The Importance of AI in Banking – Hands-On Artificial Intelligence for Banking

The Importance of AI in Banking

Artificial intelligence, commonly known as AI, is a very powerful technology. A thoughtful implementation of AI can do wonders in automating business functions. AI has the power to transform a wide variety of industries through its application. As computer systems have evolved over time, they have become very powerful. Consequently, machines have also become very powerful and can perform many complicated tasks with ease. For example, Optical Character Recognition (OCR) is a task that even personal computers can perform easily with the help of software. However, OCR requires intelligence to translate dots from an image into characters. So, in an ideal case, OCR will be considered an area of AI. However, because of the power of machines, we tend to not consider it as an application of AI.

In this chapter, our focus is to understand what AI is and its application in banking. Banking is an industry or domain that is extremely diversified and complex. To simplify complex banking functions, the banking industry requires a constant supply of advanced technological solutions. As shown in a recent analysis conducted by Forbes (https://www.forbes.com/sites/forbestechcouncil/2018/12/05/how-artificial-intelligence-is-helping-financial-institutions/#2e989fae460a), the implementation of AI in various banking processes will save the industry more than $1 trillion by 2030. Consequently, the banking industry will benefit the most from AI systems in the near future.

We will begin with a brief introduction to AI and banking as an industry. Here, we will define the methods of implementing AI in software systems. We will also learn how the banking industry can benefit from the application of AI. There will be many more topics to cover before we complete this chapter. So, instead of simply discussing what you can expect from this chapter, let's jump straight into it!

In this chapter, we'll focus on the following topics:

  • What is AI?
  • Understanding the banking sector
  • Importance of accessible banking
  • Application of AI in banking

What is AI?

AI, also known as machine intelligence, is all about creating machines that demonstrate the intelligence that is usually displayed by humans in the form of natural intelligence. John McCarthy coined the termartificial intelligence in 1955.

AI has witnessed two winters so far: once in the 1970s with the reduction of funding by the Defense Advanced Research Projects Agency or DARPA (https://www.darpa.mil/), then known as ARPA, and another time with the abandonment of an expert system by major IT corporates such as Texas Instruments (http://www.ti.com/) and Xerox (https://www.xerox.com/).

In a way, AI aids in the process of transferring decision making from humans to machines, based on predefined rules. In the field of computer science, AI is also defined as the study of intelligent agents. An intelligent agent is any device that learns from the environment and makes decisions based on what it has learned to maximize the probability of achieving its predefined goals.

AI is capable of solving an extremely broad range of problems. These problems include, but are not limited to, simple mathematical puzzles, finding the best route from one location to another, understanding human language, and processing huge amounts of research data to produce meaningful reports. The following is a list of capabilities that the system must have in order to solve these problems along with a brief description of what each means:

  • Reasoning: The ability to solve puzzles and make logic-based deductions
  • Knowledge representation: The ability to process knowledge collected by researchers and experts
  • Planning: The ability to set goals and define ways to successfully achieve them
  • Learning: The ability to improve algorithms by experience
  • Natural Language Processing (NLP): The ability to understand human language
  • Perception: The ability to use sensors and devices, such as cameras, microphones, and more, in order to acquire enough input to understand and interpret different features of the environment
  • Motion: The ability to move around

How does a machine learn?

Let's take a quick look at the basics of machine learning. There are three methods that a machine can use in order to learn: supervised learning, unsupervised learning, and reinforcement learning, as described in the following list:

  • Supervised learning is based on the concept of mining labeled training data. The training data is represented as a pair consisting of the supplied input (also known as a feature vector—this is a vector of numbers that can represent the inputted data numerically as features) and the expected output data (also known as labels). Each pair is taggedwith a label. Thefollowing diagram illustrates the supervised learning method:

  • Unsupervised learning is based on a situation where the training data is provided without any underlying information about the data, which means the training data is not labeled. The unsupervised learning algorithm will try to find the hidden meaning for this training data. The following diagram illustrates the unsupervised learning method:

  • Reinforcement learning is a machine learning technique that does not have training data. This method is based on two things—an agent and a reward for that agent. The agent is expected to draw on its experience in order to get a reward. The following diagram depicts the reinforcement learning method:

Software requirements for the implementation of AI

The open source movement (which will be discussed in the Importance of accessible banking section) propels software development. The movement is coupled with the improvement of hardware (for example, GPU, CPU, storage, and network hardware). It is also supported by countless heroes who work on improving hardware performance and internet connectivity. These technicians have developed the AI algorithm to the point where it delivers near-human performance.

