The role of A.I. in Fintech
Master in International Finance Amsterdam Business School
Studentnumber: 10839569 Author: Manolito Gobind Supervisor: Torsten Jochem
Date:31st of August 2017
"The
best way to predict the future is to invent it."
, Alan KayIndex
Introduction 3
Chapter 1: Fintech 4
Finance and technology 4
The Fintech ecosystem 5
Chapter 2: Artificial Intelligence (A.I.) 6
A.I. methods 9
Human expertise versus Algorithms 11
A.I. Risks and safety issues 13
Chapter 3: Fintech and A.I. 18
Payments and savings 19
Market research and sentiment analysis 20
Credit scoring and direct lending 20
Insurance 21
Business finance and expense reporting 21
General purpose and predictive analytics 22
Regulatory, compliance, and fraud detection 22
Asset management 23 Chapter 4: Discussion 25 Chapter 5: Conclusion 28 Future Directions 28 References 29 2 of 31
Introduction
The financial industry is changing fast. The arrival of personal computers, smartphones and the internet have opened up new doors for financial services. These innovations have improved accessibility, transparency, speed and cost reductions. (Pan, 2016) (Makridakis, 2017) The financial services industry has high expectations from the next wave of innovations driven by artificial intelligence. (Upchurch, 2017) This thesis sets out to understand the added value of artificial intelligence in its current state, and its future outlook for the financial industry. Chapter 1 provides an overview of the financial industry, its main innovation drivers, and its ecosystem. In Chapter 2 we discuss the different perspectives on A.I. to (1) better understand the core concepts behind this technology, and (2) the added value, that can be created, by exploring its general benefits and disadvantages. In addition we address concerns about trust, privacy, liability and regulations. Chapter 3 provides an overview of Fintech companies that indicate they have adopted A.I. technologies. We discuss how the A.I. technology is adding value, compared to their non-A.I. driven counterparts. Chapter 4 discusses the findings and provides insights about the challenges ahead for Fintech and the adoption of A.I.
Chapter
1: Fintech
The purpose of this chapter is to provide an introduction to the Fintech industry. The chapter starts with a definition of Fintech. We then continue with an overview of financial products throughout our lives, to emphasize our requirements for a diversity of financial products and services. Finally, we discuss the Fintech ecosystem to understand the different stakeholders that are participating in creating new Fintech solutions.
Finance
and technology
Fintech refers to startups, technology companies, and legacy banks that operate on the intersection of finance and technology. (PWC, 2016) They provide the game changing disruptive and innovative services we utilize (Lee, 2016) to pay for products and services, send money to others, borrow, lend, and invest. Their focus is on offering trust, transparency, technology, information, analytics, and advice. (Menat, R., in Chrishti et al, 2015) The way we handle our finances has been changing dramatically since the introduction of computers, smartphones and the internet. (Pan, 2016) These financial and technological innovations bring convenience in ways we have never seen before. It should be no surprise that the preferred research method for millennials is using the web to find, compare and buy products and services. (Waupsch, 2017) An overview of typical financial service requirements, in a person’s life, is presented in table 1.
Life phases Financial services
Birth to 5 (Preschool) Parents start saving for the child’s future and education.
5 to 11 (Elementary School) money management through allowances.
12 to 15 (Middle School) spend money individually using mobile phones, Internet, Facebook
15 to 18 (High School) having jobs and pay taxes, but financially dependent on parents
18 to 23+ (College Students) private financial control, integration with parents for transfers.
24+ (First Full-Time Job) investments, mortgages, retirement products, loans, savings for particular purposes such as travelling.
Table 1: Financial services typically required in a person’s life
Research shows it is increasingly important to focus on customer behavior that tailor to the requirements from customers. This will allow companies to create effective financial products and services. (Fünfgeld, 2009) The Fintech ecosystem caters to these needs and will be discussed in the next paragraph.
The
Fintech ecosystem
The financial products and services industry has been transformed by 5 main drivers: (1) the changing role of information technology, enables the convergence of digital technologies such as social computing, big data, cloud computing, the Internet of Things (IoT) (Puschman, 2017). The result is a new generation of products and services, which are addressing particular customer segments in the financial services long tail. These innovations are (2) changing customer behavior and their preferences to use digital channels over traditional local bank branches, because of 24/7 availability, self servicing functionality, (Puschman, 2017) offering more personalized services, lowering operating costs, and targeting more niche markets. (Lee, 2016) On the governmental side, the financial crisis caused major (3) changes in regulations such as an increase in policies and legislation. This increased the complexity of regulations and difficulty of introducing new product and services. But at the same time, many governments also introduced initiatives to support Fintech startups by: creating experimental hub environments, providing funding, and assisting in market development. (Gnirck et al., in Christhi et al, 2017) Another important driver has been the (4) changing financial ecosystem; As an example in the past banks developed their proprietary banking solutions inhouse. These complex legacy systems make it difficult for established banks to innovate. (Dudley et al., 2017) Nowadays, there is a clear trend of outsourcing and partnering with specialized third parties. In addition, banks partner with companies outside the Fintech industry. Some examples are telecom operators for handling mobile payments (Puschman, 2017), social media experts for market research, and encryption specialists for digital security. (Sironi, 2016) These developments have contributed to the expansion of participants in the Fintech ecosystem (see figure 1).
Figure 1: Fintech Ecosystem (Lee, 2016)
The Fintech ecosystem is expected to be disrupted again. This time the disruption is driven by A.I. technology (Upchurch, 2017). The rise of artificial intelligence technology will be discussed in the next chapter.
Chapter
2: Artificial Intelligence (A.I.)
Artificial Intelligence (A.I.) is a controversial topic because it is aimed at substituting, supplementing and amplifying all human tasks (Makridakis, 2017) This chapter discusses the origin of A.I. and the different perspectives on adopting A.I. technologies. The process of A.I. is explained and the short and long term expectations from researchers are discussed. The chapter is concluded with a discussion on social themes in A.I.
