IMPROVE ENTREPRENEURIAL FUNDING SCREENING AND EVALUATION:

BUSINESS SUCCESS PREDICTION WITH MACHINE LEARNING

W
IE
A DISSERTATION SUBMITTED TO
EV
THE DEPARTMENT OF MANAGEMENT SCIENCE AND ENGINEERING
AND THE COMMITTEE ON GRADUATE STUDIES
OF STANFORD UNIVERSITY
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
PR

FOR THE DEGREE OF

DOCTOR OF PHILOSOPHY

Chenchen Pan
December 2021
 © 2021 by Chenchen Pan. All Rights Reserved.
Re-distributed by Stanford University under license with the author.

This work is licensed under a Creative Commons Attribution-

Noncommercial 3.0 United States License.
http://creativecommons.org/licenses/by-nc/3.0/us/

W
IE
This dissertation is online at: https://purl.stanford.edu/ps743yv6430
EV
PR

ii
I certify that I have read this dissertation and that, in my opinion, it is fully adequate
in scope and quality as a dissertation for the degree of Doctor of Philosophy.

Edison Tse, Primary Adviser

I certify that I have read this dissertation and that, in my opinion, it is fully adequate
in scope and quality as a dissertation for the degree of Doctor of Philosophy.

Charles Eesley

I certify that I have read this dissertation and that, in my opinion, it is fully adequate
in scope and quality as a dissertation for the degree of Doctor of Philosophy.

W
Yinyu Ye

IE
Approved for the Stanford University Committee on Graduate Studies.
EV
Stacey F. Bent, Vice Provost for Graduate Education
PR

This signature page was generated electronically upon submission of this dissertation in
electronic format. An original signed hard copy of the signature page is on file in
University Archives.

iii
Abstract

As entrepreneurial funding supports the growth of entrepreneurial firms, it can be viewed

as the fuel that enhances the creation, development, and growth of new technologies, industries,

W
and markets. However, investing in entrepreneurial firms is highly risky. For example, Venture
Capitalists (VCs) may receive a large number of business plans/proposals every year but only a few
IE
of them can be successful. Since VCs typically employ a small number of people, they do require a
more e↵ective and efficient screening and evaluation process. Prior research has mainly focused on
identifying evaluation criteria to help VCs predict the business success of these firms. Nevertheless,
EV
relying on VCs’ self-reporting and small regional datasets, these earlier studies have little agreement
on the evaluation criteria. Therefore, it is difficult to employ those criteria to predict business success
in practice. In this work, we propose data-driven approaches using machine learning methods as a
complementary methodology to help VCs predict companies’ success when they screen and evaluate
PR

investment deals. We start by verifying new evaluation criteria with large datasets and then focus
on applying machine learning methods to predict business success. We compare di↵erent machine
learning methods and discuss how VCs can benefit from the prediction. We also apply deep neural
networks and few-shot learning methods to two challenging scenarios faced by the VCs: (1) when
the companies are just founded and don’t have any funding history; (2) when the companies are
from an emerging industry that doesn’t have a lot of historical data to learn from.

iv
Acknowledgments

I feel very fortunate to meet many wonderful people at Stanford who have made my PhD
journey a great experience to grow personally and professionally. Without them, this moment would

W
not be possible.
I would like first to give huge thanks to my advisor Professor Edison Tse. Edison has been
IE
giving me invaluable advice and support since my first day at Stanford. He is an excellent advisor,
teacher, and friend. I thank him for introducing me to the research world, for discussing with me
regularly, and for supporting me to explore diverse methodological approaches. He has o↵ered me
EV
a lot of guidance ranging from research directions to life advice. I truly learned and benefited a lot
from him.
I would also like to thank Professor Charles Eesley. He has opened my eyes to the field of
entrepreneurship and innovation. I appreciate his vision, and really enjoy the insightful discussions
PR

with him. In addition, I am indebted to Professor Yingyu Ye for his valuable advice and comments.
His optimization classes help me a lot when I explored machine learning methods. I am grateful to
Professor Ramesh Johari for serving on my committee. My thanks also go to Professor Tze Lai for
kindly chairing my dissertation defense.
My deep thanks go to the members and alumni of the Department of Management Science
& Engineering (MS&E). I would like to thank June Lee, You Wu, Xinyi Yang, and Xianling Long
for sharing your encouragement and wisdom. Thanks to Lori Cottle for all the help. I feel grateful
to all my close friends in and outside of MS&E who went through the ups and downs with me and
left many amazing memories in my life. Thanks to my mentors and peers at Microsoft for making
the internship a great experience.
Finally, I am deeply grateful to my family. I want to thank my husband, Chen Liang, for
always supporting me and being with me through the best and worst times. Thanks to my mother,
Tong Wu, for everything she has done for me.

