The AI Book: The Artificial Intelligence Handbook for Investors, Entrepreneurs and FinTech Visionaries

Part 1

AI: Need to Know

Artificial intelligence (AI) is poised to disrupt lives, businesses, whole economies and even the international geopolitical order. As such, it has never been more important to have a clear understanding of what AI is and the ramifications of its mass adoption, particularly in the financial services sector. However, the inherent complexity of the topic is often intimidating to non-specialists, and the absence of broad-based dialogue on the topic of AI is hindering business decision-making related to its application.

What exactly is AI; how is it being used in financial services; what is at stake; who are the major players; and what lies over the horizon?

In Part 1, we will explore all these questions and more. By delving into the detail behind the hype, readers will gain a firm understanding of the different type of technologies that fall under the more general, and somewhat opaque, AI heading. We will have the opportunity to look at how nation states are jostling for position and international competitive advantage relative to their peers through their national AI strategies and action plans. We will also have a chance to learn about tried-and-tested recommendations for successfully embedding AI into the daily operations of financial services firms, while avoiding the myriad pitfalls that still unfortunately get in the way of firms reaping the full advantage of their AI investments.

Finally, we will take a close look at the human aspects of AI and examine the reasons why, in the face of the growing sophistication of algorithmic systems, the exercise of sound human judgement, governance and control has never been more important. We will look at the role of boards and directors in the formulation and execution of AI strategy within firms, and we will see how artificial intelligence systems that complement human cognition have the potential to deliver maximized value.

CHAPTER 1

The Future of AI in Finance

By Chee-We Ng¹

¹Venture Capitalist, Oak Seed Ventures

How will artificial intelligence (AI) transform finance? What can AI do and how can we get it to work? What do we need to do to regulate AI in finance? These are questions at the forefront of many minds as we try to investigate the future of finance.

AI, a loosely defined set of technologies that try to mimic human judgement and interaction, has been in use in banking and finance since its inception in the 1950s. AI encompasses everything from rule-based technologies and probability-based methods that detect fraud, through to primitive neural networks for optical recognition and automatic stock and option trading. Collectively, these technologies automate processes that were previously undertaken by human beings, often improving accuracy and efficiency. One might argue that none of these traditional AI technologies is truly intelligent; AI merely automates what was previously performed manually.

The Promise of Deep Learning

The recent excitement around AI has tended to be linked to deep learning in its various forms. To understand why deep learning technologies simultaneously inspire excitement among researchers (who believe that deep learning is the breakthrough in AI everyone has been waiting for), and fear among tech leaders and politicians, it is important to place deep learning in the context of what its component technologies have achieved in the past 6 years.

The most recent wave of deep learning began in 2012 when Geoffrey Hinton and his students used deep convolutional neural networks (CNN) to tackle image recognition, a problem that has baffled scientists and engineers for many years. By achieving significantly higher detection rates and smaller false positives without having to write complicated code, Geoffrey Hinton was able to teach computers how to classify images just by showing many labelled samples, hence the term machine learning. AI was taken to new heights in 2017, when Google’s AlphaGo, and subsequently AlphaGo Zero, beat the world’s best Go player, Hanjin Lee. Using reinforcement learning, AlphaGo Zero learnt how to play by playing against itself without having been provided any instruction on how to play. Not only did it teach itself Go strategies humans had developed over hundreds, and possibly thousands of years, it developed strategies that no human had ever conceived of previously.

Meanwhile, recurrent neural networks (RNN), and variations like long short-term memory (LSTM), improved machine translation significantly, while generative adversarial networks (GANs) succeeded in restoring colour photographs from old black and white ones, creating cartoons and oil paintings from photographs and even making fake videos and photographs. In a matter of years, deep learning has demonstrated, at least under certain conditions, that it can learn better than humans (without being taught) and be capable of mimicking humans themselves.

Business Applications in Finance

Today, AI and deep learning have broad ranging applications in deposits and lending, insurance, payments to investment management and capital markets. Deep learning methods are now better than probability-based methods in fraud detection. Like image recognition, fraud detection is a classification problem. Instead of creating static rules which struggle with keeping up and are not sufficiently discerning at times, deep learning solves the classification problem by letting the machine learn by itself. Similar technologies are used in assessing the right premiums for insurance markets and making predictions about stock market prices based on a large number of variables, which can then be used for automated trading.