To understand the origins of A.I. technology, we have to go back in history. Alan Turing was the first to build a computer in 1936. The computer was used to decipher German military messages which were encrypted using Enigma machines. The Enigma encryption algorithm was thought to be the most secure coding scheme available and unbreakable by humans. The German army, under the command of Hitler, used the Enigma machine to encode the communication of its secretive strategic plans. The Allies tried to decipher the Enigma code manually, but failed. The reason being that there were seemingly endless combinations, and each day the encryption key was changed by the Germans. This was until Turing was able
to decipher the enigma code. He did so, not by trying to decode each message by hand, but by building a machine that could automatically decipher these messages. Turing showed that his Universal Turing Machine (UTM) can in principle perform any mathematical computation. (Boden, 2016) The main breakthrough was achieved when his machine deciphered a fixed greeting pattern at the end of each message. This was just a first step towards solving real world challenges with computers instead of the human mind. The machine created by Turing was just the beginning of adopting A.I. to analyse and learn from real world data to perform tasks that were previously done by people. (Dunis et al, 2016) Many definitions and interpretations of A.I. exist. (Boden, 2016) A definition of Artificial Intelligence is the ability of machines to understand, think, and learn in a similar way to how human beings use their mind (McCarthy, 2007) to function appropriately and with foresight in its environment. (Stanford, 2016) The field of A.I. is depicted in figure 2. (Duin et al, in Deloitte, 2017) In the next paragraph we discuss the different perspectives on Artificial Intelligence, before continuing our discussion on Machine Learning.
Figure 2: Field of Artificial Intelligence (adoption from Duin et al, in Deloitte, 2017 ) 1
1
https://www2.deloitte.com/nl/nl/pages/deloitte-analytics/articles/part-1-artificial-intelligence-defined.ht ml
Different perspectives on an A.I. driven world
Recent A.I. developments are pointing towards a fully automated A.I. driven world. (Pan, 2016) In literature there are four different perspectives on the development of A.I. (1) the optimists provide an utopian view from scientists such as Ray Kurzweil. They think of A.I. as a key technology enabling revolutions in the fields of Genetics, Nanotechnology, and Robotics (GNR). Technologies that will be helping humanity in eliminating diseases, old age, and human work respectively. This utopian view thinks of robots of performing all the work, and humans enjoying leisure and creative fulfillment. (Goertzel et al, 2017) On a critical note; questions could be asked about who will pay for the robots, and who will pay for housing and other basic needs. The “Singularity” book written by Kurzweil only answers this question in terms of reduced costs, which would make robot services available for the masses. (2) The pessimists, view A.I. as the important decision maker in all aspects of human life. This would make people less confident in making their own decisions. An example of this viewpoint is the introduction of autonomous cars, planes, and boats that would leave many millions of drivers, pilots, and captains unemployed. Simply because autonomous vehicles will be safer and cost efficient. People would become mere observers of the decisions made by A.I. This would lead to less jobs, but cheaper products and services. (3) The pragmatists believe that A.I. should be used for human intelligence augmentation. Allowing humans to be one step ahead. They also think there should be worldwide initiatives, such as OpenAI, to focus on regulations. As a safety measure, A.I. scientists argue that a kill switch should be built-in and available. (Makridakis, 2017) Researchers suggest that there needs to be global cooperation in creating and implementing these A.I. safeguards, which is arguably the hardest thing to do in the A.I. field. (Agrawal, 2016) (4) The doubters do not believe that A.I. is able to achieve the creative level of the human mind. Because this would require breaking the rules and creativity is regarded as an anti-algorithmic process. In essence humans are regarded to be essential for creative challenges such as innovative breakthroughs, strategic thinking, entrepreneurship, and risk taking. At least in the foreseeable future. (Makridakis, 2017) The doubters could be incorrect about A.I. not being a creative entity. Because there are different types of creativity which can result in incremental and radical innovation, simply by combining different concepts. The A.I. would be able to conceive new combinations and concepts based on available data. Similarly to how music can now be created by A.I. 2 (Boden, 2016) or how autonomous cars have learned to drive based on the driving habits of
2 http://www.flow-machines.com/
people. (Agrawal, 2016) The next paragraph will explore how A.I. works in order to understand its current possibilities and limitations.
Optimists Pessimists Pragmatists Doubters
A.I. solves human challenges
Humans as worry free beings taken care by A.I.
A.I. as the decision maker.
Humans as mere observers of A.I decisions.
A.I. under control of humans; off switch in case of danger. Humans augmented by A.I. to improve human intelligence.
A.I. not capable of human creativity. Humans have a creativity edge compared to A.I.
Table 2: The four different perspectives on A.I. (Makridakis, 2017)
A.I.
methods
A.I., particularly Machine Learning uses algorithms to analyze data. The purpose is to learn from the data and use the results to describe, diagnose, prescribe, or make predictions. (Copeland, 2016) with feedback loops. (Walker et al., 2017) There are three main machine learning methods. (1) supervised learning, where the user already has an idea about the results it expects from the model. e.g. “what percentage of our customers will suspend their subscription based on historical data?”. The method works by identifying a sample set of customers and training the model. Then applying the model to the full set and determine if the predictions are accurate. (2) unsupervised learning, a method where the user does not provide desired outcomes. The assumption is that previously correlated features will occur together again. (3) reinforcement learning, is driven by reward and punishment indicators. These indicators are providing continuous feedback whether an action turns out to be good or bad. The A.I. can take corrective actions to keep improving its rewards. (Boden, 2016) The general process for learning is explained in the next paragraph.
The Machine Learning process
A.I. can be applied in a range of analytic settings. The main categories according to Gartner are presented in figure 3 and discussed here. 1) describing what happened based on historic data, and 2) diagnosing why something occurred. 3) predicting events that will happen in the future, and 4) prescribing what should be done or taking automated actions. (Gartner, 2016) However, it must be noted that the distinction between these 4 categories is difficult. Because all four analytics types are based on historic data. To provide an example; let us assume that our analysis provides an answer to the question ‘what happened’, the same
retrospective data is used in determining ‘what will happen’. Or in human terms, making a decision based on our past experience.
