In this paper, Dan Barnes and Mike O’Hara of The Realization Group investigate the real-world applications of AI within financial business today, taking use cases from experts in the field and exploring the choices firms need to face in order to overcome hurdles in implementation. Speaking with Rael Cline of MediaGamma, Jonty Field at Quantitative Brokers, Chuck Ocheret at NEX Optimisation, Massimi Morini at Banca IMI Intesa San Paolo and Software AG’s Michael Zeller and Charles Platt, we have built a picture of how firms are operationalising AI to achieve new approaches to service and develop truly digital business models.
The opportunity to grow business through artificial intelligence (AI) is closer than you think. Those technologies that enable applied AI, including pattern recognition and machine learning (ML) capabilities, are established at low levels within many technologies in financial services. Analysts see the real value for the future. Analyst firm Tractica has estimated the market for AI systems for enterprise applications could increase from US$202.5 million in 2015 to US$11.1 billion by 2024E, expanding at a compound annual growth rate of 56.1%. Sarj Nahal, equity analyst at Bank of America Merrill Lynch, has estimated that the adoption of robots and AI could boost productivity by 30% in many industries, while cutting manufacturing labour costs by 18-33%.
In a 2015 note entitled ‘The real consequences of artificial intelligence’, Goldman Sachs Global Investment Research observed that within financial services, the simple algorithms guiding robo-advisors are having a real impact on the costs for retail investors. Their all-in fee of 25 basis points (bps), or even free at broker Schwab, is on average 50 bp lower than the fee rate that non-robo incumbents charge, and is also 30 bp lower than the cost basis that incumbent firms face.
AI, as the operationalised model of these technologies, is developing at a breakneck pace but its potential and implications – including issues around decision making – create challenges that firms need to address before they can embed AI within their organisation. AI can handle data sets that humans cannot. It can find patterns within this information with speed and precision. Leaders in the space are already taking advantage of these abilities to build more elegant risk modelling algorithms. Traders can make fewer, better-informed choices whilst seeing a more complete picture of the market.
Financial services firms are employing machine-learning technology within their operations today, they are seeing the benefits today, and they are increasing their use in readiness for tomorrow. AI, machine learning models and predictive analytics are revolutionising business as the PC did in the early 1980s. The level of data generated by automating business has been revealing, yet the picture it has created is not always visible to the human eye. From anti-fraud systems, identifying everything from faces to patterns of behaviour, to trading platforms that determine how best to place an order for a securities trade, data is being mined by enterprises seeking competitive advantage, increased efficiency and more rigorous risk management. It is a natural extension of the digital revolution. Placing a PC on every desk, and then building that into a network, created flows of information. For functions based upon mathematical modelling, the development of these flows led to better understanding of actual behaviour, beyond reported behaviour.
“I don’t think about our data storage as static repositories of data, I think of them as multistage workflows, so that as you ingest data and as you transform it, you always persist all the intermediate states all along the way, and you never throw anything out,” says Chuck Ocheret, chief innovation officer for Optimisation at NEX Group. Having established pools of compute power and rivers of data, the next step has been to increase the complexity of the processing so the technology can gain not only quantitative but qualitative advantages from the information.
“This data aqueduct is constantly flowing, and you can always reroute some flow to somewhere else. This requires a tremendous amount of automation, sophistication, and versioning capability.”
Chuck Ocheret, Nex Optimisation
“This data aqueduct is constantly flowing, and you can always reroute some flow to somewhere else,” Ocheret says. “Of course, all this requires a tremendous amount of automation, sophistication, and versioning capability.”
A bottleneck in this process can be created by the capacity of humans to cope with the scale of data. Where machine learning comes into its own, is that the user can teach a system to understand the data, and build models based on that understanding. At that point solutions which humans could never think of are possible.
“We’re effectively experts in a generalised sense of analysing huge, huge datasets, finding signals in the noise and being able to output a decision in five milliseconds,” says Rael Cline, CEO and co-founder of MediaGamma, a London-based machine learning company.
There are two levels of AI, based upon the categorisation set by John Searle, Slusser Professor of Philosophy at the University of California in the 1980s. One is full AI that is truly conscious. The other is a level of AI that lacks true conscious understanding but that can follow the cognitive patterns of recognition and processing found in the human mind.
In a September 2016 research note, analysts at Morgan Stanley asserted that, “There is currently lots of hope around the Fourth Industrial Revolution and full Artificial Intelligence, which in a bull case could potentially lead to decades of global GDP growth above 3%.”