Just like how AlphaGo Zero taught itself strategies of Go that humans haven’t discovered, deep learning is now used in finance to make connections between large numbers of seemingly unconnected events and variables to make predictions for fraud detection, insurance pricing and trading stock. With strides in natural language processing (NLP) achieved by deep learning, chatbots are also used in banking and finance to do preliminary sales and improve customer service, replacing human customer service agents.

Time for a Reality Check

Despite having made significant breakthroughs, deep learning nonetheless has limitations. These limitations can present themselves in the form of implementation challenges, unintended consequences and ethical issues. In order to implement deep learning technologies well, large quantities of labelled and clean data are often required. Picking the right neural network architecture and the number of layers is largely an art today and performance and robustness varies with architecture. To obtain large volumes of clean labelled data often requires significant effort on the part of firms in consolidating, fusing and cleaning large volumes of source data.

Data needs to be unbiased, or otherwise the machine will learn the bias that is inherently embedded in the data. It is a known fact that many facial recognition algorithms work well with certain races but much less reliably in other races and gender. It is also known that language models today are sexist or discriminatory because of biases engrained in the training data. When such biases exist in finance, it means that certain races or gender may be subject to lower approval rates for loans, or higher interest for mortgages or higher premiums for insurance.

Furthermore, because deep learning is essentially still a black box, it can fail catastrophically in unexpected ways. Studies have shown how when noise imperceptible to the eye is added to images, deep learning can recognize a panda as a cat with high confidence. It has also been demonstrated that deep learning algorithms used in autonomous cars to recognize road signs can be easily tricked.

As deep learning learns patterns and correlations without understanding causality, its classification result may be based on the wrong features, or features that are only temporal, or even features that coincide but actually do not mean anything. When deep learning is applied to finance, it can mean that loans could be rejected unfairly for a reason that is hard to decipher and explain to customers. Meanwhile, it is also plausible that a smart attacker could fool a deep learning model used to detect fraudulent activity.

Safeguards and Systemic Risk

When AI is used in isolation, the impact of major failures could be large but contained. However, as AI is being used more and more in connected systems such as in the stock market for automated trading, unexpected catastrophic failures could lead to the widespread failure of entire systems. We don’t need to go very far back in history to recall how credit default swaps caused the financial crisis of 2008 and the valuation of Russia’s ruble led to the 1998 crash of Long Term Capital Management (LTCM) – a $126 billion hedge fund – that subsequently required a bailout from the US Fed. Will the use of more AI in financial markets lead to similar catastrophic failures in the future?

Finally, there are ethical issues associated with the use of AI in finance, particularly issues linked to privacy and the use of personal data. For example, do insurance companies have the right to use data related to places customers go to frequently, or their DNA profile, to optimize the pricing of insurance premiums? Other issues are linked to questions of fairness. Today, insurance premiums and mortgage rates may already be biased for people of certain ethnic origins; however, with the use of deep learning to discover connections between multiple sources of data, we may end up faced with quotes and premiums that depend on factors that we would typically consider unfair and unjust from an ethical perspective.

The question is, will AI cause our moral compass to shift course?

AI is the new electricity, and with great opportunity comes great responsibility. AI is not perfect and can be harmful if used improperly. What it certain is that AI will expose us to immensely challenging questions related to ethics and accountability, and we will need to leverage the very best of our humanity if we are to find the answers we need.

CHAPTER 2

What Is AI and How to Make It Work for You

By Terence Tse¹, Mark Esposito² and Danny Goh³

¹Co-Founder and Executive Director, Nexus FrontierTech

²Co-Founder and Chief Learning Officer, Nexus FrontierTech

³CEO, Nexus FrontierTech

Let us start with a fact: there is really no intelligence in artificial intelligence (AI). If anything, the term has been so overused recently that the hype is reminiscent of the dot-com boom in the late 1990s. The problem back then – as now – was that many companies and opportunists were making exaggerated claims about what technology can really do; so much so, that a recent study found that a staggering 45% of companies in Europe claiming to do AI actually operate businesses that have nothing to do with AI.1