Figure 3: Data Analytics maturity model according to Gartner
Our daily actions are part of a continuous feedback process. The process is based on: data, prediction, judgment, actions, and outcomes. As an example, when we drive a car our goal is to travel to a location. The process can be described as follows: (1) analyzing the roads, signs, and our fellow drivers (data), (2) diagnosing the situations we encounter such as “if there is a traffic jam take road A, then we predict outcome X, but if we drive on road B, then we predict outcome Y”. (prediction) (3) weighing options: “if I leave to work early there will be less traffic jams and I will be able to leave from work early.”, “but if I leave late I could risk missing an important meeting, which would have to be rescheduled.” (judgment) (4) making the decision to drive on road A (action), and (5) arriving on time without traffic jams (outcome) (Agrawal, 2016) Next, we look at the process of training an A.I. The training of an A.I. robots requires extensive human guidance from programmers and data scientists. (Austin, 2017) The general ML process is very similar as shown in figure 4 and is based on the following 4 steps; (1) Analyze the problem that needs to be solved in a certain domain. (2) create the machine learning algorithm to find optimal routes and train it using data about available routes.
Figure 4: The process of Machine Learning (adapted from Geron, 2017)
(3) inspect and analyze the solution to find the fastest route (4) improve the understanding of the problem, (Geron et al., 2017) which has to be performed by humans. Because A.I.’s cannot think, understand, or reason. They only appear to be able to do so, because of their programming by data scientists to keep them up-to-date with the latest body of knowledge. (Austin, 2017) Researchers argue it could be necessary to simplify the learning process of an A.I. by focussing on a particular domain field. (Kurzweil, 2005). Because a major challenge is to determine the variety of factors that have an impact on data observations. (Goodfellow et al, 2016) A.I. is becoming more efficient in analyzing data than human beings. The consequences will be discussed in the next paragraph, which will discuss the efficiency of human expertise versus the performance of algorithms.
Human
expertise versus Algorithms
Humans with domain-specific expertise are average in predicting future events compared to algorithms (Ayoub et al., 2016) A.I.’s can outperforming humans on narrow and typically repetitive tasks. The reasons could be that A.I.’s are not biased by 1) self esteem pressure 2) risk attitude 3) isolated opinions 4) incomplete analogies 5) limited processing capacity and speed. If a person has to choose between an algorithm and an expert person then it is more likely that the person will be chosen. This phenomenon is referred to as algorithm aversion. Because humans have a higher risk tolerance for other people than for algorithms. But people are willing to use algorithms if they do not see them, or when its predictions are nearly perfect. (Dietvoorst et al, 2015) However, there are no guarantees that A.I. algorithms and their decisions are error free, simply because of their dependence on the breadth, depth
and quality of the data and machine learning code. (Ayoub et al., 2016) Regardless, in the future the choices people make will be largely based on A.I. recommendations. The results from a recent Baker and McKinsey survey indicates that people expect limited change in their jobs due to A.I. in the coming three years. But they do expect the impact to accelerate at an incredible pace and run up to a 20 fold impact on their jobs in just 10 years time. (Upchurch, 2017) Another research survey, which was held among published A.I. researchers, indicates when they expect a particular type of work to be replaced by A.I. technology. See figure 5 for an overview. (Grace et al. 2017) The challenge in reaching the milestones could be (1) the availability of sufficient funding, (2) finding sufficient trained A.I. engineers, (3) market adoption because businesses could be hesitant to implement these relatively new technologies. (4) employees could try to resist the adoption of A.I.’s to protect their jobs (5) customers might need time to get accustomed to being assisted by A.I. robots instead of human beings, and thus delay the speed of technology adoption.
Figure 5: timeline indicating when A.I. can perform the indicated task (Grace et al, 2017)
China is expected to reach these A.I. milestones faster than the U.S. They have already published more A.I. research papers than the U.S in 2013. The main reason for the 3 research intensity is that A.I. could add 0.8 to 1.4 percentage points to China’s GDP annually. To put this in perspective, in 2016 China’s GDP value is 18.06 percent of the world economy, representing 111119.15 billion US dollars. These growth estimates provide an4 incentive to push A.I. initiatives forward. Researchers warn that A.I. should not be rushed to the market because of risks and safety issues. (Baker et al., 2017) We explain why, in the following paragraphs.
3 https://www.wired.com/story/america-china-ai-ascension/ 4 https://tradingeconomics.com/china/gdp
A.I.
Risks and safety issues
This paragraph discusses A.I. risks and safety issues. The emerging social themes in A.I. risk and safety are (1) privacy, (2) fair treatment, (3) security and integrity, (4) preventing harm to people, (5) understanding what an A.I. is doing, (6) proper social and economic policies. (Amodei et al, 2016), and (7) responsibility and liability. (Upchurch, 2017) These topics are discussed in the following paragraphs.