v
Contents

Abstract iv

W
Acknowledgments v

1 Introduction IE 1

2 Background 6
2.1 Entrepreneurial Funding Landscape . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
EV
2.1.1 Traditional Forms of Entrepreneurial Funding: VC, CVC, and Angel Investor 7
2.1.2 Emerging Forms of Entrepreneurial Funding: Crowdfunding, Incubators and
Accelerators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 Investment Decision Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
PR

2.2.1 Stages in VCs Decision Process . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2.2.2 Screening and Evaluation Process . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.3 Issues in Evaluation Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3.2 Supervised Classification Algorithm in Machine Learning . . . . . . . . . . . 16
2.3.3 Deep learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.4 Few-shot Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3 Study of the Incubators’ Impact 28

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.3 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

vi
 3.4 Data Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.6 Conclusion and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

4 Prediction of Companies’ Business Success 43

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.2 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.3 Dataset and Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.3.1 Data Cleaning and Labeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.3.2 Feature Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.3.3 Feature Exploration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

W
4.4 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.4.1 Logistic Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.4.2
4.4.3
IE
Random Forest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
K-Nearest Neighbors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
53
55
4.4.4 Selected Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
EV
4.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.5.1 Pre-processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.5.2 Hyperparameter Tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
PR

4.6.1 Results Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

4.6.2 Implications for VCs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.6.3 Feature Importance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.7 Conclusion and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

5 Prediction of Crowdfunding Projects’ Success 67

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.2 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.3 Dataset and Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.3.1 Data Labeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.3.2 Feature Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.4 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.4.1 Shallow Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

vii
 5.4.2 Recurrent Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.5.1 Pre-processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.5.2 Hyperparameter Tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.6 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.6.1 Ablation Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.6.2 Error Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.6.3 Implications for VCs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.7 Conclusion and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

6 Few-Shot Business Success Prediction 83

W
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.2 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.3
6.4
IE
Dataset and Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
88
90
6.4.1 Baseline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
EV
6.4.2 Joint Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
6.4.3 First-Order MAML/FOMAML . . . . . . . . . . . . . . . . . . . . . . . . . . 92
6.4.4 Reptile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
PR

6.6 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

6.6.1 Implications for VCs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
6.7 Conclusion and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

7 Conclusion 98
7.1 Main Findings and Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
7.2 Limitations and Future Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

A Loss Function and Gradient Descent 102

A.1 Loss Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
A.2 Gradient Descent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

B List of Figures 106

viii
List of Tables

2.1 Stages of Venture Capitalists’ Management Process . . . . . . . . . . . . . . . . . . . 10

2.2 Comparison of Venture Capital Evaluation Criteria . . . . . . . . . . . . . . . . . . . 13

W
2.3 Comparison of Machine Learning Technologies . . . . . . . . . . . . . . . . . . . . . 16

3.1 3-stage Development of Incubation Models . . . . . . . . . . . . . . . . . . . . . . . .

IE 31
3.2 Value-added View - Incubatee’s Point of View (Ewan, 2010) . . . . . . . . . . . . . . 34
3.3 3-level Incubator Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
EV
4.1 Dataset split and up-sample . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.2 Setting of hyperparameters tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.3 Summary of hyperparameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.4 Metrics Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
PR

5.1 Dataset Split . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

5.2 The Setting of Hyperparameters Tuning . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.3 Summary of Hyperparameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.4 Experiment Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.5 Ablation Study: Compare Textual and Metadata Features . . . . . . . . . . . . . . . 80
5.6 Compare Results of Di↵erent Training Size . . . . . . . . . . . . . . . . . . . . . . . 81

6.1 Summary of Selected Hyperparameters . . . . . . . . . . . . . . . . . . . . . . . . . 94

6.2 Experiment Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

ix
List of Figures

2.1 Process of Supervised Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2.2 Supervised Classification Algorithm Workflow . . . . . . . . . . . . . . . . . . . . . . 18

W
2.3 Structure of A Single Node . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.4 Structure of MLP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.5
2.6
IE
Multilayer Perceptron Forward Pass and Backpropagation . . . . . . . . . . . . . . .
Unrolled Structure of RNN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
22
23
2.7 Caption for LOF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
EV
2.8 Caption for LOF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.9 Caption for LOF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