Sure, machines can solve problems. Yet, while they can perform complicated mathematical calculations with a speed that no human can match, they are still unable to do something as simple as visually distinguishing between a dog and a cat, something that a 3-year-old child can do effortlessly. Viewed from this vantage point, AI can at best solve clearly defined problems and help with automating time-consuming, repetitive and labour-intensive tasks, such as reading standard documents to onboard new customers and entering customer details into IT systems. Furthermore, the term machine learning is somewhat misleading, as machines do not learn like human beings. They often learn by gradually improving their ability and accuracy so that, as more data is fed into them, they guess the right answer with increasing frequency. Through such training, they can come to recognize – but not understand – what they are looking at and are still very far away from comprehending the nuances of context. This is like when we text on our smartphones: often the right words will be presented for us to choose from. While remembering what we have typed in the past, our smartphones can guess the right words to complete a sentence to a reasonably accurate degree; this doesn’t imply that our phones actually understand the meaning of the words or sentences we type.

So, all in all, and for the moment at least, AI resembles much more a mindless robot and much less a thinking machine. This, in turn, means a bit of presence of mind is required when leveraging AI in business activities. The following five action points can help.

1. Be Narrow Minded

AI is currently most effective in dealing with very narrow tasks in well-defined circumstances. It is therefore important to narrow your scope when thinking about what you would like to use AI to achieve in your business. It is also paramount to know the exact business objective you want to achieve. Labour-intensive and time-consuming standardized tasks are particularly ripe for automation using AI.

2. Weigh the Risk

When it comes to AI, humans need to get comfortable with the idea of relinquishing some control. Once you implement AI in your business, it is important for everybody – human and machine – to stay in their respective lane: there are certain things people will be responsible for and certain things that will be best left to the machines. One of the biggest issues that people have with AI is the idea of letting machines make decisions for them. It can be a scary prospect, but it doesn’t have to be all or nothing. If the decisions are minor, and the machine can improve its accuracy over time, then the risk is minor, and the best path is in letting the machine continue autonomously. However, if the decisions have major repercussions, then it is probably advisable to have humans involved in the decision-making, with AI assisting by processing data in a way that helps inform those decisions.

3. Get the Last Mile Right

Even if 99% of a job is automated, there will always be 1% that needs to be handled by humans. There are three main reasons why it is important to think through this last mile carefully and the manner in which it should be integrated into workflows and procedures. Firstly, it remains important to have a human checking the work of machines, particularly those that carry potentially large financial risk. The second reason is that while certain tasks can be automated, there are still a lot of tasks that are best left to humans, such as customer-facing work that involves selling complicated financial products. Thirdly, there are tasks that machines are simply unable to take over from human beings, particularly the many physical activities still requiring human intervention, such as quality control.

4. Consider That Less Data May Mean More

Contrary to what many people believe, the idea of the more data the better is often a misconception. Not all goals need to be achieved with 100% accuracy. The important thing to understand is the minimum level of accuracy needed to do a job. If this baseline level is low, then less data would be needed to train the AI models. While there are times when 100% accuracy is not needed to solve every problem, at other times problems can be so complex that not even a machine can solve them with perfect accuracy. In this case, no matter how much data is available, it will not help reach the objective. By contrast, where the task is easily definable and straightforward, it is possible to achieve near 100% accuracy even with only a small training set. Furthermore, machine models decay over time because data sets evolve and become outdated. If a firm has a massive store of data on selling mortgage products, for example, the same huge set of data will not be of much help in making an AI model to improve the selling of insurance products.

5. Do Your Homework

If we take as a given that there is only so much AI can actually do, then that means that firms must take on the burden of effort involved in laying the groundwork for putting AI to work. Executives often do not know what they really want or only have an abstract idea of a goal they would like to set for their business. They may know, for example, that they want to reduce costs, but they do not know how to go about doing so. It is important to know that AI is not built to serve abstract purposes. To get results, firms need to have a very clear idea about what it is they want to achieve. Another must-do piece of homework is to map out current workflows and processes. This is because technology must be supported by the robust workflows and processes to maximize its potential. In turn, the workflows and processes should be backed by broad-based staff buy-in, both managerial and IT. While this may sound like stating the obvious, it is often surprising to see people’s enthusiasm in taking on AI wane as soon as they are asked to plot out the existing workflows and processes they seek to improve.