Privacy
The sharing of private information is easier than ever. Technology companies such as Amazon, Facebook, Apple, and Google are famous for storing, tracking and analyzing personal information. (Esposito, 2017) The general rule of business analytics states that the more personal data is collected, the better business decisions can be made. (Provost et al., 2013) These companies argue that customers benefit from a personalized experience. However, most people do not read the privacy notification. Thus providing approval for the collection, sharing and possible trading of personal information. (Marreirros et al., 2017) The consent of an internet user to allow a monitoring cookie, provides instant benefits such as free access to online newspapers and social media. The risk of sharing their personal information which could lead to possible identity theft, exposure of personal information, or being spammed by personalized advertisements (Marreirros et al., 2017) The spamming could continue, even after buying the product or service. Indicating an insufficient understanding of the customer. There have also been signals around the globe that A.I. already understands our behaviour better than we understand ourselves. A famous example is a teenager that got profiled as being pregnant based on her buying habits by the shop ‘Target’. Whilst her parents did not yet know about her pregnancy. The teen was informed by mail from Target with a message stating “congratulations with your baby.”. Her parents accidently read the e-mail, causing much controversy over the need for further privacy regulations. Interestingly, the company Target changed their personal advertising and5 randomized other products into their person focussed advertising. The result is, according to Target, that people do not feel as if they are being spied on, anymore. (Sterne, 2017)
5
https://www.forbes.com/sites/kashmirhill/2012/02/16/how-target-figured-out-a-teen-girl-was-pregnant-before-her-father-did/
Any savvy internet users can use open source ad- and privacy blocker. However, there are6 websites that block access to their services, if a user has a privacy blocker installed. Similarly, Google and Apple block apps such as Been.Choice which blocks in-app advertisements. This example clearly indicates that who has control of an ecosystem, also7 has control over its participants. These events should be concerning because it could cause isolation or unwanted visibility for privacy concerned people. An example is a Spanish citizen that sued Google stating that links to his private information should be removed from the search engine. The data by itself might not seem meaningful, but when combined with other personal information it could become harmful. Particularly when analyzed by A.I. robots to find patterns and correlations. However, Google responded that their search engine algorithms process website data, but that it has no knowledge or control over the website data. This is exactly the reason why the EU had to intervene and created regulations referred to as the right to be forgotten. (Esposito, 2017)
Security and Integrity
The implications of being hacked and security breaches are detrimental for corporations. The financial implications are substantial, business operations could be crippled for days, and could cause bankruptcies. The trust of current and future customers could be8 profoundly damaged by such events.9 10 An A.I. could be trained to scan and recognize security hacks; in essence adapt to future cyber attacks on a 24/7 basis. It would grow and understand how to counter security breaches, alert security people, and instantly limit access to systems for particular users. An A.I. could theoretically be hacked, because it is a software program. This could lead to the leaking of private company and customer information. Hackers could also take control of A.I. robots and harm the company by using it to their own advantage. (Wakeham, in Christhi, 2017)
Preventing harm to people
Regulators have the important role of (1) providing users protection against unsafe financial products and services, (2) assessing the likelihood of failure of the financial product provider, and (3) protecting users from being exploited. In order to fulfill this task the regulators perform solvency assessments, monitor the organization of the entity, analyse market 6 https://www.ublock.org/ 7 http://been.mobi/ 8 http://www.reuters.com/article/us-cyber-attack-ukraine-idUSKBN19K1WI 9 https://www.wsj.com/articles/cfpb-fines-Fintech-firm-dwolla-over-data-security-practices-1456956326 10 http://www.bbc.com/news/business-39544762 14 of 31
conduct, and investigate the products and services. This investigation involves assessing the algorithms based on information about the underlying models, data selection, expected outcomes, and evidence that they can be achieved by using test cases. Furthermore, the company has to show how success is measured, and what the response is to success or failure, the considered alternatives, and evidence from specific domain knowledge experts (Baker et al, 2017) All these measures are intended to prevent harm to people. Thus, it becomes necessary to know what an A.I. is thinking.
Understanding A.I. logic and Fair treatment
There is a black box issue with A.I. solutions. An example of this issue is that U.S. Regulators have legislation determining that loan decisions can’t be made blindly; A.I. require supportive reasons for any loan application to be declined. This is why A.I. had a challenge in the lending business, said ZestFinance CEO Douglas Merrill. Similarly, the EU is expected to implement regulations in 2018 that requires companies to provide their users explanations on how their A.I. solutions reached certain decisions. (Knight, 2017) But there is no obvious way to design A.I., so that it could always explain why. Note that this example 11 only requires explanation when the loan is declined. But what if a consumer receives a loan offer at a higher interest rate than its peers? Should the A.I. explain why the person was assigned a higher risk and thus higher interest rate? Could this lead to unintended discrimination? We believe it could, because A.I. bases its decisions on data. If the data contains biased information about factors that increased the interest rate on a loan to an ethnic minority when loans were still provisioned by human bankers, then the A.I. could conclude it must do the same again. (Barton, 2017) To understand what an A.I. is doing there must be 1) transparency in an A.I.’s decision making algorithms and processing steps e.g. a white box 2) the data used in the decision making has to be available to reproduce the results 3) people that are capable of refining the A.I and its data. In other words; an A.I. should be understandable to its creators and accountable to its users. In addition it should adhere to social and economic policies.
Proper social and economic policies.
The social and economic policies will need to be adjusted as A.I. adoption is increasing. Elon Musk also thinks most jobs will be robotized and that it will be necessary to provide everyone with a universal basic income. 12 Bill Gates argues that robotic driven solutions should be
11 https://www.technologyreview.com/s/604087/the-dark-secret-at-the-heart-of-ai/ 12 https://www.businessinsider.nl/elon-musk-universal-basic-income-2017-2
taxed similarly as human work is taxed. The tax can then be used to provide a basic income to people.13 Governments have yet to take a stance on the issue. There are also A.I. governance movements in the A.I. market from non-governmental organisations. These are companies which are taking the responsibility to create a form of public policy stance on A.I. (Grace, 2017) e.g. the Partnership on A.I. is focussing on sharing best practices on A.I. to14 accelerate the development of A.I. technology through an open discussion platform of A.I. influences on people and society. The company OpenA.I. ; focuses on discovering and15 enacting the path to safe Artificial Intelligence as per their mission statement. The reason behind this initiative is that the rise of A.I. is seen as inevitable and A.I. could at some point become intelligent enough to reinvent itself continuously. Thus, accelerating the speed of A.I. evolution, which will be many times faster than the evolution of humans. (Kurzweil, 2005) This could pose existential questions for humankind. If there is A.I. in every corner of our lives, then what do we still have to do that matters, and cannot already be performed better by an A.I.? Finally, there are also movements to create brain to computer interfaces, by directly connecting cpu’s into our brains. An example is the Neuralink company, founded by Elon Musk. 16
Responsibility and liability
The legal risks associated with new technologies is barely understood by organisations. A question that should be answered by corporations is: “Who should be held responsible in what situation?”. (Upchurch, 2017) Because an A.I. could make incorrect decisions due to bad data, incorrect analysis, or malfunctioning algorithms. (Amodei, 2016) The responsible entity could be the financial institution, the programmers of machine learning, the data providers or others. (Upchurch, 2017) A practical example is the use of machine learning in antivirus software. It analyzes a computer by scanning for virus patterns, and provides software heuristics/behavior analysis to determine if there are threats. The possible results can be allocated into the four categories of a confusion matrix (Provost et al., 2013).