3.1 4 Key Practices Provided by Value-added Incubators (Campbell, 1985) . . . . . . . . 32

PR

3.2 Resource Flow (Chandra & Chao, 2011) . . . . . . . . . . . . . . . . . . . . . . . . . 33

3.3 Geographic Distribution of Successful Companies . . . . . . . . . . . . . . . . . . . . 38
3.4 Industry Distribution of Successful Companies . . . . . . . . . . . . . . . . . . . . . 39
3.5 IPO VS Acquired . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.6 Di↵erence in Maturity Period of Incubated and Non-incubated Companies . . . . . . 40
3.7 Delta MP for Di↵erent Industries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.8 Incubated and Non-incubated Successful Companies Comparison . . . . . . . . . . . 42

4.1 Imbalanced dataset: 1 = IPO or acquired, 0 = closed or operating . . . . . . . . . . 49

4.2 Selected Features and Corresponding Examples . . . . . . . . . . . . . . . . . . . . . 51
4.3 Relationship of Numerical Features and Target Labels . . . . . . . . . . . . . . . . . 52
4.4 Random Forest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.5 K-Nearest Neighbors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.6 ROC Curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

x
4.7 Confusion Matrix of Random Forest . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.8 Confusion Matrix of KNN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.9 VCs Decision Boundary and ROC Curve (Scenario 1) . . . . . . . . . . . . . . . . . 63
4.10 VCs Decision Boundary and ROC Curve (Scenario 2) . . . . . . . . . . . . . . . . . 64
4.11 Feature Importance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

5.1 Final Status of Projects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

5.2 Data Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.3 Shallow Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.4 RNN Model Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

W
6.1 Meta-learning vs Joint Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.2 Industry Distribution: Original and Log Y . . . . . . . . . . . . . . . . . . . . . . . . 88
6.3 Data Preview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
IE 89
6.4 Meta-learning Workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
6.5 Data Split for Few-Shot Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
EV
6.6 An Example of Industry: FOMAML Consistently Outperform Joint Training . . . . 96

A.1 An Example of Sigmoid Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

B.1 Industry Rank Snapshot for CrunchBase Data . . . . . . . . . . . . . . . . . . . . . 107

PR

B.2 Industries: FOMAML Outperform Joint Training . . . . . . . . . . . . . . . . . . . . 108

B.3 Industries: Joint Training outperform FOMAML . . . . . . . . . . . . . . . . . . . . 109

xi
Chapter 1

Introduction

W
Entrepreneurial firms are usually viewed as the backbone of economies and drivers of both
IE
economic development and employment since young and innovative entrepreneurial firms are closely
related to the creation, development, and growth of new technologies, industries, and markets and
create the most jobs (Megginson, 2004). Entrepreneurial funding, such as venture capital (VC),
EV
works as the fuel to help these high-growth-potential ventures by providing financial support. Thus,
successful investment of entrepreneurial funding can benefit not only the investors but also the
economy and society. However, millions of companies are emerging around the world each year.
Among them, some become successful via acquisition or initial public o↵ering (IPO) in that they can
PR

provide high returns to the entrepreneurial funding investors, while others may vanish and disappear
so the investors waste all their money. Better understanding and a more accurate prediction of the
business success of an entrepreneurial firm become very crucial.
Previous research has focused on identifying the evaluation criteria to help investors, such
as venture capitalists (VCs) to predict business success, however, these earlier studies have little
agreement on the evaluation criteria. The limitations, such as the unequal influence of respondents,
problems of retrospective reporting, and VC self-reporting (MacMillan et al., 1987; Robinson Jr,
1987; Tyebjee and Bruno, 1984), let the previous research open to errors and biases. For example,
it has been found that relying on VCs’ self-reporting to investigate how they make decisions and
what criteria they use was not reliable (Sandberg et al., 1989; Shepherd, 1999; Zacharakis and
Meyer, 1998), since VCs are not good enough in self-introspection and have limited insight into their
decision-making process, thereby are not good at predicting the investment outcomes (Khan, 1987;
Pfe↵er, 1989; Sandberg et al., 1989; Shepherd, 1999; Zacharakis and Meyer, 1998). In addition,

1
CHAPTER 1. INTRODUCTION 2

these studies only rely on small and regional datasets which leads to some conflicting conclusions
on evaluation criteria. Therefore, it is difficult to employ those criteria to predict business success
in practice. In this work, we proposed data-driven approaches using machine learning methods as a
complementary methodology to help VCs screen and evaluate companies’ success.
In Chapter 2, We o↵er a brief background review. It start by reviewing the landscape and
introducing key concepts of entrepreneurial funding. In addition, we briefly introduce the process
and di↵erent stages when VCs makes an investment decision. Specifically, we focus on how VCs
make their decision when screening and evaluating investment opportunities and illustrate some
issues in the previous research that we attempt to overcome in this dissertation. Finally, we describe
the machine learning methods that are employed in this dissertation.