It is easy to be overwhelmed by the sheer possibilities offered by AI. Many companies make the mistake of thinking too big when the scope of impact of AI is, in fact, quite small. They overgeneralize and exaggerate the impact AI could have on their company. The unassailable truth is that AI is most effective in narrow, well-defined and specific circumstances. When approaching the question of implementing AI in your company, it is important to stay grounded in the overarching goals and missions of your business. AI cannot define or replace your business strategy. AI for AI’s sake is neither useful nor cost efficient; however, employing AI strategically to help advance your business can be a boon for everyone.

Note

1 Olson Parmy, Nearly Half of All ‘AI Startups’ Are Cashing In On Hype, Forbes.com, 4 May 2019: www.forbes.com/sites/parmyolson/2019/03/04/nearly-half-of-all-ai-startups-are-cashing-in-on-hype/#61d8a348d022.

CHAPTER 3

Getting to Day Zero: Let’s Get the Foundation Right

By Matt Allan¹

¹Founder, Fintech Sandpit

Artificial intelligence (AI) is new, exciting…and difficult. McKinsey & Co estimates the technology will generate $13 trillion of revenue over the next decade, affording global GDP over 1.2%.1 New use cases consistently promise not only efficiency improvements but also deeper experiences for customers. However, the road to AI adoption is windy, unmarked and filled with potholes of resistance. Herein lies the root of the problem. Banks have traditionally placed themselves at the centre of their own universe, while the customer came second. The customer only had access to the bank during opening hours, and always had to play by the bank’s rules. This is incompatible with a successful deployment of AI, which requires an obsession with customers, their data and its quality.2 Many banks are missing fundamental aspects of reliability, scalability and security within their existing architectures, and yet are committing to build AI systems that promise to be reliable for customers, adaptable for the future and secure for data. We must grow legs before we can crawl, walk or run, and banks must create a firm digital foundation before they rush to deploy AI. This chapter teases out some of the challenges that firms must overcome and the opportunities available to maximize the value of artificial intelligence.

Challenge 1: A House Built on Sand

As customers began to demand new forms of interaction via internet and anywhere/anytime mobile banking, the requirements placed on legacy core banking systems grew exponentially. Instead of renewing these systems, most banks layered additional products on top of the base system to satisfy new business demands, creating additional layers of operational complexity, cost and risks. Core technologies were expanded to do things they were never designed to do. From a cost perspective, this means most banks today spend between3 80% and 90% of their IT budget simply to keep the lights on. Some banks still use core banking software purchased 30 or more years ago. Forcing these systems to support real-time mobile banking or open banking application programming interfaces (APIs) creates a massively complex architecture that is extremely fragile. Advisory firm KPMG4 suggests that the status quo is slowly changing, and that CIOs are becoming more proactive in solving their data access and operational resilience problem. They know that in order to take advantage of technologies like AI, there is an urgent need to revitalize legacy systems, move workloads to the cloud and open their core for collaboration.5

Challenge 2: The Digital Transformation Dilemma

Banks are struggling to adopt AI because they have not traditionally placed customer data at the centre of their operations. Digital transformation is an attempt to rid themselves of the constraints of their legacy technologies and to re-establish the customer as their primary focus. However, undergoing a full technology refresh while continuing to serve millions of customers is like trying to replace the engines on an airplane while 30,000 feet in the air. Banks are stuck in a classic catch-22 dilemma: their ongoing commitment to serve their customers is preventing them from succeeding in the timely deployment of emerging technologies (such as AI), that would ultimately improve how effectively they can serve their customers. Above all else, customers value service availability and want unfettered access to their money at every second of every day. To complicate matters further, banks must continue to support the needs of an increasingly diverse customer base. Challenger banks have started their businesses with a greenfield approach and are unencumbered by the maintenance burden of old products (like cheques) or legacy account structures. Unfortunately for themselves, incumbent banks do not have that luxury.

Banks operate in a web of social friction that they must skilfully navigate during their deployment of emerging technologies like AI. The best example of this is in the debate of whether to close brick-and-mortar branches. Despite declining utilization and significant cost, closing branches in rural towns often escalates to a social or political issue.6 Despite the resistance, banks must continue to overcome the challenges that prevent their digitalization. The applications of artificial intelligence discussed in later chapters will only be realized if banks can successfully navigate this transition successfully.