However, it also noted that resource constraints make that next stage of development beyond our reach at present: “We are still in the tail of the Third Industrial or Digital Revolution where investment in digitalisation could drive significant productivity gains.”
“We’re effectively experts in a generalised sense of analysing huge, huge datasets, finding signals in the noise and being able to output a decision in five milliseconds.”
Rael Cline, MediaGamma
Digitalisation uses the level of AI which includes machine learning engines and advanced analytics. “Predictive analytics is a very good place to start, moving up into machine learning and AI,” says Charles Platt, Head of Financial Services at Software AG. “What you need is a platform that can continue to grow in all of those stages as you go through.”
AI systems can enhance productivity via two approximate routes. One is to replace the processing that older systems were capable of with qualitatively better processing; the other is to enhance the decision-making that people are capable of. The operational challenge is to embed AI in such a way that it functions reliably and yet can scale within the organisation.
“Some of those techniques can be very helpful to refine an existing idea,” says Jonty Field, head of EMEA, at Quantitative Brokers. “There’s some value in portfolio construction, and for stressing portfolio construction.”
There must be a strong connection between the people generating the ideas and the technology itself. Field gives the example of the codebreakers at Bletchley Park cracking the Enigma code. While they had the machines to do the calculations they lacked the power to crack the code via ‘brute force’. But by figuring out that every morning and evening the messages would start with weather reports, and being able to actually see the weather, they would know approximately what the message was.
“AI is helpful where the human who has the domain expertise, be it an experienced trader, portfolio manager or risk analyst, can identify a clue and then use AI to dig deeper into it”
Jonty Field, Quantitative Brokers
“AI is helpful where the human who has the domain expertise, be it an experienced trader, portfolio manager or risk analyst, can identify a clue and then use AI to dig deeper into it” says Field.
Understanding the use cases is key to firms deciding to invest in AI. For senior management to see how the business may be transformed, they need to know precisely what the advantages are and how they can be operationalised in order to invest in these technologies. “How do you choose to invest ahead of time before you know that gold is buried there?” says Ocheret. “If you have a very specific problem you’re trying to address, then you can be more focused. That requires investment in infrastructure and space in order to orchestrate this and find the opportunity that exists.”
NEX Group’s post-trade operations have focused on high volume activity such as FX that requires a lot of operational automation. Where there is a need to correlate information from natural language documents with more structured data natural language understanding, AI systems can take context into account.
“We’re interacting with a lot of companies that are working on natural language,” says Ocheret. “It is most interesting for decision-making or decision enhancement and augmentation. How do you make people more effective at making decisions? How do you bring the right things to their attention, based on their activity, get them the information that is going to make the most difference to them, thus finding a good outcome?”
Another key potential for the application of AI is to allow manual processing to transition, by becoming learning systems, into automated systems. If it can be observed that a task is handled repetitively or that the same decision is always made, then either a system can make recommendations so decisions are faster, but approved, or just become fully automated because the same choice is always made.
Introducing such powerful digital technology into a business can be transformative, but it should not be disruptive. The business – encompassing the people behind it – must be carried through the transformation process. To operationalise models for AI, including predictive analytics, firms also need to integrate the technology into existing systems.
“It shouldn’t be a rip and replace; it’s really an enhancement of many of the existing business processes. Where you inject agents to make smarter and better decisions, they can be focused on risk or they can be focused on fraud or on customer experience.”
Michael Zeller, Software AG
“There has to be an extra enhancement of existing systems,” says Michael Zeller, senior vice president for AI Strategy & Innovation at Software AG, “It shouldn’t be a rip and replace; it’s really an enhancement of many of the existing business processes. Where you inject, let’s say, little agents to make smarter and better decisions, they can be focused on risk or they can be focused on fraud or on customer experience.”
Traditional risk modelling, such as the Black-Scholes model, are products of their time. They often contain assumptions, such as positive interest rates, and historically models such as Value-at-Risk (VaR) have been found wanting when stretched beyond their original purpose. For firms that need to create risk models for more complex scenarios, AI opens up new possibilities in algorithmic development.
Massimi Morini, head of Interest Rates, Credit and Inflation Models at Banca IMI Intesa San Paolo. says, “When I have to perform data selection, or when I have to optimise my algorithm, I’m interested in the cleverness of the algorithms I am using.”
To calibrate a model to different asset classes for a large portfolio, the amount of data is no longer manageable for the mathematical models that Banca IMI Intesa San Paolo has used in the past.
“Using the traditional algorithms you have a crazy amount of inputs and a crazy amount of calculations,” says Morini.