W
In Chapter 3, we focus on successful companies and use a recent large dataset of real
companies, CrunchBase, to analyze a new evaluation criterion for VCs. Over the past decades, a
wide variety of incubators have been introduced to support and accelerate the creation of successful
IE
ventures. Therefore, we attempt to verify whether backed by an incubator and/or accelerator can
be a useful evaluation criterion to predict the business success of a venture. By studying the impact
of incubators and/or accelerators, we find that (1) on average, incubators can accelerate the devel-
EV
opment of companies by 25%, around 1 year, in terms of the time to be acquired or IPO; (2) some
industries, such as Location-Based Services and FinTech, can significantly benefit from incubators
and/or accelerators because the percentage of incubated companies is higher and the time to succeed
PR

is shorter than the average level; while some industries, such as BioTechnology, doesn’t benefit much
from it. Based on these findings, we can conclude that whether the venture has been backed by an
incubator or accelerator is an important factor for VCs to make a better decision. However, this
factor alone cannot be a reliable evaluation criterion to predict business success since companies
from di↵erent industries benefit di↵erently from it.
The rest of this dissertation focuses on improving VCs’ screening and evaluation pro-
cess that is the first step of VCs’ decision-making process. Given the large number of business
plans/proposals VCs may receive every year and the small number of people they typically employ,
VCs do require a more e↵ective and efficient screening and evaluation process (Golis, 1998). As a
data-driven approach, we apply machine learning methods to large datasets to better predict the
business success. We first introduce a study in Chapter 4 where we compare di↵erent machine learn-
ing methods and discuss how VCs can benefit from the prediction. We then move on to investigate
two challenging scenarios faced by the VCs in Chapter 5 and 6: (1) when the companies are just
CHAPTER 1. INTRODUCTION 3

founded and don’t have any funding history; (2) when the companies are from an emerging industry
that doesn’t have a lot of historical data to learn from.
In Chapter 4, we use CrunchBase data and attempt to apply machine learning methods to
predict business success which can help VCs quickly screen and evaluate a large number of ventures.
Specifically, we use the company’s basic profile information, such as industry category, and the
company’s funding history, such as the number of investors, to predict the company’s success. In
addition to accuracy, we introduce the area under receiver operating characteristic curve (ROC AUC
score), which is independent of the classifier decision threshold, as the metric to compare di↵erent
machine learning models and find that K-Nearest Neighbor (KNN) model achieving a 79.89% ROC
AUC score and 73.33% accuracy has better overall performance than the Logistic Regression and

W
Random Forest model across di↵erent classifier thresholds. This result demonstrates the potential of
this data-driven approach to help VCs screen and select a subset of promising companies. In addition,
we propose a method to help VCs choose the best model in terms of the expected investment return.
IE
For example, if the VCs have the knowledge of the average profit of the successful investment and the
average cost of the non-successful investment, they can calculate the expected investment returen
to determine their decision boundary, and then use the di↵erence between the decision boundary
EV
and the model’s ROC curve to find out which model can achieve the best performance and the
corresponding threshold. Last, by analyzing the feature importance for di↵erent models, we find
that the company’s funding history is more important than the company’s profile information for
PR

predicting its business success, and more investors’ endorsements may lead to a positive prediction
on the company’s success. The intuition behind these findings is that the funding history contains
(1) the extra information provided by previous investors via their evaluation and due diligence
process, and (2) the added values contributed by previous investors when they mentor and monitor
the business health.
While the funding history of a company is an important factor to predict business success,
many new companies are just founded and have no funding history. Therefore, VCs also have to
screen and evaluate a large number of companies without funding history. This situation leads us
to the question: How to predict a new company’s success without funding history? In practice,
VCs usually rely on business plans/proposals to make investment decision under this situation.
Inspired by this, we study the prediction of business success with business description information.
Since CrunchBase data doesn’t include any business description, we instead use the Kickstarter
dataset that contains the information of more than 100K crowdfunding projects with the metadata
CHAPTER 1. INTRODUCTION 4