It’s time to take advantage. Although banks are playing catch-up when it comes to laying a firm foundation for AI, there are several opportunities banks can take advantage of today in order to prepare for tomorrow.

Opportunity 1: Share Your Data with the World

Banks with an open attitude to data sharing will realize much more value from AI than those who seek to lock their data away. Banks who have adopted an open data or platform strategy like Starling, a UK challenger bank, are generating much more revenue from their data than those who remain closed. Starling’s developer APIs go well beyond that mandated by open banking and PSD2, which enable FinTech companies to solve problems for their users and integrate directly with their platform. Accenture7 found that 71% of banking practitioners believe that organisations that embrace open banking will reduce their time to market, streamline their operational costs, and offer better experiences for their customers. Banks who prioritize their data openness will be in a much better position to consistently leverage AI as opportunities arise.

Opportunity 2: The Alternative Data Revolution

Never before have firms had access to so much contextual and supplementary data about their customers as they have today. Unstructured data contained in news articles, broker research and/or written documents is becoming much easier to ingest in automated systems due to advances in convolutional neural networks. Firms that can synthesize and mobilize the information contained in health, geographical, transaction, credit and social media data sets will be at a significant competitive advantage. To benefit from data availability, banks must pay close attention to improving data quality. Refinitiv,8 a banking consultancy, found that 43% of banking executives see data quality as the biggest hindrance to benefiting from AI. When it comes to machine learning, the adage garbage in, garbage out has never been more pertinent. As future chapters will explore, alternative data alongside AI will be the catalyst to streamline back-office processes for legal and compliance purposes. Banks must keep asking how they can become more efficient, and then use the data available to execute.

A Bright Future

It is not the strongest of the species that survives, or the most intelligent, but the one most adaptable to change.

Charles Darwin

AI is an incredible set of technologies, yet most banks are still unable to fully realize its potential. Most are trying desperately to move faster, yet the complexity and cost of keeping the lights on continue to hold them back. In nature, as in business, those most responsive to change will survive. Financial institutions that are proactive in taking advantage of alternative data and understanding their own data will be better positioned to capitalize on the new and exciting opportunities presented to them.

At Fintech Sandpit, we are helping financial institutions securely work with fintechs to realise the potential of their data. Banks use our digital sandbox to do proof-of-concepts in order to find the best partner to work with. Get in touch to instantly test any fintech on our marketplace.

Notes

1www.mckinsey.com/featured-insights/artificial-intelligence/notes-from-the-ai-frontier-modeling-the-impact-of-ai-on-the-world-economy.

2www.forbes.com/sites/willemsundbladeurope/2018/10/18/data-is-the-foundation-for-artificial-intelligence-and-machine-learning/#13a77be751b4.

3www.bobsguide.com/guide/news/2019/Jun/5/banks-marooned-by-mainstream-technology-development/.

4https://assets.kpmg/content/dam/kpmg/nl/pdf/2018/sector/banken/banking-systems-survey-20172018.pdf.

5www.fintechfutures.com/2019/06/core-modernisation-is-essential-for-truly-digital-banking/.

6www.theguardian.com/business/2018/may/01/rbs-to-close-162-branches-with-loss-of-800-jobs.

7www.accenture.com/_acnmedia/PDF-71/Accenture-Brave-New-World-Open-Banking.pdf#zoom=50.

8www.refinitiv.com/en/resources/special-report/refinitiv-2019-artificial-intelligence-machine-learning-global-study.

CHAPTER 4

Navigating a Sea of Information, News and Opinion with Augmented Human Intelligence

By Andreas Pusch¹

¹Founder and CEO, YUKKA Lab AG

We are drowning in an endless sea of new information. Google reported that the number of web pages has grown from 1 trillion in 2008 to a whopping 130 trillion in 2016.1 Even though a majority of these might be irrelevant to your or your company’s interests and operations, it is equally true that well-informed decisions will have a major impact on the success of any project or investment. While the amount of information is growing, the human ability to read and digest information has stayed rather stagnant in absolute terms, and even decreased in relative terms. Research has shown that an average professional reads 5–15 articles per day from 1–3 sources. Not only is this a negligible amount compared to the actual volume of news published each day, but there is also an undeniable information bias driven by personal preferences and valuable time lost reading nonsense. It is thus easy to get lost in this growing sea of information, which leads to wasted time and prevents well-informed decision-making.

