新創數據分析:新創成功預測與新創募資金額預測技術

Abstract

全球商業格局因新創的興起而經歷深遠的變革。這些新創在製藥、生物技術、資訊與通信技術及軟體等多個產業中，扮演推動創新與經濟增長的關鍵角色。儘管擁有顯著潛力，新創仍面臨技術不確定性、市場陌生與資源有限等挑戰。因此，對於創業者、投資者及政策制定者等利益關係人而言，預測新創企業的成功與募資金額顯得尤為重要。本論文針對這些研究議題，通過開發有效的新創分析技術，旨在支援創業生態系統中利益關係人的投資決策與策略制定。 本研究第一章闡述研究背景，探討新創在全球經濟中日益增長的影響力以及新創數據分析在其中的重要性。此章為專注於預測高科技新創成功與募資金額的研究重點奠定基礎。 第二章提出創新的新創成功預測方法。該研究不僅利用常見的基本變數（如新創公司概況及募資輪次相關資訊），還融入了與高科技新創密切相關的技術及創投相關變數。研究利用SDC VentureXpert資料庫及USPTO資料庫所收集的4,415個新創案例及其相關變數值，對兩個時間點（即新創成立年份與成立三年後）的預測模型進行評估。結果顯示，將技術和創投相關變數納入模型能顯著提升預測準確性，其中創投相關特徵的影響尤為突出。此外，對深度學習方法的探索研究指出，雖然在新創早期階段使用深度學習（如圖卷積網絡，GCN）自動提取創投相關變數未能提升預測效果，但在後期預測中，由於投資新創的創投增加，深度學習方法表現出相較於傳統統計及機器學習方法的潛在優勢。 第三章將重點轉向新創募資金額的預測。現有文獻主要採用解釋性模型，而非預測性模型，且多數集中於新創早期階段。因此，本研究提出一種基於機器學習的方法，旨在預測高科技新創各輪次的募資金額。此方法綜合考量了基本變數、募資相關變數及創投相關變數，所使用的數據來自VentureXpert資料庫，涵蓋11,365家新創的23,201輪募資。實驗結果顯示，所提方法相比僅依賴基本變數的基線方法具有更佳之預測性能。此外，募資相關變數在預測中展現出比創投相關特徵和基本特徵更高的影響力。 第四章於第三章提出的傳統機器學習方法之基礎上，探討多任務學習框架與深度學習模型在提升新創募資金額預測的效能。本研究利用長短期記憶網絡來處理時間序列數據，以捕捉募資輪次間的時間依賴性，同時採用兩項輔助任務：預測創投的數量和創投的再投資率。該多任務學習框架透過硬參數共享，實現對相關任務同步訓練，藉由知識共享和降低過擬合風險來增強主任務之預測表現。本章納入公司概況、募資輪次資訊、資金相關以及VC相關之變數，評估該模型相較於傳統方法是否提供更優越的預測性能。實驗結果顯示，雖然所提出之模型在均方根誤差（RMSE）與平均絕對誤差（MAE）未具優勢，但其平均絕對百分比誤差（MAPE）較基準模型表現更佳，顯示該模型在相對準確性上的提升。 第五章進行總結並探討其學術與實務意涵。總體而言，本研究透過開發高效的新創成功及募資金額預測模型，為新創數據分析領域做出顯著的貢獻。這些模型不僅提升利益相關人在複雜且多變的創業生態系統中進行決策之準確性與前瞻性，還為策略制定提供重要的依據和洞見。此舉有助於各方更為精準地掌握創業風險與機遇，進而促進創業活動的成功與可持續性發展。

The global business landscape has been significantly transformed by the emergence of startups, which drive innovation and economic growth across diverse industries such as pharmaceuticals, biotechnology, information and communications technology (ICT), and software. Despite their potential, startups often face considerable risks due to technological uncertainties, unfamiliar markets, and limited resources. Consequently, predicting startup success and startup funding size has become crucial for stakeholders, including entrepreneurs, investors, and policymakers. This dissertation addresses these research challenges by developing effective startup analytics methods, aiming to guide investment decisions and facilitating strategy formulation for stakeholders within the entrepreneurial ecosystem. The first chapter establishes the context by discussing the growing importance of startups in the global economy and the critical role of startup analytics. The chapter also sets the stage for the research focus on predicting startup success and startup funding size, particularly within high-tech sectors. Chapter 2 is dedicated to the prediction of startup success. The study leverages not only commonly used basic features (e.g., company profile and funding information) but also integrates the proposed technological and venture-capital-related (VC-related) features that are particularly relevant to high-tech startups. Drawing on a dataset comprising 4,415 startup cases and their corresponding feature values collected from the SDC’s VentureXpert database and the USPTO database, the study evaluates prediction models across two time points (i.e., startup’s founding year and three years after). Results indicate that integrating technological and VC-related features enhances prediction performance, with VC-related features proving particularly influential. Moreover, the exploratory study of the deep learning approach suggests that employing deep learning (e.g., graph convolutional network, GCN) to automatically extract VC features may not improve prediction effectiveness at the very early stages of startups. However, it demonstrates a potential advantage over statistical and machine learning approaches at a later prediction time point due to the increased number of VCs investing in the startups. Chapter 3 shifts focus to the prediction of startup funding size. Existing research primarily adopts an explanatory approach rather than a predictive one, often concentrating on early rounds. To fill these gaps, this study propose a machine learning-based method to predict funding size across various rounds for high-tech startups, by using basic, funding-related and VC-related features. A dataset comprising 23,201 funding rounds from 11,365 startups is collected from the VentureXpert database. The proposed method outperforms the baseline (comprising solely the basic features), suggesting its predictive value. Additionally, the funding-related features emerge as more salient compared to VC-related and basic features. Chapter 4 builds on the traditional machine-learning approach introduced in Chapter 3 by exploring a multi-task learning (MTL) framework with deep learning model to enhance effectiveness of the startup funding size prediction. this study leverages LSTM to handle time-series data, capturing temporal dependencies across funding rounds and integrating two auxiliary tasks: predicting the number of VCs and the VC reinvestment rate. The MTL framework, using hard parameter sharing, allows for simultaneous training on these related tasks, enhancing the primary prediction task through shared knowledge and reduced overfitting risks. By incorporating features related to company profiles, round information, funding-related, and VC-related, the chapter evaluates whether this model offers superior predictive performance compared to traditional methods. The experimental results demonstrate that although the proposed model does not outperform the benchmark in terms of RMSE and MAE, it exhibits better performance in MAPE. This indicates an improvement in the model’s relative predictive accuracy. This dissertation concludes in Chapter 5 with a summary and discussion of contributions. Overall, it contributes to the field of startup analytics by developing effective predictive models of startup success and funding size that empower stakeholders to navigate the dynamic landscape of startup ecosystem with greater precision and foresight.