AI in Financial Services - A buzzword that is here to stay!

In a few of my most recent blogs I tried to demystify some of the buzzwords (like blockchain, Low- and No-Code platforms, RPA…​), which are commonly used in the financial services industry. These buzzwords often entail interesting innovations, but contrary to their promise, they are not silver bullets solving any problem.

Another such buzzword is AI (or also referred to as Machine Learning, Deep Learning, Enforced Learning…​ - the difference between those terms put aside). Again this term is also seriously hyped, creating unrealistic expectations, but contrary to many other buzzwords, this is something I truly believe will have a much larger impact on the financial services industry than many other buzzwords.
This opinion is backed by a study of McKinsey and PWC indicating that 72% of company leaders consider that AI will be the most competitive advantage of the future and that this technology will be the most disruptive force in the decades to come.

Deep Learning (= DL) is a subset of Machine Learning (= ML), which is also a subset of Artificial Intelligence (= AI). AI is the all-encompassing concept that initially erupted.

Nonetheless to correctly grasp the impacts of AI, it is also important to first clear the smokescreen and understand that there is no magic under the hood. AI is (today) nothing more than a mathematical technique (based on the development of neural networks) for advanced pattern recognition, where statistical correlations identified in large sets of data, are modelled in an advanced way. High level it all comes down to calculating the right weights in an enormous mathematical matrix, which maps the available inputs on the outputs (at least in supervised learning, which is the most used in the industry).

In supervised learning, a model is trained based on an existing training set (of in- and outputs). The trained model is then applied to other data. The other type is unsupervised or reinforcement learning, which works with a cost function (i.e. the objective) which should gradually be optimized by adapting the model. Unsupervised learning has the advantage of not requiring large labelled sets of training data, but the disadvantage is that you need quick and good feedback on the resulting cost function (objective). This means the technique is very good for e.g. recommendation engines (as you can see immediately if a recommendation is clicked on or not), but not so good for credit risk scoring, as a credit default is often months (if not years) away from the credit decision moment.

This technique was till recently only used in academic circles, but has now become widely accessible in the industry, thanks to:

• User-friendly tooling (like TensorFlow, PyTorch, Keras…​), abstracting away the complexity of the mathematical modeling behind.

• The exponential increase of available computing power (mainly due to the rise of the cloud providers), allowing to train much more complicated models

• The available data sets that increased enormously, as more activities are digitized, but also as more data can be stored in a cheap and efficient way.

Nonetheless the current AI techniques (sometimes referred to as "narrow AI" or "applied AI", as the model only performs well for a very narrowly defined task) are still very far away from the concept of General AI, which can apply intelligence to any problem.

More than 15 years ago, I did my thesis on Non-linear system identification, which was not very different from the current AI techniques, except that we derived much smaller models with only a few dozens of weights to estimate. The aim however was the same, i.e. model an unknown system - handled as a black-box - based on a number of well-defined inputs provided to the system and their associated measured outputs.
In the same year, a friend of mine who was studying economics, asked me for some help with his exam on "Multi-variate analysis", which also came down to pretty much the same thing.
It is interesting to remember this basis, because these "simpler" models apply similar principles, but are much easier to understand. E.g. you can give a meaning (a physical explanation) on every calculated weight in the model, which is no longer possible now (as the models are too complex to understand).

Although AI is very similar to these simpler modelling techniques, AI is applied on a much larger scale and is fully automated, meaning AI can help us to model large amounts of (unstructured) data and can evolve the derived model automatically based on new (changing) data.
Compared to a human being, speed, accuracy and lack of bias are some of the main advantages.

Nonetheless this unknown behavior of the model (i.e. the fact that a human being can no longer explain the model) gives also some crazy results:

• Ability to trap the model, e.g. by changing only 1 pixel in certain images an AI model to identify faces can be misled

• Certain conclusions cannot be explained. There are some examples of AI correctly identifying heart conditions based on ECGs, but cardiologists and researchers unable to explain on which basis the AI system came to this conclusion.
  For some industries, the importance of the prediction’s explanation might even surpass the importance of the prediction itself, often it is even a legal explanation. E.g. GDPR gives the right to every person to obtain an explanation about an automated decision. Transparency on AI models will become crucial, hence the rise of explainable AI models.

• Discrimination is very hard to avoid. Even when you avoid providing discriminatory inputs to the model (like gender, race, age…​), the model might still be discriminative, as it might derive the discriminatory characteristics from other inputs. Furthermore, even if the model is not discriminatory, it is still nearly impossible to proof to external auditors and regulatory authorities that this claim is justified.

• Techniques exist to know if an input was used as data to setup the model or not. This is of course info you don’t want to share with every user of the AI model.

Additionally, due to the internal complexity (and abstraction of this complexity) of AI models, some basic principles of mathematical modelling tend to get overlooked, despite that those principles apply just as well to AI modelling:

• With statistics you can proof everything: it is important to have sufficient data for each category/segment, otherwise for certain categories/segments the modelling will only be based on a few data points, which can give very wrong statistical results. As in statistics, the choice of a sufficiently diverse data set, is therefore very important.

• Avoid over-fitting: the number of data points should be considerably more than the number of coefficients (weights) to be estimated in your model. As AI models are very large (thousands of weights), you need hundreds of thousands of data points in order to avoid over-fitting.