13 http://www.businessinsider.com/bill-gates-robot-tax-brighter-future-2017-3 14 https://www.partnershiponai.org/
15 https://openai.com/
16 https://www.neuralink.com/
Confusion Matrix Predicted Condition
Antivirus detected threat Antivirus didn’t detect threat True condition A virus True positive
(virus is caught)
False Negative
(virus should have been caught) No virus False positive
(quarantined a normal file)
True negative (no action required)
Table 3: The Confusion matrix for an antivirus solution As can be seen in the antivirus example there could be unintentional mistakes. Should the antivirus company be blamed, or is it reasonable to assume that mistakes will be made. This is probably the reason why firms have to be (1) careful in their statements, and (2) should not rush their solutions to the market. (Upchurch, 2017). The Tesla corporation has experienced at least two types of liability claims regarding their A.I. based self driving cars: 1) a class action lawsuit claiming the autopilot functionality is not providing the marketed autonomous driving functionality. (e.g. the A.I. is not good enough) 2) drivers claiming their17 autonomic car caused an accident or should have prevented one. (e.g. is the person or the A.I. responsible) 18 Both lawsuits were unsuccessful and dismissed. But it remains easy to see that mistakes could be made, be it by humans or A.I.’s, and that the question remains who is liable. A final point regarding the responsibility and liability, it is interesting to see that large established banks are buying smaller innovative Fintech’s. This allows the Fintech to scale up under the umbrella of an established and trusted name. 19
Now that we have explored the field of Fintech and A.I. separately, it is time to discuss the how A.I. robots are currently adding value to the Fintech industry, in the next chapter.
17 https://electrek.co/2017/03/22/tesla-autopilot-class-action-lawsuit-self-driving/ 18 https://www.forbes.com/sites/patricklin/2017/04/05/heres-how-tesla-solves-a-self-driving-crash-dilem ma/ 19 https://globenewswire.com/news-release/2017/03/17/940463/0/en/BinckBank-acquires-FinTech-firm-Pritle.html 17 of 31
Chapter
3: Fintech and A.I.
The financial industry and its customers can benefit from A.I. technologies. (Dunis et al., 2016) But according to Gartner, in order to create an impact the main focus for businesses should be to offer solutions for problems, instead of just the latest A.I. technologies (Patrizio, 2017)20 Indicating that many businesses poach their A.I. solutions instead of solving problems. Research indicates that the expected impact of A.I. will be the largest on the provisioning of loans, asset management, but also stock and trading exchanges. The corresponding financial areas that are expected to be most automated by A.I. are risk assessment (49%), financial analysis and research (45%), investment and portfolio management (37%).
Financial companies that are looking to automate their business processes are expected to benefit the most from technologies in the field of big data and advanced analytics (52%), artificial intelligence (18%), trading platforms (14%). (Upchurch et al., 2017)
Overall, the composition of the Fintech industry is expected to be more diverse due to the entry of more small entrepreneurial companies. At the same time Fintech is expected to be dominated by large players (24%). The same concentration and domination, we discussed in terms of A.I., by the largest technology players is also expected in the financial industry. The financial provider that is able to eliminate human cost and leverage the potential of A.I., is expected to profit the most and achieve the largest market share, if not market dominance. Regarding the workforce, Gartner expects that by 2019 already 10% of IT hires will be writing scripts for customer service bots. This could be a conservative outlook when looking at A.I. solutions like Flow.ai, which enables anyone to create an A.I. chatbot in minutes. 21 Surprisingly, Gartner also expects that startups will overtake Amazon, Google, IBM, and Microsoft in driving the A.I. economy. These startups are expected to cannibalize 30% of revenues from market leading companies. This is an interesting estimate when analysing the numerous A.I. takeovers performed by these technological behemoths. The organisations22 that are able to adopt A.I. technology are expected to be up to four times more successful than others by 2020. A year in which 20% of businesses are expected to dedicate workers to monitor and guide neural networks. (Andrews et al, 2016)
Regulations are probably one of the biggest challenges to Fintechs. As an example the U.S. 20 http://www.networkworld.com/article/3209713/software/beware-of-companies-claiming-products-have-ai-capabilities.html 21 https://flow.ai/ 22 http://www.simplybusiness.co.uk/microsites/hungry-tech/ 18 of 31
market is a single market, yet with complex multi level regulations. On the other side we have the European market, a complex market with some harmonized regulations. And finally we have the Asian market, which is regarded to be a very fragmented market. (Christi et al., 2016) Robo advisors would have to comply with the local regulations, or fines could be invoked by local authorities which could run up into the millions of dollars.
In the next paragraph we discuss how fintech business models are currently changing thanks to A.I. technologies.
Fintech
Business Models
In this paragraph we provide an overview of companies that focus on Fintech and A.I. market. The list is based on a selection from the CB Fintech market research list. We discuss the Fintechs to understand the added value of A.I. technology in each segment. The following is presented 1) description of the segment, 2) successful Fintech examples, 3) the role of A.I. in the segment.