The following diagram depicts the typical technology stack that we should consider whenever we implement software to perform machine learning projects:

The following table breaks down several key technologies that contribute to the different software components mentioned in the preceding diagram:

Serial no.

Components

Software/package name

Software/package description

1

User interface/application programming interface

API/Python

API: An application programming interface is a type of interface that allows a program to interact with another program using the internet protocol. In comparison to the UI, the API is meant for a robot. It will be used to pull data from data sources throughout the coding chapters of this book, where we will create consumer banking services for an open bank project.

2

Machine learning and analysis

TensorFlow, scikit-learn, and ImageNet

Google's TensorFlow (https://www.tensorflow.org/) has been one of the most popular frameworks for deep learning since 2017. Scikit-learn (https://scikit-learn.org/stable/) is a handy machine learning package that delivers lots of useful functionalities in machine learning pipelines. TensorFlow and Keras (https://keras.io/) will be used when we work on deep neural networks, while we will use scikit-learn in less complex networks and data preparation works. These libraries will be used throughout the book, from chapter 2 to 9, to build machine learning models. ImageNet (http://www.image-net.org/) was created by Princeton University in 2009 to aid researchers in testing and building a deep learning model based on a dataset, which led to flourishing research on image recognition using deep learning networks. We will be converting an image recognition network to identify stock trends in Chapter 6, Automated Portfolio Management Using Treynor Black Model and ResNet.

3

Data structure

Pandas and NumPy

Pandas (https://pandas.pydata.org/) and NumPy (http://www.numpy.org/) are data structures that allow Python to manipulate data. They are used throughout this book's coding samples. These libraries are one of the key reasons for Python's popularity among data scientists. These libraries are used from chapter 2 to 9.

4

3D acceleration

Nvidia

The computation performance of Keras-related coding, such as the coding found in Chapter 3, Using Features and Reinforcement Learning to Automate Bank Decisions, will be enhanced if 3D acceleration (such as the software and hardware provided by Nvidia (https://www.nvidia.com/en-us/)) is used in the backend by TensorFlow. The driver will help to improve certain elements of GPU performance.

5

Operation systems

Ubuntu

This is a free, open source operating system that is compatible with most of the Python libraries we will use in this book. It is arguably the operating system of choice for the AI community.

6

Programming languages and development environment

Python and IDLE

Python programming is the language of AI. Python's existence is due to funding by DARPA in 1999, which was granted in order to provide a common programming language in a plain, readable style. It is open source. IDLE is a development environment that lies within the Python package. It allows programs to be written, debugged, and run. However, there are many more environments available for developers to code in, such as Jupyter Notebook, Spyder, and more. We will use Python and the Integrated Development and Learning Environment (IDLE) for easier code development (you can find them at https://docs.python.org/3/library/idle.html).

7

Version control

GitHub

GitHub is one of the most popular cloud-based collaboration sites. It was made possible because of the proliferation of cloud technologies, which enable scalable computing and storage. This is where our code base will be housed and exchanged.

With our brief introduction to the tools, technologies, and packages that we will use throughout the course of this book complete, let's now move on to explore an important area of AI—deep learning. The following section will explain deep learning and neural networks in detail.

Neural networks and deep learning

In addition to the open source movement, research breakthroughs in neural networks have played a big role in improving the accuracy of decision making in AI algorithms. You can refer to Deep Learning (https://www.deeplearningbook.org/)by Ian Goodfellow, Yoshua Benjio, and Aaron Courville for a more mathematical and formal introduction, and you can refer to Deep Learning with Keras (https://www.packtpub.com/big-data-and-business-intelligence/deep-learning-keras?utm_source=github&utm_medium=repository&utm_campaign=9781787128422) by Antonio Gulli and Sujit Pal for a concise analysis for developers.