It has to take a different approach, bringing AI to bear on the creation of specific risk models. Initially it used to find patterns in data to reduce its complexity.
“First of all, we take the inputs of these optimisation algorithms that often use something like hundreds or thousands of inputs, and we use unsupervised learning techniques that look at the similarities in the data, and transform the original features into a reduced set,” says Morini.
“We take the inputs of optimisation algorithms that often use hundreds or thousands of inputs, and we use unsupervised learning techniques that look at the similarities in the data, and transform the original features into a reduced set.”
Massimo Morini, Banca IMI Intesa San Paolo
When looking at a specific problem that needs to be solved, a further dimension reduction is made that moves the data set from hundreds or thousands of data elements to a few dozens or less. At that point the machine learning system, which can be as complex as a neural network or an algorithm constructed as a decision tree, can be fed with this reduced set of inputs to look for the patterns in the full data set.
“Then you have to train this machine onto a larger set of samples that have been generated with traditional algorithms, so that the machine learns how to solve a specific problem, looking at the outputs computed with the very general mathematical algorithms,” he says. “But the machine does it faster.”
Algorithms developed in this way can be applied to calculating pricing and evaluation, or can be used for in some cases for portfolio organisation or collateral optimisation. The key for the bank is to optimise the calculations.
“Machine learning in this case is replacing the job of other machines, that are less smart in the sense that they are based on stable mathematical relationships,” Morini explains. “With machine learning you can replace these, using dynamic relations that are captured from statistical rather than a mathematical point of view.”
What may seem like a quantum leap in capabilities can be achieved by adding to the existing business. Engaging with a trusted partner can ensure that the right functionality can be teased out from other intertwined functions and clearly defined, a real challenge where parts of manual processes need to be automated.
“It requires engineering and a long-term view; all the usual things that happen in a big organisation that make it difficult to push these initiatives forward,” explains Ocheret.
From a technology point of view, a firm cannot rely on a single cognitive solution; rather, it should take an open and vendor neutral approach. While there are industry standards such as PMML (Predictive Modelling Markup Language), individual firms develop particular technologies and avoiding barriers to development is optimal.
At an architectural level, the more the firm can standardise in its IT environment, the more easily solutions can be deployed, helping the organisation to become more agile.
In the financial industry, AI can be applied to complex risk models, it can be applied to marketing-related models and fraud models. That means using systems that are flexible and allow the conceptual application of AI to be delivered effectively. If systems are too hardwired or too specific, the business will not be able to redeploy them; the technology cannot be deployed rapidly and the firm risks losing opportunity in the market.
“In the early algorithmic trading days, quants would come up with a great trading idea but then it would take six months to code it into the order management system to execute an algorithm in that way,” says Platt. “To overcome that quants would come up with the alpha-generating algorithm, and then plug it into the market via an existing system – Apama – which would be looking at markets in real-time and analysing them.”
“In the early algorithmic trading days, quants would come up with a great trading idea but then it would take six months to code it into the order management system to execute an algorithm in that way.”
Charles Platt, Software AG
The most effective model is to insert tools within the organisation that allow data scientists to deploy AI rapidly based on the data that exists within the business, as the system will need to be trained with that data. The quality of data will affect the quality of output, therefore systems should be trained on data that the firm has used and validated, in order to trust the outputs of the system.
“Just taking a dataset and applying these techniques to it is dangerous, unless you’ve actively been using that dataset and understand its idiosyncrasies,” Field warns.
Ensuring that the initial concept has been reflected in the technology is also key. Having created a model based on historical data and put it into the market, coding errors can occur in translation from statistical or machine-learning experts to the computer scientist that is implementing the model. Thorough testing can help to eliminate errors.
“That’s why you see six to nine to twelve months of a deployment cycle, it goes back and forth,” explains Zeller. “There is overhead of communication and testing and validation, before this model ever gets into production.”
Real-time monitoring of the systems is also important in order to understand how it learns, and ensure it stays within expected boundaries.
Within financial services there are enormous opportunities for driving growth and increasing productivity through the use of artificial intelligence. Much of the potential that firms harbour resides within expanding data sets; AI is a chance to harness that opportunity and make a business more than it has been before.
From enhanced decision making to machine-led data analysis, the front, middle and back office functions can be qualitatively improved, a change that will be reflected in service levels to customers and transformational for a business model.
That change must be made with care, in order to allow the business to grow into new ways of working that outstrip anything that non-AI firms can achieve. Expert help can allow a business to grow this capacity on top of the enterprise, rather than reducing it. In that sense AI will truly boost productivity and growth.