But what if we could be one step ahead? What if we could objectively analyse hundreds of thousands of articles from thousands of professional sources within a matter of minutes? This would not only give us the opportunity to understand an industry, a firm or another entity in greater depth while minimizing information bias. It would also enable us to save multiple hours per day that are otherwise wasted reading news or researching disclosures from a company. The solution to this exact problem can be found in augmented human intelligence and specifically in two key underlying techniques, which can be used to increase process efficiency in many information driven industries, such as financial services and management consulting.

Making Sense out of Complex Text through Natural Language Processing (NLP)

In order to extract the most value out of incoming data for it to be useful for future purposes, we need to first analyse and make sense out of it. Here, NLP comes into play and performs a linguistic analysis of the text at different levels of increasing complexity. At the lowest level, NLP performs actions to make sentences and words digestible, understandable and comparable. Initially, information is used to obtain a syntactic-semantic representation of the sentences (a representation of their meaning). The ultimate goal is for the system is to gain a deeper understanding of individual words and sentences (similar to a child learning to speak). Furthermore, NLP needs to be able to detect entities that are mentioned implicitly through pronouns or general expressions (such as the company). Building upon this, an application-dependent analysis can be performed such as through sentiment and target recognition, which allows the NLP system to detect the polarity of sentences (positive, neutral, negative) and the respective target entity. This entity recognition not only expands to company names, but also to C-level executives, subsidiaries, etc., as discussed in the next section.

In so doing, deep learning and neural networks set the baseline for this type of next-generation machine learning. Neural networks can almost limitlessly expand their learning capability without requiring significant pre-processing, since they are able to learn language structures from sentences and their context alone. The readily available and continuous data flow from numerous news organizations enables the algorithm to make use of a vast resource pool.

Ontologies Link Entities and Thus Create Valuable Connections

As previously discussed, it is not enough to merely sort and analyse data according to their explicitly named entity. In this context, ontologies are used to extract an exhaustive set of meaningful and valuable information. Ontologies represent a working model designed to provide classification of the relations between various concepts in a particular knowledge domain. Ontologies are all around us and are not only used by major firms (such as Amazon, to classify products into categories) but are also wired into our understanding of language. For example, if a person mentions Mount Everest the first thing that pops into your mind is probably mountain or high. Similarly, for iPhone this is likely to be Apple or smartphone. In essence, this reflects the simple fact that our brains have a tendency to categorize raw information, so that we can remember it and draw connections between certain subjects. Writers make use of this: a headline stating Sales forecast for Model X lowered is enough for us to assume that this is bad news for Tesla. AI ontologies replicate these connections and use them to understand relations to the same extent as we do.

How Augmented Human Intelligence Will Change the Way We Read News and Inform Ourselves

NLP and ontologies, along with several other related techniques, will have a huge impact on how we process daily information. AI allows businesses to gain both an information and time advantage, as news articles are analysed, categorized and updated in real time. Up until now, getting an overview of the current situation of an entity was a labour-intensive task. Simultaneously, feeling like you were always up to date with the latest trends was near impossible due to the rapid inflow of fresh information. Completing these tasks with the help of an augmented human intelligence offers the benefit of staying on top of the news while simultaneously freeing up time and making better informed business decisions.

While previously getting an overview meant reading the first 2–3 articles in the newspaper, in the future it means looking at insightful visual analytics such as Tag Clouds, trend signals or data networks. Important information can be spotted within seconds, while still offering the capability to delve deeper into topics of your interest. Information bias is minimized, as keywords and sentiments are curated and computed from thousands of trusted, global sources and insights can be explained and shared more easily. All of this is possible within minutes, since as humans, we tend to remember and recall information better when it is presented visually rather than through plain black-white walls of text.

While augmented human intelligence can have an impact in many verticals, the most profound impact will be in research-intensive sectors. In these verticals, much time is spent trying to assess companies’ past operations and spotting trends going forward. If this time can be saved and thus reallocated from less repetitive work to more high-value, unstructured business processes, there will be a measurable increase in productivity for the individual user and for the company.

Augmented human intelligence thus acts as a technology compass, helping firms reach their goals in the most efficient manner, free from any unnecessary information distractions along the