Deep learning is a special subfield or branch of machine learning. The deep learning methodology is inspired by a computer system that is modeled on the human brain, known as a neural network.

Online customer support by banks via a mobile or web application chatbot is an excellent example of deep learning in banking. Such applications (that is, chatbots) are powerful when it comes to understanding the context of customer requests, preferences, and interests. The chatbot is connected to backend applications that interact with data stores. Based on the customer's inputs or selection of services, the chatbot presents to the customer various alternative sub-services to choose from.

The chatbot or deep learning applications work in layers. It can be compared to learning a language. For instance, once a person masters the alphabet by rigorously learning how to identify each letter uniquely, they will be eligible to move on to the next layer of complexity—words. The person will start learning small words and then long words. Upon mastering words, the person will start forming sentences, understanding grammatical concepts at different layers of complexity. Once they reach the top of this hierarchy of layers of complexity, the person will be able to master the language.

You might have noticed that in each phase or layer of the hierarchy, the learning becomes more complex. Each layer is built based on the learning or knowledge gathered from the previous layer of complexity. This is how deep learning works. The program keeps on learning, forming more knowledge with new layers of complexity based on the knowledge received from the previous layer. The layered complexity is where the word deep was taken from. Deep learning is a type of unsupervised learning, so it is much faster than supervised learning.

The major impact of deep learning is that the performance of the model is better as it can accommodate more complex reasoning. We want financial decisions to be made accurately. This means that it will be more cost-effective to give theshareholders of banks a reasonable return while balancing the interests of the bank's clients.

What we expect from a smart machine is as simple asinput,process, andoutput, as shown in the following diagram:

In most financial use cases, we deploy supervised learning, which resembles the process of training an animal—here, you provide a reward for a correct outcome and discourage an incorrect outcome. That's why we need to have the outcome (that is, the target variable) for training to happen.

Hardware requirements for the implementation of AI

While setting the budget for the hardware required by a bank, you need to ensure that it encapsulates the right configurations. This will allow you to deliver the promised results in terms of financial results or time to market, especially now that you are about to start a bank from scratch!

You'd better be sure that every penny works, given that the economic pressures on banks are pretty high. In order to do any of this, we need to understand the contribution that hardware makes to AI in order to ensure we have the right resources.

Graphics processing units

Besides the software and algorithms, the use of a Graphics Processing Unit(GPU) and Solid-State Drive(SSD) helps to speed up machine learning. The use of GPUs and SSDs makes it possible for a computer to think intelligently.

A GPU is a specially designed circuit that can process calculations in a parallel manner. This applies to computer graphic processing, where each of the pixels needs to be processed simultaneously in order to produce a full picture. To visualize this, suppose that there are 10 pixels to be processed. We can either process each of the 10 pixels one by one, or we can process them in 10 processes simultaneously.

TheCPU has the unique strength of having a fast processing time per pixel, while the GPU has the strength of multiple threads to handle flat data all at once. Both CPUs and GPUs can do parallel data processing with varying degrees. The following table shows the difference between sequential and parallel data processing:

Sequential data processing

Parallel data processing

Data comes in sequences, which requires a longer time to complete the computation.

Data comes in parallel, which improves the processing time.

Aside from being great at processing images, a GPU is also leveraged for deep learning. Although deep learning describes the number of layers the neural network has, deep neural networks are often characterized as having a wide record and lots of variables to describe the input.

When used in combination with a GPU, the SSD also improves the speed to read and write data to the CPU/GPU for processing.

Solid-state drives

Another hardware requirement for machine learning is a storage device called an SSD. The traditional hard disk has a mechanical motor to place the head that reads or writes data at a designated location on the magnetic tape or disk. In contrast to this, the SSD reads and writes data using an electric current on a circuit without the movement of a motor. Comparing the mechanical movement of motors with the electric current onboard, an SSD has a data retrieval speed that is 20 times faster.

