Simulate me: embracing the next evolution of predictive analytics for financial services

Imagine, if you will, anticipating a customer’s default risk years in advance of their approval for a home loan or credit card. What if a health insurer could tailor insurance products in real-time based on the future lifestyle choices of their client? 

The rapid proliferation of artificial intelligence (AI) technologies has made what was once fanciful prophesying into a predictable reality.

Countless financial services institutions (FSIs), particularly in the wealth and asset management space, have implemented AI-backed predictive modelling engines to make informed projections on market trends and capital risk; increasingly, banks and insurers are also turning AI technologies outward, combining rich data assets with ground-breaking algorithms to develop today’s most advanced customer interaction platforms – including the ever ubiquitous ‘chatbot’.                                                

Tapping a vast repository of data – sourced anywhere from ATMs and mobile banking platforms at the front end to systems performance logs at the back – FSIs can now employ an array of statistical, modelling, data mining, and machine learning (ML) techniques to make informed predictions on the future course of the business – a process known collectively as ‘predictive analytics’.

Yet, while predictive modelling engines are forwards by intent, their methods are inherently backwards by design. Predictive analytics engines, by their fundamental makeup, employ a ‘rear view mirror’ approach to situational analysis, using historical datasets to inform decision-making and drive retrospective improvements to both digital infrastructure and processes.
 
Increasingly, however, with break-neck advances in AI and ML technologies, business leaders expect insights from data to not only correct past errors and refashion outmoded infrastructure, but to also prepare organisations for the known unknowns of fast-moving industry trends and evolving customer expectations.

Dr. Richard L. Harmon, Global Industry Leader – Financial Services, Cloudera, believes the inevitable evolution of today’s ‘predictive analytics’ to tomorrow’s ‘prescriptive analytics’ will deliver “a much richer way for machines” to extract information from their big data assets and deliver meaningful and actionable insight for decision makers.

Indeed, prescriptive analytics goes above and beyond the prognostic capabilities of predictive modelling, empowering machines to not only forecast future outcomes, but to also recommend multiple courses of action and present likely outcomes for each course. 

Using next-generation algorithms and computational modelling technologies, these ‘intelligent machines’ could deliver immediate practical benefits for FSIs, from pioneering advances in corporate strategy and business decisioning, to a rigorous, evidence-backed testing methodology for user experience and product design.

“Once you have a set of rules or models that capture that fundamental understanding of their decisioning, you can actually simulate hundreds of thousands of potential future scenarios,” Dr. Harmon says.

“From here, you get a much more robust view of danger points and profitable points for introducing a new product.” 

Beyond predictable limits

Traditional ML and AI engines are, nevertheless, poorly equipped to deliver simulated forward projections demanded of prescriptive analytics models, Harmon believes. 

Indeed, when straying beyond the limits of equation-based historical data analyses – necessary for deciphering complex and highly variable human behaviours – ML and AI programs become increasingly unreliable predictors. 

“While machine learning techniques are good at fitting a non-linear model to historical data, ML cannot model situations it has never seen, i.e. those that go beyond historical data,” he says. “Once the realm of forecast moves beyond historical data, the confidence intervals of those models dramatically widen.”

Enter Agent-based modelling (ABM). Used extensively in the manufacturing and health sciences sectors, as well as by physicists to model once ‘untestable’ quantum theories, ABM is a computational modelling program that simulates interactions between individual agents – be they animals, people, bacteria, or even subatomic particles – within a unique environment. 

Utilised within a business context, ABM allows organisations to effectively “test drive decisions and generate future outcomes that go beyond the limits of historical data,” Dr. Harmon says. 

Behavioural scientists have made ready use of ABM to decipher anything from message propagation and diffusion in social networking platforms (for instance, how and why certain posts can spread ‘virally’), to decoding the complex interplay between environmental and human variables that cause traffic congestion on metropolitan roads.

Because ABMs can be tailored to fit discrete situational contexts, financial services organisations are able to deploy the technology to model situations “at the micro-foundations without fitting those models exclusively to historical data,” Dr. Harmon says. This effectively allows FSIs to simulate arbitrary variables, including customer and staff decision-making processes, and analyse such behaviours at a far more granular level.

“[Using ABM], you’re calibrating models to understand the behaviour of whatever you’re trying to model,” Dr. Harmon says. “If you’re trying to understand the behaviour of customers, you might ask: ‘Why would they default on a loan? Why would they refinance a loan? Why do they take out a new credit card or additional credit cards?’ These are questions you can simulate with ABM.”

Harmon compares ABM-derived simulations in the corporate world to flight simulators in the airline industry, which test pilots’ aptitude in potentially life-threatening scenarios – from inclement weather to engine failures – without exposing them to real danger.

“Being able to do very robust scenario analysis – for example, if you’re trying to introduce a new product or design a new product – and being able to look at all different characteristics to know how it affects different types of customers – how your competitors will respond to it, how the market responds to it, or even the future benefits of cross-selling it – is a very powerful tool.” 

“These are the types of questions banks and insurers are trying to answer today. However, without a simulated environment, that’s very difficult to do.”

As keepers of vast repositories of customer and transactional data, financial services also have a unique advantage in this space. FSIs’ substantial data assets make complex simulations “readily doable”, Dr. Harmon says, not only “transforming the decision-making process, but also offering [FSIs] the ability to innovate quickly, evaluate, and test-drive what [they’re] doing with less risk.”

Yet, while ABM promises much in decoding the mercurial behaviours of customers, Harmon cautions against starry-eyed expectations of the technology – indeed, any technology – as a panacea for the industry.

Such models, he argues, are ultimately “an abstraction of reality”, and it remains incumbent on organisations to defer to human analysis, particularly when simulations trend outside the norm.               

“While it’s usually a very good simplification of reality, it’s not perfect,” Dr. Harmon says. “It is therefore critically important to understand when the models should be used or when they should be used with a high degree of caution.”
 
“In the machine learning space, when you hit areas beyond historical experience, your warning bells should sound. Humans need to be deeply involved, because it’s out of the realm of what the models were trained on.”

“Sometimes the models will do extremely well because the projection is linear it captures something outside the historical experience. But, as we’ve seen in other cases, these models can dramatically fail – it may give you a false sense of confidence that your risk exposures are properly measured or the product is right for the customer.”

For Harmon, machine simulation should – for now – remain a practicable complement to  human strategising, providing “an additional capability that enables a much more robust and prescriptive analytics,” as well as offering “guidance in circumstances that are beyond the realm of historical experience.”

Ultimately, these next-generation prediction technologies empower FSIs to make evidence-backed decisions that take full account of the variable, and often unpredictable, nature of human behaviour – behaviours that are seldom borne out in historical datasets.

“When going beyond the historical data experience, simulation tools can become critically important,” Dr. Harmon says. “It’s less about giving you better predictions of the future, and far more about offering a range of likely outcomes from which better decisions can be made.” 

To learn more about next-generation machine learning and simulation technologies uniquely tailored to your business, contact Cloudera: 
anz@cloudera.com