預測募資結果的研究感到興趣，然而多數的預測文獻為靜態預測，對於募資期間變化進行分析的動態預測卻十分稀少；另方面過往動態預測文獻所使用的演算法往往需要大量歷程樣本訓練、且其多缺乏易於理解的預測規則，對於沒有相關團隊及資料的提案人而言，過往文獻缺乏立即實務運用的價值。故本研究蒐集flyingV募資專案的歷程資料，並以決策樹作為預測演算法，運用其易懂好用的預測規則，提供給台灣募資提案人可直接運用的預測判別準則。為了證明決策樹與過往文獻建議的演算法有相當的預測能力，以及盡可能提供簡單好用的預測規則，本研究設計了三個模型： (1) 決策樹模型與KNN模型的預測準確率比較 (2) 以募資第N天做為動態預測時間軸的決策樹 (3) 接續模型二，並僅以募資達成率作為規則用的決策樹 藉此依序證明決策樹有與KNN演算法相當的預測能力、使用募資第N天作為動態預測時間軸有更好的預測能力，以及僅用募資達成率做為預測規則已有足夠的預測能力。模型實驗結果如下： (1) KNN模型在預測準確率上大致略高於決策樹模型，本研究透過兩相依母體期望值差T檢定證明兩者預測準確率上沒有顯著差異。 (2) 相對過往文獻以募資天數進度百分比，以募資第N天作為動態預測時間軸的設計排除單位時間上不同的誤差，而有較佳的預測準確率。 (3) 僅以募資達成率作為規則的決策樹，演算出只要判斷第一天募資達成率是否超越4%，就有近80%的預測準確率。 本研究證實了可以僅用募資達成率作規則的決策樹能有80%以上的預測準確率，並進一步整理出募資前14天準確率的變化及對應的募資達成率規則閾值，除了給予提案人簡易現成的預測規則、可作為募資後每日的關鍵績效指標衡量判斷外，本研究也發現前14天的募資達成率閾值皆不到20%，卻仍有相當高的預測準確率，推測多數成功募資案為慢熱型，前期應為口碑推廣的醞釀期。

With the rising of the crowdfunding, more and more professors do the research in the prediction of crowdfunding. However, most of their studies are static prediction, but not the dynamic prediction with the changes during the funding period. On the other hand, the algorithm the past dynamic prediction researches use needs large training sample and its predictive rules are not easy-to-use, providing less practical values for funders who may not have relating experts as well as a bunch of samples. This study uses the decision tree, reputed for its easy-understanding predictive rules, as algorithm with the daily data from flyingV to offer the predictive criteria for funders in Taiwan. To prove that the ability of the decision tree is comparable with the algorithm the past researches recommended and to provide simple and useful predictive rules, this study designed the three models: (1) The comparison for accuracy between decision tree model and KNN model (2) The decision tree model using daily funding as the time series for dynamic prediction (3) Same as model 2, but only using the percentage of fund-raising as input variable Through these models, we can prove sequentially that decision tree is comparable with KNN algorithm, using daily funding as the time series has better predictive performance, and that only using the percentage of fund-raising as predictive rules is enough to predict results. The results are following: (1) Though the performance of KNN is slightly higher than decision tree’s, we proved there is not significant difference between them with paired samples t-tests. (2) Compared using the percentage of funding duration as the time series for dynamic prediction in the past studies, using the daily funding excludes the error within different durations among funding projects, having better performance. (3) The decision tree only using the percentage of fund-raising as rules shows that only we check if the percentage of fund-raising at the 1st day is larger than 4%, we have nearly 80% for accuracy. This study shows that the decision tree only using the percentage of fund-raising as rules can reach above 80% for accuracy, and provides the accuracy chart and paired threshold rules for the percentage of fund-raising for the beginning 14 days. This study provides simple predictive rules as daily KPI for funders, and found the thresholds for the beginning 14 days are under 20% but with relatively high accuracy, implying that most successful crowdfunding projects are slow-to-warm-up, and the early period should be the brewing period for word-of-mouth.