群眾募資平台動態預測之研究：以Kickstarter平台為例

Research on Dynamic Prediction of Crowdfunding Platform: An Empirical Study of Kickstarter platform

群眾募資是指一個創新的項目透過網路吸引大眾共同參與投資的過程，根據統計資料顯示，群眾募資僅81%可以達到資金目標的20%，其餘19%連目標的20%都達不到。本研究使用C#和Python程式撰寫爬蟲程式，蒐集國際知名Kickstarter群眾募資平台上科技類提案的動態數據，六個月（2019/2/1～2019/7/1）期間，每四個小時蒐集一次資料，然後使用ARIMA、類神經網路、決策樹、支持向量機以及隨機森林建立群眾募資金額動態預測模型，以平均絕對百分比誤差（MAPE）作為預測模型的比較，可提前掌握那些專案較容易達成募資目標。各預測模型結果顯示，使用ARIMA在中高價提案有不錯的準確度，隨機森林在低價提案、中高價提案以及高價提案的表現最佳，而類神經網路則是在中低價提案的預測表現較佳，整體綜合評估仍是以隨機森林的預測表現最好。此一研究成果可以讓群眾募資的投資者對於募資專案的掌握度和可預期程度有所提升，而募資者除了專注於技術開發以及創意發想之外，更應該關注募資專案在平台上的回應以及經營，對於專案的募資成功率將有所提升，而平台經營者可以在網站的功能和服務上進行加值，提供更即時且更精準地提供專案分析和動態訊息服務給投資者和募資者。

Crowdfunding is defined as a project or business process requires investment, and requires a large group of people to provide this investment. In the past few years, this phenomenon has grown exponentially in the index, is seeking ways and means of funding for entrepreneurs and designers. Statistics show that the vast majority of the people have not been successful fund-raising activities, which is only 81% to 20% funding target. Crowdfunding proposal of sorts, in addition to the commodity, content creators, covered the public issues, campaigning, art, invention, design, and scientific research, public disaster reconstruction can initiate fund-raising in fund-raising platform. This study uses C# and Python programs to write web crawlers and collects dynamic data on technology proposals from the world's best-known Kickstarter crowdfunding platform. The data collection period is six months (2019/2/1~2019/7/1). Dynamic data is collected every four hours. After data consolidation and collation, ARIMA, neural network, decision tree, vector support machine, and random forest are used to establish a dynamic prediction model for the amount of funds raised. And the mean absolute percent error (MAPE)of the predicted results is compared as a different prediction model. According to the research results, it can be known in advance that those projects are easier to achieve the goal of fundraising, and the follow-up proposers can put the factors of past fundraising success into their own proposals to improve the success rate of fundraising, and the sponsors can also use This research explores potential innovative commodities and projects, and then invests in the commodity early to develop cooperation and develop markets. Results show that ARIMA model has low prediction error in mid-high price proposal; random forest prediction model has low error in low price proposal, mid-high price proposal and high price proposal; neural network is in the forecast mid-low price proposal. The overall is still the best prediction error random forests than now. The research results can help the investors who raise funds for the masses to improve the mastery and predictability of fundraising projects. In addition to focusing on technology development and creative ideas, fundraisers should pay attention to the response and operation of the fundraising project on the platform, which will improve the project's fundraising success rate. Platform operators can add value to the functions and services of the website, providing project analysis and dynamic information services to investors and fundraisers in a more timely and accurate manner.