For students in operation research, comparing the two is as simple as identifying the hardware capacity, which is akin to how we design a factory—find the capacity and reduce the bottlenecks as much as possible!

Modeling approach—CRISP-DM

CRISP-DM refers to a cross-industry standard process for data mining. Data mining is the process of exploring large amounts of data to identify any patterns to be applied to the next set of data to generate the desired output. To create the models in this book, we will use the CRISP-DM modeling approach. This will help us to maintain a uniform method of implementing machine learning projects. The following diagram depicts the project execution using the CRISP-DM approach in a machine learning project:

As you can see in the preceding diagram, there are various phases of the CRISP-DM approach. We can explain them in detail, as follows:

  1. Business Understanding: This phase involves defining the business objectives for the project. During this phase, you clarify the queries related to the core business objectives. For example, a core business objective may be to predict when the customers leave a particular website using the historical data of the customer's interaction with the website. The relevant query to clarify might be whether the payment interface currently in place is the reason for customers navigating off the website. Business success criteria are also laid out during this phase of the project execution.
  2. Data Understanding: This phase involves understanding historical data that is mined in the database or data store. The data is analyzed for its size, format, quantity, number of records, significance in relation to business, fields, source of data, and more.
  1. Data Preparation: This phase involves raising the quality of the data to the level required for the machine learning algorithms to process it. Examples of data preparation include formatting data in the desired format, rounding the numbers to an acceptable degree of precision, and preparing derived attributes.
  2. Modeling: This phase involves selecting a modeling technique or algorithm to be applied. A modeling algorithm is used to find a function that, when applied to an input, produces the desired output.
  3. Evaluation: This phase involves assessing the accuracy of the training model that was built in the previous phase. Any required revisions to the model are made in order to increase efficiency and accuracy.
  4. Deployment: This phase involves defining a deployment strategy for the training model in the live environment to work on new data. The models are monitored for accuracy.

After roughly covering what AI is, how machines learn, and the methods of AI implementation, it is now time to look at banking as a sector or industry. In the following section, we will explore the various types of banking and the challenges involved in the banking sector.

Understanding the banking sector

The banking sector is defined as a devoted economy for holding specific types of financial assets using methods that will make said assets grow financially over a period of time. Banking sectors are governed by rules imposed by governments or similar bodies.

Renowned author and financial consultantStephen Valdez described in his work, Introduction to Global Financial Markets (please visit https://www.macmillanihe.com/companion/Valdez-Introduction-To-Global-Financial-Markets-8th-Edition/about-this-book/), the different types of banking in the global financial markets. These are commercial banking, investment banking, securities firms, asset management, insurance, and shadow banking.

These types of banking are required to fulfill the needs of a wide variety of customers, ranging from large organizations to individual customers. The following is a description of these various types of banking based on the needs of customers:

  • Commercial bankingcan be retail (serving consumers) or wholesale (serving companies). Essentially, banks focus on taking deposits from savers and lending them to borrowers by charging interest. Commercial banks thrive on their ability to assess the riskiness of the loan extended to borrowers. Any failure to accurately assess the risk can lead to bankruptcy due to the failure to return money to the depositors. Many banks have failed in financial crises, including Washington Mutual in the US.
  • Investment bankingincludes advisory businesses and security trading businesses. Advisory businesses deal with the buying and selling of companies, also known asmergers and acquisitions(M&A), debt and equity capital raising (for example, listing companies on the New York Stock Exchange), and security trading businesses. The security trading businesses deal with the trading of stocks, fixed income, commodities, and currencies. Securities trading involves a buyer who is willing to buy a security, a seller who is willing to sell a security, and a broker who facilitates the buying and selling of a security.

The advisory businesses hinge on creating value for companies by combining or spinning off businesses. This process optimizes the organizational performance for M&A activities. It also optimizes the cost of capital for clients into a standardized borrowing structure (such as bonds). The clients can do more investment by issuing new shares or canceling existing company shares (equity) to financial market participants.