information, such as fundraising goals and the country, and the textual information, such as project
description where the project’s owner summarizes their idea, the plan of how to use the funds, and
other important information of the project in a piece of text. Chapter 5 focuses on the pre-posting
prediction, i.e., predicting the success of the project without any funding history, and explores the
deep learning methods to better utilize the textual information. Specifically, we employ shallow
neural network (SNN) and recurrent neural network (RNN) to improve the prediction accuracy. In
particular, we apply the sequential model, Long short-term memory (LSTM), on the textual data and
use the pre-trained word embeddings to initialize word representations/vectors and achieve 73.03%
accuracy. In addition, our ablation study shows that textual features, such as project description, are
more important than all the other metadata information for predicting the project’s success. Despite

W
the di↵erences between crowdfunding and VC investments, the crowdfunding project description
plays a similar role as the startups’ business plan/proposal that describes the business idea and
other key information for the investors, especially during VCs’ screening when the business plan
IE
is the main document for communicating such information. Our findings implies that the business
plans contain important information for predicting the business success and the appropriate machine
learning methods can extract and make use of such information.
EV
The two chapters above reveal the potential that machine learning and deep learning
methods can predict the business success reasonably well. However, to obtain a good prediction
accuracy on new examples, these methods typically require that a large amount of training data is
PR

available and that the new examples are from the same distribution as the training data. These
requirements imply that the models usually perform well in the developed industries where there
are enough data samples to learn from but may fail in the emerging industries where there are only
a few companies can be learned. In practice, VCs pay more attention to the prediction for emerging
industries since the emerging industries usually contain more high-risk high-return opportunities.
This fact leads us to another challenging question: how to predict the company’s success in emerging
industries? This question has been rarely touched in previous studies. we will initiate our approaches
to address this question in Chapter 6. Concretely, we define a few-shot business success prediction
task and create a corresponding dataset using the companies from the CrunchBase dataset. Our
analysis shows that the meta-learning methods can not only capture the uniqueness of the emerging
industries but also learn the shared trends across di↵erent industries. Specifically, we find that First-
Order Model-Agnostic Meta-Learning (FOMAML) method can consistently outperform in some
industries, such as Online Video Advertising, which contain unique companies that behave di↵erently
CHAPTER 1. INTRODUCTION 5

from the majority of companies in other industries. It implies that meta-learning model can better
predict the business success in disruptive industries.
Finally, Chapter 7 summarizes our findings and contributions, discusses the possible limi-
tations to take into consideration if VCs adopt the data-driven approach in real investment decisions,
and recommends future areas of study.

W
IE
EV
PR
Chapter 2

Background

W
In this chapter, we first overview the landscape of entrepreneurial funding. We briefly
IE
introduce the traditional forms of entrepreneurial funding, such as venture capital (VC), corporate
venture capital (CVC), and angel investment, and the emerging forms of entrepreneurial funding,
such as crowdfunding and accelerators. Second, we introduce the decision making process of en-
EV
trepreneurial funding. Specifically, we focus on the screening and evaluation process and point
out some limitations in the previous research. Finally, we briefly introduce some machine learning
methods that we will apply in this work.
PR

6
CHAPTER 2. BACKGROUND 7

2.1 Entrepreneurial Funding Landscape

Entrepreneurial funding primarily includes venture capital (VC), corporate venture capital
(CVC), angel investors, crowdfunding, incubators and accelerators. Compared to VC, CVC, and
angel investors, crowdfunding, incubators and accelerators are a set of relatively “new” sources of
financing which have emerged in recent years (Bruton et al., 2015). While both the “traditional”
and the “new” entrepreneurial funding represent the key sources of capital that fuels innovation
and development, these two types of funding source are di↵erent on the stage of investment focus,
amounts invested, strategic objectives, geographic concentration, and the nature of involvement
beyond the provision of capital (Drover et al., 2017). Based on this distinctiveness, this section will
o↵er a brief overview for each type of funding source.