Payments
and savings
The payments sector is the least regulated. Consumers are most interested in payment speed, convenience, and accessibility. Typical payment solutions include mobile wallets, peer-to-peer mobile payments, foreign exchange and remittances, realtime payments, and digital currency solutions. (Lee, 2016) The business models in this segment are typically be based on a subscription fee, payment fees, and bid / ask spreads when paying in other currencies. The payments segment focuses primarily on A.I. chatbots and mobile assistants to monitor personal finances. A business example is the company Digit. The Fintech has built an savings chatbot assistant that is A.I. driven and accumulates small amounts to your savings account at a fee of $2.99 dollar per month. At a savings of 100 dollar per month this would be 3% extra fees. Its doubtful that the Digit A.I. would indicate to the user that it is not a useful savings app in this particular situation. An example of a Fintech startup in this segment is Kasisto. The company offers a natural language chatbot to be able to perform normal banking services; get balance information, send money, ask service and analytic related questions. Remittance services is another category of payment banking. The purpose of remittance fintechs is to make it easier, faster and cheaper to send money to relatives in other countries. The Fintech RemitRadar provides this service by aggregating remittance rates from different banks to provide consumers with the best available rates. The service is accessible through an A.I. chatbot on Facebook, but can also be integrated into
other websites. Another example is Transferwise. Their remittance solution assesses the funds that need to be transferred from country A to B and from B to A, but does not actually transfer the funds through international banking intermediaries. Which would incur all types of transaction costs. Instead, the funds are redistributed locally. The use of A.I. technology is said to be focussed on their chatbot.
Market
research and sentiment analysis
This segment focuses on market research and measures social media sentiment. The Fintech Dataminr uses their A.I. technology to retrieve the latest news from social media in real time, and plots the news on geographical maps. This could help businesses understand customer feedback, events and trends. The Fintech Alphasense offers an A.I. based solution to easily search and browse corporate data and news using natural language questions to perform A.I. driven in-depth research. The main focus of these A.I solutions collect and analyze big data in realtime to present relevant information to their users.
Credit
scoring and direct lending
Fintech in the peer to peer lending area allow individuals and businesses to lend and borrow from each other directly. The business model is typically based on commission fees. These platforms use of A.I. to analyse online and offline indicators to assess the creditworthiness of customers. Depending on the country there are business model preferences, mostly influenced by the local legislations. (Dudley et al, 2017) We expect that in the future the Fintech’s will be providing loans more independently. Thanks to realtime and accurate credit risk and return information. Some examples of active Fintech’s in this industry segment are now discussed. The Fintech Affirm is an A.I. based credit scoring service to provide direct loans in the 10% to 30% APR range through an app to buy products in affiliated stores. This is, unfortunately, similar to the high fees of credit cards. The Fintech Zestfinance focusses more on the credit assessment backend by using A.I. to assess credit scores based on 3000 variables and determine underwriting fees for insurances. The Fintech TrueAccord has created an A.I. based debt collector. The main advantage of this A.I. driven service is automatically initiating notifications to the debtor using mail and smartphone, making it easy to repay the debt with easy payment options, and see the results of the collection process. The Fintech CollectAI is a direct competitor by focussing on collecting the debt and
incorporating the case handling into its A.I. learning cycle. This enables the A.I. to continuously learn and improve the processing and handling of similar cases in the future.
Insurance
Insurance companies connect the insurer and the insuree to provide insurance coverage for car, life, health care, or casualty insurance. The focus on improving risk analysis using traditional and non-traditional sources of information.If we would ask the question how can autonomous cars have an impact on the insurance industry? The answer is simple, because A.I. is expected to be a better driver than humans can be. Because these autonomous cars can see 360 degrees and learn from the collective driving data created by human and autonomous drivers, and have a faster response time. Resulting in safer driving which will 23 have severe consequence for insurance companies. Because there will be less accidents, and this will in turn reduce insurance fees. Interestingly, even when people are insured 24 they are not always aware about the claims that they are eligible for. Furthermore, people could be over-insured by having different insurances that cover the same e.g. car incident. Example is a car insurance and a travel insurance. Last but not least, more insurances are luring customers with low insurance fees. However, this typically implies high “own risk” fees when a claim is made, which have to be paid by the insured person. Fintech’s could provide assistance in selecting the right insurance, and provide personal advice when claiming compensation. We discuss several Fintech’s in this segment. The Fintech Lemonade has created an A.I. chatbot which is able to insure people by interacting in natural language to determine their insurance requirements. The company earns 20% fees on the insurance fees as an intermediary between the originating insurer and the consumer. The Fintech Cape Analytics utilizes satellite images to detect real estate attributes, with the help of A.I., such as the size of the land, the home, solar panels, parking space and others. The information can be used to determine the homeowners insurance fee.
Business
finance and expense reporting
Fintech’s in this segment improve the process of business accounting. This could save time and money for companies, because 70% of all businesses are still managing their invoices via paper. (Christi et al, 2015) The company AppZen provides an A.I. robot for Time & 23 http://www.businessinsider.com/tesla-avoids-accident-before-happens-2016-12?IR=T
24
http://www.marketwatch.com/story/why-robots-may-pay-less-for-car-insurance-than-you-will-2016-06-01
Expense reporting. The A.I. robot reduces the time spent on needless auditing of expenses. The Firm Zeitgold aims to artificially handle bookkeeping, payrolling, tax filings, payment notifications, and debt collecting. The customer segment for their product seems to be small and medium enterprises who cannot afford to have separate people for these functions. This is a general A.I. adoption strategy, to start with SME’s as early adopters, for new Fintech entrants because the technology has to be trained and proven first, before it will be adopted by larger enterprises.
General
purpose and predictive analytics
The organisations in this Fintech segment focus on general purpose A.I. technology. Their expertise is rooted in semantic and natural language applications as well as broadly applied predictive analytics. The company Opera Solutions offers analysis software which can be customized for different purposes. The concept is based on signals, these are on KPI’s derived from big data. By combining the KPI’s, the A.I. robot is able to predict and prescribe actions. The Fintech Kensho Technologies aims to use their A.I. based analytics in large enterprises. Examples are banks and the healthcare sector, but unfortunately Kensho does not disclose sufficient information to understand what it is that makes their A.I. different from competitors.
Regulatory,
compliance, and fraud detection
The purpose of companies in this Fintech segment is to detect fraud and abnormal financial behavior. They use A.I. to improve general regulatory compliance matters and workflows. The Trifacta organisation offers an A.I. driven data wrangler to easily clean and manage dataflows. In addition, it enables users to assess large amounts of data by graphically representing data patterns such as normal distributions using smart detection algorithms. The Fintech Digital Reasoning Systems provides an A.I. that is able to understand the context of natural language data in a regulatory and legislative context.