All of the aforementioned activities create value with the correct evaluation of the companies given by the participants of the markets, which are driven by moods and more rational concerns.

  • Asset managementincludes funds of all kinds—mutual funds, exchange-traded funds, hedge funds, private equity, and more. Asset management companies invest in various types of financial assets and the various life stages of a corporation using different investment strategies (a combination of buying and selling decisions). A critical decision made in this industry also falls under the umbrella of proper valuation, with regard to an investment's future values.

Asset management participants have a hunger for generating returns to meet various purposes, from the protection of asset values to appreciation. They are typically referred to as the buy side, which represents the asset owners, while the banking services that help the buy side are referred to as the sell side, which typically includes securities sales (client-facing, gathering orders), trading (executing the orders), and research (evaluating the securities).

  • Insuranceincludes general insurance and life insurance. Life insurance protects buyers from mortality risks (consequences of death), and non-life insurance covers everything else, such as loss due to disasters, the loss of luggage, the loss of rockets (for example, Elon Musk's SpaceX loss) and vessels, system breaches due to hacking or viruses, and more.

The core function of insurance is to estimate the risk profile of borrowers. On the other hand, the ability to generate investment returns to cover losses can be important as well. The stronger the investment performance of the insurer, the more aggressive the pricing of insurance it can offer and the more competitive it becomes. That's one of thereasons why Berkshire Hathaway can provide competitive insurance pricing—due to its superior investment performance.

  • Consumer banking isrepresented by the asset size of consumer debts,which focuses on the mortgage, auto, and personal loans, and credit card businesses that we might need at various points in our life.
  • Shadow bankingis a lending settlement involving activities outside the regular banking system. It refers to alternative investment funds, such as bitcoin investment funds, broker-dealers in securities, and consumer and mortgage finance companies that provide lending to consumers.

The size of banking relative tothe world's economies

By comparing the sheer size of the finance industry with the world's annual income from production, we get a fair sense of how the world uses banking services to support itself. However, it is rather abstract to only show the statistics. Let's say the world is a person. How does finance fit into this person's life? The following is a list of points to consider:

  • Annual income:The productivity and, therefore, income of the global economy as gauged by the World Bank was $86 trillion in 2018. Roughly, one-fifth (19%) of the annual income comes from trading across borders (where export trade volume is at $15 trillion).
  • Wealth:The global person has approximately 4.4 years equivalent of annual income (annual GDP). A breakdown of the annual GDP can be found in the table at the end of this section. The information on annual income has been derived from various sources by comparing the activities with the size of the GDP. These 4.6 years can be bifurcated as follows:
  • 0.9 years has been with the asset manager.
  • 0.9 years has been deposited in banks.
  • 0.8 years has been in the stock markets.
  • 2.3 years has been funded by credit/borrowing (1.17 through debts, 1.0 through bank loans, 0.5 through shadow banks, and 0.03 through consumer credits).

Of course, this is a simplified treatment of global wealth; some figures could be double-counted, and the stock market figure could include deposits placed by listed companies that are accounted for by bank liabilities. However, given that we want to understand the relative size of various financial activities and their importance, we've just taken a shortcut to show the figures as they are.

  • Insurance:To protect against any kind of undesirable risks derived from productive or investment activities,6% of the global person's annual income was spent on the insurance that covers 1.45 times their equivalent income. The premium will be used to buy the underlying financial assets to generate income to offset any undesirable risks.
  • Derivatives:As a risk-protection instrument, besides buying insurance, banks can also offer derivatives as a financial instrument to offer risk protection. The termderivativesrefer to the agreement between two parties to pay or receive economic benefits under certain conditions of underlying assets. The underlying assets vary fromfixed income and currency to commodities (FICC).

Fixed income includes the interestrate and credit derivatives. Currency refers to foreign exchange derivatives, and commodities refer to commodity derivatives. Foreign exchange came in second with $87 trillion of outstanding exposure, which is roughly equal to the world's GDP. Commodity, credit, and equity derivatives have smaller shares, with each at around 2% to 9% equivalent of GDP. When accounting for derivatives as a risk-protection instrument,we exclude a form of derivatives called the interest rate over-the-counter (OTC), which is equal to 6 times the annual income—this is far more than the annual income that our wealth requires for protection. Indeed, some investors take the interest rate OTC as an investment. We carve out this instrument for our overall understanding of insurance. OTC refers to the bilateral agreements between banks and bank customers.