W
2.1.1 Traditional Forms of Entrepreneurial Funding: VC, CVC, and An-
gel Investor
IE
Traditional entrepreneurial funding typically formed by equity investors who trade capital
for a portion of company ownership. Over the last decades, entrepreneurial funding literature
EV
has emphasized the importance of entrepreneurial equity funding which includes venture capital
investors, corporate investors and business angels, since it shows that new entrepreneurial firms
heavily rely on these traditional external equity finance. Compared with the other two types of
PR

investors, VC receives much more discussions by the scholars across di↵erent disciplines.
VC tends to be the most common form of entrepreneurial equity financing, while only
funding a small portion of startups. Venture capitalists (VCs) raise money from a set of limited
partners (e.g., university endowments, pension funds, etc.) and aim to give a return to these partners
by investing a portfolio of young, innovative companies (Gompers and Lerner, 2000). VC firms
are typically small, geographically clustered entities, often working closely with the ventures that
they invest to provide guidance and value beyond capital (Sapienza, 1992; Sørensen, 2007). VCs
commonly participate in mid-stage to late-stage investment activities, and continue drifting to larger
and later-stage investments (Hellmann and Thiele, 2015). Because VCs provide returns to their
limited partners within 10 or so years, there is often a focus on realizing timely exits via an acquisition
or initial public o↵ering (IPO) (Drover et al., 2017).
CVC denotes the investment vehicles established by corporations making equity invest-
ments in entrepreneurial ventures. CVC is distinct from traditional VC in that funds are invested
CHAPTER 2. BACKGROUND 8

by arms of corporations as extensions of their primary focus (e.g., Google Ventures and Microsoft
M12). Corporate investors are usually looking to bring long-term value to entrepreneurial firms and
focus more on early-stage to mid-stage ventures. They usually have a significant impact on invested
companies by providing capital, complementary assets, shared industry knowledge, and access to
customers as well as on shaping innovation strategies for their own firms (Drover et al., 2017).
Angel investors (also known as private investors, seed investors) are accredited high-net-
worth individuals who provide financial backing for young ventures with their own personal capital.
Angels are often former entrepreneurs who seek to invest and add value/guidance to investee firms in
their area of expertise. Such individuals tend to take a less formal approach to investing, so they are
often found among an entrepreneur’s family and friends. They usually invest at the earliest stage

W
of the venture life cycle (Drover et al., 2017). Most recently, angels have enhanced their impact
by forming angel investor groups (Kerr et al., 2014) and by providing platforms, both online and
in-person, for individual angels to evaluate and invest in high-potential deal flow collectively (e.g.,
IE
Tech Coast Angels).

2.1.2 Emerging Forms of Entrepreneurial Funding: Crowdfunding, Incu-

EV
bators and Accelerators

In recent years, crowdfunding platforms, incubators and accelerators emerge as the early-
stage entrepreneurial funding sources alternative to VC, CVC and angel investors. Unlike the tradi-
PR

tional o✏ine funding, crowdfunding activities take place on internet-based marketplaces, including
equity-based platforms, lending platforms, reward-based platforms, and charity platforms.
While the majority of existing entrepreneurial funding studies focus on VCs, only a small
proportion of startups receive VC funding. Instead, a larger number of firms receive fundings from
incubators, accelerators and crowdfunding platforms. As accelerator can be viewed as a new form
of incubators, we can combine these two forms together. While most incubator/accelerator studies
examine the e↵ects of accelerators on venture performance (Cohen and Hochberg, 2014; Gonzalez-
Uribe and Leatherbee, 2016; Hallen and Pahnke, 2016; Yu, 2016), crowdfunding studies mainly focus
on the investment process (Burtch et al., 2015; Gorbatai and Nelson, 2015; Greenberg and Mollick,
2017; Mollick and Nanda, 2016).
Crowdfunding platforms are internet-based marketplaces that connect entrepreneurs
seeking funding with a large number of supporters, each providing a small fraction of the total
amount required (Fleming and Sorenson, 2016). Specifically, crowdfunding activities take place on
CHAPTER 2. BACKGROUND 9

four types of platforms, including equity-based platforms(e.g., AngelList), lending platforms (e.g.,
Lending Club), reward-based platforms (e.g., Kickstarter and Indiegogo), and charity platforms (e.g.,
Causes). Equity-based crowdfunding can be understood as the VC and angel investments online,
but the syndication structure and lead investors’ incentives on crowdfunding platforms tend to di↵er
from traditional VC syndication or angel groups (Agrawal et al., 2014). While equity crowdfunding
initially faced significant legal challenges, it is experiencing rapid growth as legal barriers are being
relaxed in many countries (Ahlers et al., 2015). Compared with the other three types, reward-based
crowdfunding receives much more discussions in entrepreneurial financing research.
Incubators/Accelerators are service-based, fixed-term programs that provide ventures
with a configuration of mentorship, work space, and/or a small amount of funding, often in ex-