Asset
management
Asset management is a form of portfolio management. The funds can be stocks, bonds, or cash. The purpose of the Asset Management industry segment is to maximize risk adjusted returns from investments. The investment algorithms suggest an investment mix based on personal preferences, risk attitude, and matching investment opportunities. (Lee, 2016) Asset management companies are downsizing on human asset managers and moving towards robo advisors. A robo advisor automatically assesses investment goals and preferences to rebalance portfolios using A.I. algorithms (Sironi, 2016) A critical question in the asset management industry could be whether the customer actually understands the type of investment proposition. (Fein, 2015) The asset management segment benefits from automated and passive investment strategies, a transparent fee structure, and low or no investment minimums. There are good reasons for companies and individuals to be interested in asset management services. Because as research shows the more funds you have, the easier it is to make even more money. This is referred to as the “increasing returns” phenomenon. (Piketty, 2014) Indicating that funds achieved via financial investment, will over longer periods outsize fund returns achieved by directly creating products or executing services. (Goertzel et al, 2017) An example is the Numerai company, which is a hedge fund with a radical different approach to traditional hedge funds. The purpose of the Numarai platform is for data scientists across the world to compete and together create the best hedge fund A.I. algorithms. The best contributors are awarded the crypto currency Numeraie. The crypto currency cannot be mined like Bitcoins, but can be traded on cryptocurrency exchanges. The general crypto currency rage has made many people very rich, but its volatility and intrinsic value make it an unstable, not transparent, and unpredictable investment.25 Competition is fierce in the asset management field with companies like Blackrock spending millions on automating their asset management processes.26 It is likely that the collective intelligence of the crowd, such as the Numarai initiative, will create better asset management algorithms than individual companies. However, commercial hedge fund employees could easily participate and incorporate knowledge from the Numarai platform into their employer's proprietary product. Because there are no measures to protect the Intellectual Property from commercial application.
25 http://www.investopedia.com/articles/investing/052014/why-bitcoins-value-so-volatile.asp 26 http://fortune.com/2017/03/30/blackrock-robots-layoffs-artificial-intelligence-ai-hedge-fund/
A summary of these Fintech initiatives and the role of A.I. is presented in table 4. Interestingly, the A.I. assistants are focussed on understanding natural language to enable interaction with people. It will just be a matter of time before A.I. learns to answer questions in a particular finance domain better than any human could. These A.I.’s will be able to autonomously provide financial products and services.
Fintech segment Role of A.I.
Payments and savings financial advisor assistant
Market research and sentiment analysis big data processing assistant
realtime corporate data analysis assistant market analysis assistant
Credit scoring and lending assess customer credit ratings recover customer loans
Insurance insurance advisor assistant
visual real estate analytics Business finance and expense reporting administrative assistants General purpose and predictive analysis data wrangler
business analytics Regulatory, compliance, and fraud detection audit data tooling
legal advisor
Asset management hedge fund analysis
investment assistant
Table 4: Summary of Fintech and A.I. solutions
Chapter
4: Discussion
The very nature and role of finance in our lives makes it an important and intimate subject. People require financial services and products throughout their lives. In the past, it might have felt personal and intimate to go to the bank and have a bank clerk attend to your banking needs. But truth is that the bank understood less about you and your finances than it does now. (Waupsch, 2017) Businesses are collecting information about us in the real world and online. A.I. robots are the next step in better profiling and understanding our needs and behavior throughout life. To hopefully provide better, faster, and cost efficient financial products.
The Fintech landscape is adopting A.I. technology, and finding new ways to utilize its potential. There are many Fintech companies that are rushing and boasting their A.I. driven solutions. But any algorithm that has descriptive capability could be referred to as A.I. technology. This could raise questions about whether or not the concept of A.I. is overhyped. At the same time, it is clear that A.I. technology has tremendous potential.
The Fintech companies show that A.I. is able to analyze big data, find patterns we would not have found, and interact with customers. A.I. enables Fintech companies to understand customers better than ever before. The customer is likely to be unaware what amounts and types of information financial service providers are collecting about them. Or how personal information is being shared and used. Fintech’s must be vigilant that their A.I. based solutions do not become black boxes. The actions from an A.I. should be auditable, traceable, and explainable to regulators. This should also make the case for Fintech’s to be able to explain to their customers the type of personal information is being accumulated and why he or she is eligible for a certain debit or credit rate. The benefit of finance is that it can be objectively assessed. A.I. can help in this objectification. By providing pure facts and not being influenced by emotions. The danger could be that the A.I. simply analyzes the data by using the algorithms it is given, and could make the same biased decisions as his human predecessor. (Barton et al., 2017) An A.I. does not understand concepts such as privacy, ethics, or secrets as discussed in the ‘Target’ company case. Because the A.I. robot indicated that there was a baby coming in a personalized ad. The A.I. was correct, but could have been more subtle in this case. This confirms that predictions from A.I. will be readily available, but humans will have to provide judgment to determine the follow-up actions. (Agrawal, 2016)
The financial services industry is expanding horizontally (e.g. more competitors in the same niche) and vertically (e.g. more niches in the value chain). The challenge of any expert domain is to assess the different types of data that have an effect on the analysis and results. The development of Fintech and A.I. is further stimulated by the availability of open source A.I. toolboxes, big data, real time information, and software interfaces. The arrival of A.I. provides an opportunity to reinvent business models. As an example retailers could become banks, but bank are unlikely to become retailers. The monitoring of user behavior e.g. buying behavior helps in predicting future product and services needs, and provide matching offers, loyalty points, rebates, bulk discounts to keep the user incentivised. An A.I. could continuously keep track of the user behavior, and use reinforcement learning to keep the user interested. The retailer could then provide cash, or credit options to pay for these products and services for a seamless customer experience. A bank does not have sufficient detail information about the products or services bought by a consumer to provide the same tailored services. Of Course, both parties could work together to combine the behavioral information. The European PSD2 banking initiative to share banking information with third parties could be a first step into this new generation of products and services. Another practical application of A.I. is the availability of chat bots on social media that enables an user to send money to friends or family directly. The financial service is available, exactly where it is needed at that moment. Other examples can be found in the many A.I. bots that provide financial advice and coaching based on the analysis of personal financial data, such as mortgage and savings advisors. The A.I. will present the advantages and disadvantages for the proposed strategy. The A.I. will ask the user for permission to implement proposed actions. The user will have to choose between accepting and declining the advice. The A.I. solutions provide transparency and rationale in the form of decision support. The question will remain whether the user receives objective financial advice that could provide added value, or that the A.I. is outsmarting the user.