Another form of agreement can be exchange-traded agreements, referring to bank customers buying and selling products via a centralized exchange. I did not include too many exchange-traded figures, but the figures mentioned in this point for foreign exchange, commodity, credit and equity, and so on, serve the purpose of showing the relative size of the sectors.

The following table lists the GDP figures:

Trillions in USD in 2018

% of GDP

Income

75.87

100%

World's GDP (annual income generated globally)

75.87

100.00%

Global export volume

14.64

19.00%

Wealth

332.46

438%

Global asset management

69.1

91.00%

Global bank liabilities (including deposits)

58.93

78.00%

Global stock markets

79.24

104.00%

Global debt markets

57.49

76.00%

Bank loans

29.7

39.00%

Shadow banking

34

45.00%

Global Consumer Debt

4

5.00%

Global insurance (new premium written)

4.73

6.00%

Insurance coveragederivatives (ex-interest rate OTC)

110.15

145.00%

Global foreign exchange OTC + exchange-traded derivatives

87.41

115.00%

Commodity OTC contracts

1.86

2.00%

Credit OTC derivatives

9.58

13.00%

Equity-linked contracts

6.57

9.00%

Interest rate OTC contracts

461.98

609.00%

All figures were earlier reported for the full-year figures of 2018 unless otherwise stated. GDP and stock market sizes are from the World Bank; export trade data is from the World Trade Organization; new insurance premium figures are from Swiss Re Sigma for 2018; the global asset management size is from BCG Global Asset Management for 2018; all banking, debts, and derivatives statistics are from the Bank for International Settlements.

Customers in banking

Customers in the finance industry include depositors and borrowers engaged in saving and lending activities. When engaging in commercial banking activities, such as cross-border payment or trade finance, they are calledapplicants(senders of funds) andbeneficiaries(receivers of funds).

If customers are engaged in investment banking, securities, and asset management activities, they are calledinvestorsor, generally,clients. To protect buyers of insurance products from potential risks, the personbuying is called theproposer,and the item is called aninsured item. In cases where risk occurs and if/when compensation is required from the insurers, the person to be compensated is called abeneficiary.

Non-financial corporations are the real corporate clients of all financial activities and should be considered the real players of economics. They save excess cash and produce goods and services for consumers.

A message that I wish to clearly underline and highlight is that finance is a service to real economies. So why does financial sector growth surpass real economic growth? Well, as per the opinion of Cecchetti and Kharroubi, too much finance damages the real growth of economics. That is, it takes away high-quality research and development talents that could contribute to real economies. Therefore, the taking away of talented people negatively impacts production factors. You can find out more about this at https://www.bis.org/publ/work490.pdf.

Importance of accessible banking

Like electricity and water, banking should be made as widely and easily available as utilities. Only when we make banks efficient can we make them accessible and have them benefit the highest number of people possible. Banking is a service that is provided to make the best use of capital/money to generate returns for those whosaveand/orthose who need the capital to have a more productive life at an agreed risk and return.

What we want to do is to be consistent with Robert J. Shiller's sentiment in his book,Finance and the Good Society, where he indicates the necessity of information technology in finance to help achieve our goals. A step further would be to leverage open source methods and applications to solve the accessibility challenges of the banking industry. Open source software solutions tend to be cost-effective, robust, and secure.

To make banking accessible, one of the most important things to do is to have a lot of data. This will make decisions more efficient and transparent, which can help to reduce the cost of banking decisions. We will discuss the need for open source data in the next section. By virtue of the competitive banking market, the price of banking services will gradually decrease as banks with good efficiency will win a large market share.

Once implemented in the financial sector, AI will have three impacts on the sector—the job of repetitive tasks will be eliminated, there will be increased efficiency with AI augmenting human, and job creation with new AI-related tasks such as model building. Out of these three, job reduction and increased efficiency will impact existing jobs, whereas job creation will have an impact on future talent and the job market.