W
change for equity (Cohen and Hochberg, 2014; Hallen et al., 2016). The capital provided typically
is o↵ered at the very earliest stages. The core value provided by incubators/accelerators is the
educational benefits which are mainly driven by indirect learning from the advice and experience
IE
of others (Gonzalez-Uribe and Leatherbee, 2016; Hallen et al., 2016). During the program, en-
trepreneurs apply for an opportunity to develop (or “accelerate”) their concepts and/or products
on site within a fixed time period (1-3 years for incubators; 3–6 months for accelerators). Those
EV
accepted companies are provided with an immersive experience that equips entrepreneurs with many
of the essential knowledge required to succeed (Hathaway, 2016). The programs usually close with a
“demo day” or ”graduation day”, where concepts and/or products are pitched to potential investors
PR

and stakeholders.

2.2 Investment Decision Process

In this section, we will o↵er a review of the research related to the decision making pro-
cess of the entrepreneurial funding, also known as the investment decision process. As VC plays the
most important role in identifying high-growth and innovative companies, it has become a prominent
branch of entrepreneurial finance and the typical form of entrepreneurial funding. The decision mak-
ing process of venture capitalists (VCs) has emerged as an important area of research. Since “VCs
are real-world entities that closely approximate the investors of theory” (Kaplan and Stromberg,
2001), we will focus on the investment decision process of VCs, and specially focus on the screening
and evaluation phases. We will start with introducing di↵erent stages of the decision-making pro-
cess, and then address the criteria used by VCs in screening and evaluating investment proposals.
At last, we will identify some issues and limitations in this area.
CHAPTER 2. BACKGROUND 10

2.2.1 Stages in VCs Decision Process

There is a large amount of academic literature on how VCs make their investment decisions
(Ferris and Bill, 2000; Hall and Hofer, 1993; Khan, 1987; MacMillan et al., 1985, 1987; Poindexter,
1976; Robinson Jr, 1987; Sandberg et al., 1989; Shepherd et al., 2003; Timmons et al., 1987; Tyebjee
and Bruno, 1984; Wells, 1974; Zacharakis and Shepherd, 2007; Zacharakis and Meyer, 1995, 1998).
It was found that the VCs’ decision-making process includes a series of activities that starts with
the proposal of a new venture and continues until the successful exit with adequate returns (Sharma
et al., 2015), and that the VCs’ decision-making process consists of multiple stages (Hall, 1989;
Tyebjee and Bruno, 1984; Wells, 1974). Hall and Hofer (1993) summarized the related findings in
previous research and showed that the process commonly consisted of five or six stages (shown in

W
Table 2.1).

Table 2.1: Stages of Venture Capitalists’ Management Process

IE
Process Wells Tyebjee & Bruno Silver Hall
Stage (1974) (1984) (1985) (1989)
1 Search Deal origination Search Generating a deal flow
EV
2 Screening Screening Initial screening Proposal screening
X X X Proposal assessment
3 Evaluation Evaluation X Project evaluation
X X Due diligence Due diligence
4 X Deal structuring Deal structuring Deal structuring
Venture board meetings
5 Post-investment activities Monitor progress Venture operations
and operations
PR

6 Cashing out X Cashing out Cashing out

Source: Hall & Hofer 1993, Table 2

As Table 2.1 shown, in general, we can classify VCs’ decision-making process as the fol-
lowing six stages: (1) deal origination; (2) initial screening; (3) evaluation and due diligence; (4)
deal structuring; (5) post-investment activities; and (6) cashing out or exit activities (Sharma et al.,
2015).

• Deal origination is the first stage of the process and refers to generating a deal flow, in which
VCs search and establish good sources of investment opportunities.

• Initial screening involves a quick review of a large amount of investment proposals or business
plans they received. In this stage, VCs delimit investment prospects to a manageable few that
are worthy for a closer evaluation.

• Evaluation and due diligence refers to the in-depth assessment of the selected deal, in
CHAPTER 2. BACKGROUND 11

which VCs investigate the potential investee carefully and determine whether or not to invest.

• Deal structuring involves the negotiation and establishing an agreement with the entrepreneur
after the VCs decide to invest in a particular venture. A ‘Term Sheet’, as the outcome of this
stage, a document summarising all the terms and conditions, will be signed by both the VCs
and the entrepreneur.

