By Steve Wilcockson, Industry Manager – Financial Services, Mathworks.

The financial services industry has long been a hub for technological innovation. So it’s not surprising that banks, asset managers, insurers and even supervisors are exploring how they can apply machine learning.

Interest in the technology has grown rapidly over the last few years.  For example, back in 2014, we surveyed financial professionals and found only 12% were using machine learning in their workflows. When we repeated the study in 2016 with risk and quant analysts attending our London computational finance event, the number using machine learning had almost quadrupled to over 40%.  What’s more, half of those who said they weren’t doing anything today said they planned to start projects within 12 months.

This appetite to embrace machine learning is driven by a strong belief that the future of the industry will be decided by financial computing engineers, and their algorithms. But, stripping away the headlines and hyperbole, what’s the current state of play?

Today, the application of machine learning is progressing at different rates across the industry segments. Buy-side players, competing to increase returns and take (sensible) risks,are increasingly applying machine learning techniques as an extension of their factor analysis suites to differentiate from peers and rivals. One asset manager we know well uses machine learning to determine correlation and predictive trends across macroeconomic, credit, liquidity, risk and money flow factors. This allows them to better understand asset class performance trends, with some of their portfolios outperforming benchmarks by 100 basis points.

The sell-side has a more nuanced “risk management-first” view, oriented around task, skills and, importantly, regulation. Data science and big data teams, sometimes born from Quant departments and other times IT, often focus on “test” problems, such as fraud detection, money laundering investigation, or modelling customer activity. However, methodologies in some cases are impacting mainstream functions, for example as adaptable means of modelling credit risk dynamics. Regulators can be sensitive to the use of such approaches because in some instances machine learning approaches are considered “black box”, but where they are well-documented and discussed, they are increasingly accepted.

The challenges of bringing machine learning into a complex financial business remain daunting. This can be due to the quantity, type and variety of data being analysed. Other challenges might include determining the best model approach for the problem at hand, and ensuring participants are comfortable with its role in the process.

So, as the industry evolves its use of machine learning, what lessons can we learn to generate success?

1. Understand your mandate

Like any complex project, set clear goals about what you want to achieve, and the role of machine learning in achieving them. If considering developing “intelligent” robo-advisers, for example, how intelligent do they need to be, when and how will difficult questions be passed to a human, what human roles will be replaced or augmented, and how will your customers respond to robo-advice? Most importantly, what will the human/automation interface look like? The head of one Wall Street hedge fund pushing ahead with greater use of machine learning summed this up well as “No man is better than a machine. And no machine is better than a man with a machine.”

2. Pick low hanging fruit but be strategic

Running machine learning in a standalone greenfield project can be a sensible first step, but it may be hard to demonstrate sufficient gains and pitfalls. Incorporating machine learning methods into existing processes can be more demonstrable. Consider running shadow projects to compare and contrast performance with existing ones, for example to assess whether retail credit requests such as mortgage extension requests may reflect default conditions or are reasonable requests for (reasonable) asset financing. This allows relative quantifiable benchmarking. Consider multiple measures too. Accuracy is one thing, but model lifecycle might be another. One advantage to both bank and regulator for example in the adoption of so-called bagged decision trees as a means of modelling credit ratings adjustment might result in a 4x to 10x model lifetime compared with a less adaptable traditional approach. This can save development and maintenance time and effort. As one machine learning expert at a major international bank recommends, tackle low hanging fruit to achieve 80 percent of the impact in five percent of the time.

3. Together, Stronger

Machine learning might appear to be a subject for a small band of technologists and thus the preserve of the mathematician and computational elite, but recognize that others bring relevant skills.

Cross functional teams matter hugely. Software engineers bring the ability to speed up model creation and implementation gains; project managers bring project lifecycle management; model users and domain experts bring the ability to state design requirements and suggestions that lead to differentiating and useful insights. To ensure all this activity doesn’t fly apart, communication, coordination and support across the functions matter,keeping the team tight and successful.

Yes, your machine learning expert may be your team’s star player but it takes a strong performance from the whole team to ensure successful execution.  You need to prepare to manage, and overcome failure.  Similarly, it’s important to continue to learn, adapt and improve, both as individuals and as a team. Your team needs to consolidate around its own slick cross functional working environment, developing models and implementing them rapidly.

Your team also must acquire new blood, to develop, apply, implement and interpret new methodologies better, to fill gaps, acquire code and expertise from beyond your organizational boundaries. Kaggle, who run hundreds of data science competitions, can provide competitive leader boards for driving energetic internal problem solving and tapping hidden talents for a machine learning project. It’s quite possible to incorporate these contributions into your own projects.

4. Make it Human

Ultimately user engagement is key. Machine learning is useless without meaningful output. Machines may know best in many cases, but value comes from human interpretation and use. Humans are visual creatures, so consider “simple” visualization as much as you do the most complex model. Avoid misleading graphical representation and do consider subtle nuances that can communicate insights most forcefully. As one big data lead in a global bank stated,“a good font in a visualization is like being more mentally engaged when hearing a nice accent”. Such good aesthetics means key stakeholders can make a faster mental connection with what they’re being shown when sometimes fully and quickly understanding the most complex model output can be challenging.

5. No one size fits all

Every task is different, so use platforms and tools that you can immerse yourself in to understand those differences, to import whatever data-set you need, to try out multiple algorithms quickly and easily perhaps as part of near-at-hand applications. Cross-validate and test, parallelize to speed-up model execution in clouds and on GPUs, and get trustworthy solutions out to the people who need them, as web applications, in spreadsheets and as database components.

But know your risk. Bad tools add risk as much as good tools can facilitate faster time-to-market and enhance productivity. Know which tools you are using, how trustworthy the routines are, and how comprehensive the supporting help and documentation is.  Consider how efficient their algorithms are, and if faced with a difficult time-consuming task, how easily scaled is it in a distributed set-up. Software for developing and implementing machine learning and data science can be like the Texas rangers in the days of the Wild West, at worst toxic, dangerous and unpredictable and at best bring order to the frontier. As far as possible, seek your own order amidst chaos, yet also take care not to lose sight of the opportunity and the excitement of the unknown.

The financial services industry’s use of technology can be contradictory; fast-paced yet also conservative. Nonetheless, the sector is increasingly leveraging machine learning as part of a broad tool-set to address challenging problems, and build exciting new use cases.