The added value from A.I. is that it can assist in analyzing data to 1) describe, 2) diagnose, 3) prescribe, and 4) predict. An A.I. is accurate in cases 1 and 2, because they are based on historical information. However, it becomes unclear if the same is true for the third and fourth analytics categories because the future is still to come. The questions that could be asked are whether our world is really predictable? Can people be clustered in behavioral groups? Would an A.I. be able to advise and coach people in such a way that it leads to financial wealth? We like to believe it to be so, because most people are not specialized into finance and wealth management. This makes it a perfect candidate for A.I. monitoring and advising. But, at the same time people should be aware of the commercial interests. This is where
governments have an important role. Governments will have continue to develop legislations and could base their regulations on rules and guidelines provided by initiatives like OpenA.I. and Partnership on A.I. Fintech’s currently have no reason to comply with these guidelines because they are not mandated by governments. But Fintech’s would preferably have to comply with these regulations, be transparent, to become a trustworthy advisor.
Chapter
5: Conclusion
The financial industry is being disrupted by A.I. initiatives. The entire financial products and services value chain is being tested for added value that can be generated by adopting A.I. technologies. The most forthcoming role of A.I. is disintermediation of the human workforce, by replacing the human factor with A.I. technologies. A.I. technologies assist in the analysis of big data by providing opportunities for real time information, new insights, and decision support. The most observed example from analyzing the Fintech companies that have adopted A.I. technology is the usage of A.I. chatbots. This enables the use of financial products and services without requiring the assistance of a workforce involved in customer servicing. The advantage for A.I. solutions is that it can handle thousands if not millions of service requests from its users. The interactions allow the A.I. to collect data and learn best practices through a continuous feedback cycle. This allows the A.I. to profile users and make better decisions in providing customized services. Arguably, the customer will benefit the most from transparent, easy to use and A.I. technology driven solutions, at a lower cost.
Future
Directions
The thesis aims to provide a general understanding of the added value that A.I. can bring to the financial industry. There are many interesting topics that could be researched in the field of A.I. We mention several based on the research performed in this thesis. There are many companies boasting their A.I. solutions, but how can anyone determine if an A.I. solution has reached a certain maturity level? e.g. is the A.I. predicting instead of describing? or is the A.I. simply describing correlated patterns that have occurred in the past and indicating that the same will happen again. How can this be measured objectively?
Regarding the role of governments; should there be regulations that A.I. should present their findings to the user and be transparent about the costs that will be occured to make commercial interests also visible. In comparison, the EU has set restriction on credit card fees and transparency in fee structures. This is not the case in the U.S with its credit rates of up to 30% APR.
Last but not least, can customers be protected from A.I. misusage? in particular if the A.I. has commercial interests and focusses on the less experienced financial consumers (Funfgeld et al, 2009)
References
Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J., & Mané, D. (2016). Concrete Problems in A.I. Safety, 1–29.
Andrews, W., Brant, K., Revang, M., Reynolds, M., Karamouzis, F., & Hare, J. (2016). Predicts 2017: Artificial Intelligence.
Ayoub, K., & Payne, K. (2016). Strategy in the Age of Artificial Intelligence. Journal of Strategic
Studies, 39(5–6), 793–819.
Austin, T., Linden, A., & Rollings, M. (2017). Hype Hurts : Steering Clear of Dangerous AI Myths. Baker, T., & Dellaert, B. (2017). Regulating Robo advice across the financial services industry,
1–39.
Barton, D., Woetzel, J., Seong, J., & Tian, Q. (2017). Artificial Intelligence: Implications for China. McKinsey Global Institute, Discussion(April).
Bernoff, C. L. and J. (2011). Groundswell: Winning in a World Transformed by Social Technologies. Harvard Business Review Press, 286.
Boden, M. A. (2016). A.I.: Its nature and future (1st editio). Oxford: Oxford University Press. Brynjolfsson, E., & Mcafee, A. (2017). The Business of Artificial Intelligence.
Bundy, A. (2017). Smart machines are not a threat to humanity. Communications of the ACM Centineo, S., & Centineo, S. (2017). Investment innovation trends: Factor-based investing.
Serbian Journal of Management, 12(1), 65–75.
Chishti, S., & Barberis, J. (2015). The FINTECHBook: The Financial Technology Handbook for
Investors, Entrepeneurs and Visionaries (Vol. 1).
Crosman, P. (2017). How Fintechs are using A.I. to transform payday lending
Dietvorst, B. J., Simmons, J., & Massey, C. (2014). Understanding Algorithm Aversion: Forecasters Erroneously Avoid Algorithms After Seeing them Err. Academy of
Management Proceedings, 2014(1), 12227–12227.
Dudley, W., & Knot, K. (2017). Fintech Credit: Market structure, business models and financial stability implications. The Financial Stability Board.
Dunis, C., Middleton, P., Karathanasopoulos, A., & Theofilatos, K. (2016). Artificial Intelligence in
Financial Markets: Cutting-Edge Applications for Risk Management, Portfolio Optimization
and Economics. Palgrave macmillan.
Esposito, E. (2017). Algorithmic memory and the right to be forgotten on the web. Big Data & Society, 4(1)
Fein, M. L. (2015). Robo-Advisors : a Closer Look By Melanie L. Fein, 1–33. Flasiński, M. (2016). Introduction to Artificial Intelligence. Springer.