As automation and efficiency increase, existing jobs will be altered and impacted. Machines will perform day-to-day tasks with more efficiency than humans can. However, to manage, monitor, maintain, and enhance tasks performed by machines or AI, the industry will become open to skilled, techno-functional professionals who understand both banking and AI technology.

Open source software and data

The speed of technological development in the past 20 years or so has been quite rapid due to the open source movement. It started with Linux and was followed by ImageNet. ImageNet provided lots of training data. This training data fueled the activities of technicians who worked in research on developing AI algorithms. These technicians developed algorithms for deep learning and neural networks using open source libraries written in programming languages such as Python, R, scikit-learn, TensorFlow, and more.

While the open source approach encourages software development, another key ingredient of AI is data. Finding practical open data is a challenge. Banks, on the other hand, have the challenge of converting data into a machine-trainable dataset cautiously and safely to make sure that there is no breach of data that customers entrusted the bank with.

Today, in the finance and banking world, client confidentiality remains a keyobstacleto opening up data for wider research communities. Real-world problems can be more complex than what we have seen in the open data space. Opening up data stored in databases can be a practical step, while opening up images, such as documents, audio files, or voice dialogues, for example, can be challenging as this data, once masked or altered, may lose some information systematically.

In fact, the major cost of implementing real-life applications in banking also comes from the data-feed subscription. The cost of data collection and aggregation is a major challenge that you will see in this book. How our society is handling this problem and incentivizing the commercial sector to tackle it requires further discussion beyond the scope of this book. Following this same spirit, the code for this book is open source.

Why do we need AI if a good banker can do the job?

Let's consider a single financial task of matching demand for capital in the funding markets. This is a highly routine task of matching numbers. Here, it is obvious that the computer would be a better fit for the job.

The goal of employing AI is to make machines do the things that humans do right now, but with more efficiency. Many people wonder whether applying AI in banking might affect the jobs of those working in the industry.

Do remember that the aim is not to replace humans, but to augment the current human capacity to improve productivity, which has been the goal of technology throughout the history of human civilization. Humans are known to be weaker in determining accurate probability, as shown in the psychological research paper, Thinking, Fast and Slow, April 2, 2013, by Daniel Kahneman. Therefore, it is challenging to make a probability decision without a computer.

Applications of AI in banking

According to theMcKinsey Global Institute (https://www.mckinsey.com/~/media/mckinsey/industries/advanced%20electronics/our%20insights/how%20artificial%20intelligence%20can%20deliver%20real%20value%20to%20companies/mgi-artificial-intelligence-discussion-paper.ashx), outof 13 industries, financial services ranked third in AI adoption, followed by the high-tech, telecommunications, and automotive and assembly industries.

As theMckinsey report does not mention the use case in banking, with a bit of research, perhaps we can take a look at the four ways in which AI creates values, as shown in the following list:

  • Project: Forecast and anticipate demand, improve sourcing, and reduce inventory (capital).
  • Produce: Provide services at a lower cost or higher quality.
  • Promote: Provide offers for the right price with the right message for the right customers at the right time.
  • Provide: Rich, personal, and convenient user experiences.

Let's examine how each finance participant applies AI to the following aspects, as shown in the following table:

Participants

Project: better forecast

Produce: lower processing cost

Promote: personalized offer

Provide: convenience

Commercial banks

Optimize funding needs.

Using AI, trade finance processing can be automated, which will result in increased efficiency.

AI can provide a real-time quotation of export/import financing as the goods move to different stakeholders with different types and levels of risk.

Improve client services with an NLP-enabled chatbot.

Investment banks

Valuation of corporations.

With AI, it becomes faster and cheaper to reach the market signal by identifying the market's sentiments.

AI can match the needs of asset sellers and buyers through automated matching.

Mobile workforce with access to information at any time.

Asset management

Asset valuation and optimization.

AI can help here by automating trading and portfolio balancing.