• Splitting your data set: when training a model, it is important to split your data set in 2, i.e. the first half should be used for training, while the second half should be used for verification of the quality of the model. Too many people still verify the accuracy of their model with their training set, which can give very wrong conclusions.

• Garbage in, garbage out: an AI model is only as good as its training data. If this data contains too much errors, the model will also be poor. As a result, data cleansing is crucial in AI modeling.

• Don’t ignore what you do know about the system being modelled: most systems are not really black boxes, i.e. you typically know how your business conducts or how a product works, meaning you can model already a big part of the system as a white-box. It would be a pity to throw all this knowledge aside, when setting up an AI model. In many cases a rule-based system, in which you define yourself the business rules from your understanding of the business/process/product can give an equally good result if not better. It can therefore be interesting to use a combination, i.e. a rule-based system or programmed algorithm, which is fine-tuned by an AI model.

• …​

All the above arguments don’t reduce however the importance of AI as a technique to bring our digital world to a next level, as long as we understand that there is no magic and that AI is not the solution to any problem.

The technique can however have a profound impact in almost every domain. Most common examples are:

• Gaming: with the very known examples of AlphaGo of Deepmind Technologies (Google) beating the world-champion (Lee Sedol) in Go in 2016 and IBM Watson playing Jeopardy.

• Natural Language processing (NLP): used a lot in chatbots or in interpreting unstructured documents (e.g. for classification)

• Real-time translations, which not only uses NLP to understand the text to be translated, but also to compose the new text in the destination language

• Image recognition: identify patterns in pictures and videos. Used for face recognition, self-driving cars, but also in quality control in production lines

• Speech recognition: used in speech-based systems like Alexa (Amazon), Google Assistant and Siri (Apple)

• Handwriting Recognition: used to improve the quality of OCR, but also for recognizing signatures

• Prediction: predicting the future behavior of a user, e.g. for recommendations (predicting next desired action of a user), credit risk scoring (predicting if customer will not default on his credit) or insurance actuary models (predict the number and amount of insurance claims…​),

• Expert Systems: such as interpreting medical X-ray pictures, blood results or ECGs, spam filtering, targeted advertisement…​

• …​

In the financial services sector more specifically it is commonly used in:

• Robo-advise in order to assist (allowing the advisor to increase the number of portfolios he is advising) or even replace an investment portfolio advisor

• Scoring models for credit scoring and insurance underwriting

• Pricing engine for intelligent pricing, i.e. set the highest price the consumer is willing to accept and set a risk-based pricing (like for determine credit interest rate and insurance premiums)

• Recommendation engines for advising the next-best offer and proposing the best product for a specific need/project (e.g. best financing solution for a specific project)

• Ability to read and process unstructured data for credit analysis (especially for business credits) and Know Your Customer

• Anti-Money Laundering (AML) to better identify complex money laundering mechanisms via intelligent identification of cases (suspicious anomalies) and prioritizing those cases

• Know Your Customer (KYC): analyze an onboarding request and use maximum of available data to determine the client risk. AI can be used in multiple forms for automating the KYC process, e.g. face recognition (matching ID card image with camera image), NLP processing (to interpret unstructured data) and expert system to analyse the request (e.g. identify anomalies) and take a decision.

• Cyber-security: usage of AI to identify anomalies in usage of systems and access to systems, to immediately identify potential breaches (see blog https://bankloch.blogspot.com/2020/02/securing-your-doors-is-not-enough-go.html), but in risk-based authentication (see blog https://bankloch.blogspot.com/2020/02/multi-factor-authentication-and.html) to authenticate a user in a more user-friendly and secure way (e.g. via AI identify deviating/abnormal user behavior resulting in a request for additional authentication).

• Fraud detection and prevention: identify in payment transactions automatically abnormal patterns, to immediately (real-time monitoring) refuse fraudulent transactions. But also in insurance claim handling, identify potential fraudulent claims.

• Chatbots and voice assistants (conversational banking): allowing to provide 24/7 personalized advise (mimic live employees), without overloading customer care departments. A nice example is the advanced AI bot "Kate" of KBC, which will not only help customers with financial questions, but is foreseen to offer advanced concierge services in the future.

• Automating manual back-end processes, like automatic reconciliation, data cleansing, data classification…​

• AI-based PFM and BFM services, like automatically identifying and predicting budget plans and upcoming cash flows, suggest deals (coupons or cashbacks) based on insights obtained on past transactions, cash management and liquidity planning…​

• Advanced trading: with AI complex models can be setup trying to predict how the market will evolve and in this way allow to generate revenues

• Matching parties on financial marketplaces (cfr. blog https://bankloch.blogspot.com/2020/06/marketplaces-in-financial-industry-here.html), which can profit a lot of AI in optimally matching the consumer and producer on such a marketplace.

• …​

It’s clear that AI can give enormous value to a bank or insurance company, but like with any technique, it’s important to make the business case and not fall in the trap of feeling the urge to be innovative and use AI at all cost. Millions are invested in AI algorithms, but for many of them it is debatable if the investment was worth it. In many use cases AI still has to prove it gives really better results than the carefully designed algorithms or rule-based system / expert systems, used before. E.g. new alternative credit scoring algorithms using AI and alternative data sets to serve the underbanked and unbanked are already starting to show their first cracks and seem not to be better than traditional scoring systems.
Time will in any case separate the chaff from the wheat, with regards to the best use cases to apply AI on.