• Post-investment activities include, but not limited to, mentoring, monitoring, participating
in the board meetings and other management operations. In this stage, VCs usually provide
more resources and attempt to add value to the business.

• Cashing out or exit activities is the final stage of the process, in which the VCs end their

W
involvement and sell the investment to obtain a return. VCs usually play an active role in
directing the venture towards an exit strategy, such as a merger and acquisition (M&A), or
initial public o↵ering (IPO), so that they can maximize their return of the original investment.
IE
According to this staged approach to the VCs decision-making process, the previous stud-
ies categorize what VCs do for each stage (Sahlman, 1990), and identify the distinction between
EV
screening and evaluation, in which VCs could use di↵erent criteria to make their decisions (Boocock
and Woods, 1997; Fried and Hisrich, 1994; Hall and Hofer, 1993; Riquelme and Rickards, 1992). In
the next, we will focus on the screening stage (stage 2) and evaluation stage (stage 3), and specifically
address some issues in the early research.
PR

2.2.2 Screening and Evaluation Process

A relatively large number of studies has concentrated on how VCs screen and evaluate
the new ventures and what criteria they use to make the decision. It was found that, in broad
terms, VCs use the same criteria in both screening and evaluation stages, while VCs have di↵erent
focuses for each stage. The evaluation criteria used in these two stages is wide-ranging, and can be
summarized as market size and potential growth, a significant competitive advantage, uniqueness of
the product/service, management team, funding requirement, investment stage, the industry, profit
potential and economic return (Golis, 1998). The screening phase focuses on a small subset of these
criterias and follows a non-compensatory rule, which means an unacceptable value on one decision
variable cannot be o↵set by a higher value on another (Evans and Hudson, 2005). During this phase,
VCs make a quick ‘Go/No Go’ decision in around 6 minutes and the average time spent on proposal
assessment is less than 21 minutes (Hall and Hofer, 1993).
CHAPTER 2. BACKGROUND 12

While using the same criteria, the evaluation (due diligence) stage, unlike screening stage,
allows the compensatory rule, in which a low value on one criteria can be o↵set by a higher value
on another (Riquelme and Rickards, 1992). Moreover, this stage involves a greater in-depth analysis
with less subjectivity, and assesses the proposal more on financial aspects (Golis, 1998; Riquelme
and Rickards, 1992). Therefore, VCs take an average of 97 days during this stage, ranging between
52 days and 142 days (Fried and Hisrich, 1994).
The previous researchers attempted to discover the specific criteria used by VCs during
the screening and evaluation stage. However, they use di↵erent methods and datasets to approach
this question, which lead to a poor convergence of findings and unreplicated results. Zacharakis and
Meyer (1998) compared the di↵erent investment criteria found by major researchers and summarized

W
them in Table 2.2.

IE
EV
PR
 Table 2.2: Comparison of Venture Capital Evaluation Criteria
Wells Poindexter Tyebjee & Bruno MacMillam et al. MacMillam et al. Robinson Timmons et al. Hall & Hofer
Evaluation Criteria
(1974) (1976) (1984) (1985) (1987) (1987) (1987) (1993)
Personal Phone Survey Unstructured Verbal
Methodology Questionnaire Questionnaire Questionnaire Questionnaire
Interviews & Questionnaire Interviews Protocols
Sample Size 8 97 87 100 67 53 47 16
Characteristics of the Entrepreneur
- Management Skill and Experience X X X X X X X X
- Venture Team X X X X
CHAPTER 2. BACKGROUND

- Management Stake in Firm X X

- Personal Motivation X X
PR
- Entrepreneur Personality X
Product/Service Characteristics
- Product Attributes X X X X X
- Product Di↵erentiation X X
- Proprietary X X X
- Growth Potential X
- Market Acceptance X X
- Prototype X
EV
Market Characteristics
- Market Size X X X X
- Market Growth X X X X X X
- Barriers to Entry X X
- Competitive Threat X X X

Reproduced with permission of copyright owner. Further reproduction prohibited without permission.
IE
- Venture Creates New Market X
Financial Characteristics
- Cash-out Method X X X
- Expected ROI X X X X
- Expected Risk
- Percentage of Equity
X
X
W
- Investor Provisions X
- Size of Investment X X
- Liquidity X X X
OTHER
- References X X
- Venture Investment Stage X X X
- Venture Capitalist Criteria X
Source: Zacharakis and Meyer (1998)
13