AI can recommend investments to customers.

Fast and convenient portfolio updates.

Consumer banks

Project a realistic savings plan.

Personalized bot advisers can capture the data from receipts without human help.

AI can understand the right time at which consumers need financing or investment products.

Serve clients 24/7 anywhere using smart bots.

Across the board, we can now see how data is being leveraged for smart decision making in the field of finance: more data points and a higher speed of exchange can lead to a much lower cost of finance. More detailed examples will be provided in the following chapters.

How do we attain this lower cost? Essentially, we get it by having fewer hours spent working on producing an aspect of the banking service.

Impact of AI on a bank's profitability

To give you an idea of AI's impact on a bank's profitability, let's take a look at some simple estimates from two perspectives: the improvement of model accuracy and the time spent to run/train the model.

Over the past 10 years, the clock rate and the number of cores have improved tenfold, from around 300 cores to around 3,000 cores.

I have compared the shallow machine learning or statistical model I experienced a decade ago to what I see today with deep neural networks. The model accuracy of neural networks improves the model from around 80% to over 90%, with a 12.5% rate of improvement. The following table shows improvements in the memory data rate, bus width, and size:

Year

Processors

Core clock

Memory data rate

Memory bus width

Memory size

2007

8800 Ultra[42]

612 MHz

2.16 GHz

384 bit

768 MB

2018

Titan X[43]

1417 MHz

10 GHz

384 bit

12 GB

2018

GeForce RTX 2080 Ti

1545 MHz

14 GHz

352 bit

11 GB GDDR6

The following table highlights the improvement in the areas of banking:

Areas

Improvement

Areas of banking

Project: better forecast

Model forecast accuracy improves by 15%.

Risk model, pricing

Produce: lower processing cost

Automation rate of 50%.

Operations

Promote: personalized offers

Model forecast accuracy improves by 15%.

Risk model, pricing

Provide: convenience

Reduces delay by 50% if all processes are automated.

Operations

If the cost-to-income ratio is around 70% for banks, then the automation rate will likely reduce the ratios by half to 35%. However, the cost of technology investment will take up another 5-10%, taking the target cost-to-income ratios from 40% to 45% following the proliferation of AI in banking. This will impact the banks in developed countries more, as the cost of labor is quite high compared to emerging markets.

Improving the accuracy of forecasts will reduce the cost of the bank's forecast ability further, which, in turn, will reduce the cost of risk by 15%. My personal view is that, for developed countries, the cost of risk is at 50 basis points (bps) of the overall asset; a reduction of 15% on the bank's cost of risk cannot have a significant impact on their profitability.

The improvement in forecast accuracy and convenience will improve the accessibility of banks, which means they can reach a bigger market that was not considered feasible in the past. That is, the profitability ratio of return on the equity does not reflect the impact; rather, it will be shown in the size of the bank and the market capitalization of banks. It should generate an improvement of 15% given a wider reach.

All in all, there is an improvement in return by 80%, from 7.9% Return on Equity (ROE) to 14.5%. However, there will be additional capital requirements for systemically important banks from 11% to 12%, gradually, which will drag the overall ROE down to13.3%in the target AI adaption stage, with all regulations settling in.

Summary

We began this chapter by explaining what AI is all about. AI is the technology that makes machines perform tasks that humans can do, such as weather prediction, budget forecasting, and more. It enables machines to learn based on data. We looked at the various techniques of AI, such as machine learning and deep learning. Later, we looked at the complex processes of the banking domain. If we can automate them, we can reduce costs in the banking sector. We also learned about the importance of accessible banking. Later, we looked at the application of AI in the banking sector and its positive impact, with a few numbers to support it.

In the next chapter, we will continue our journey of AI in banking. As a next step, the chapter will focus on time series analysis and forecasting. It will use various Python libraries, such as scikit-learn, to perform time series analysis. The chapter will also explain how to measure the accuracy of machine learning-based forecasting. The chapter will be full of interesting content and will teach you how to combine financial ratios with machine learning models. This will provide a more in-depth look into how machine learning models can be applied to solve banking problems.