Conventional software succeeds because it’s good at many of the things we’re not so good at: executing instructions over and over, with unwavering energy and discipline, and at tremendous speed.
For 70 years, conventional software has been the brawn to our brains. A faithful companion that performs remarkable feats of computation without challenging our comforting position as the source for the code; the designer of the rules; the seat of intelligence. Well, human, say goodbye to your comfort blanket.
Machine learning doesn’t need our rules. It succeeds by being better than we are at the things we’re good at: the learned, insightful, creative, what-if, fuzzy-shades-of-grey type thinking. Machine learning software learns quickly, creates insight, is nuanced, and never forgets. Practical and affordable machine learning solutions are now overturning our expectations of computing and knowledge work. Once the domain of math geniuses and computer geeks, this technology is now popping up in all kinds of software. Machine learning engines are available via the cloud, with literally thousands of engineers at Amazon busily developing applications (to name just one company in the deep learning space). The technology is fast becoming mission critical for financial services by addressing a whole slew of previously unsolvable problems.
Yet, while almost nobody doubts that machine learning is a superior hammer, it’s not yet clear where the nails are for financial services, nor where to find the problems for this solution. This is my advice on where to aim your hammer.
What is machine learning?
Machine learning is a branch of AI (artificial intelligence), and a keystone technology for much of today’s most disruptive technology, including self-driving cars, natural language processing and genomics.
The term, coined by Arthur Lee Samuel in 1959, refers to a field of study that “gives computers the ability to learn without being explicitly programmed”. Instead of being given a set of instructions and static parameters to follow, the software learns how to operate through its own data analysis. Once taught, AI systems can generalise and derive insights that conventional computer programs can’t.
Machine learning attributes
Machine learning’s great leap forward is that it’s more organic and less rigid than conventional software. This apparent lack of structure makes it more difficult to build, but more flexible in its application. The table below compares some aspects of conventional software and machine learning to help you get a feel for why it’s uniquely placed to solve particular problems.
Table 1, Attributes of conventional software and machine learning.
|Conventional software||Machine learning|
|Operates within rigid parameters||Extrapolates and generalizes|
|Starts with tight assumptions||Starts with no assumptions|
Machine learning in your inbox
Anti-spam filters use a machine learning approach (Bayesian filtering) to let through the emails you want and weed out the ones you don’t. It’s an instructive example of what a successful application of machine learning looks like.
- It learns. Before it can stop junk mail, the spam filter must be trained using a few hundred examples of good and bad emails.
- It generalises. Once trained, the spam filter can guess with a high degree of success if an email it has never seen before is good or bad.
- It adapts. The filter updates its training when you mark emails as junk. By doing this, it adapts to spammers’ changing tactics and learns your individual preferences.
- It’s never 100% right. Without Bayesian filtering, your inbox would be unusable (about 60% of all email traffic is spam), but it cannot eliminate spam completely.
“Siri, why is machine learning suddenly so popular?”
Machine learning is exploding in popularity. In 2016, it reached the summit of Gartner’s Hype Cycle for Emerging Technologies, and according to IDC, we spent $8bn on cognitive and AI systems that year. This year, we’ll spend $12.5bn, and by 2020 this will rise to $46bn.
“This is deep learning’s Cambrian explosion”Frank Chen, Andreessen Horowitz
That investment is happening because machine learning is game-changing technology. According to Netflix, its machine-learning-based personalisation and recommendation system saves “more than $1bn per year”. Some commentators have gone so far as to suggest that AI will be the principal engine for global growth over the next 20 years.
Like all good overnight success stories, machine learning has been around for a long time. What may have changed everything was big data and the lower cost of computing power. Companies such as Google, Apple, Amazon and Facebook need machine learning to make sense of all their data, and they have the vast amounts of cheap computing power to make machine learning practical.
Financial services firms are using it too, and only the unwary will ignore how machine learning can be leveraged to create new, competitive advantage in the industry. Clearly, the use of machine learning in financial services is growing. Successful application of this specialist technology, however, requires an appreciation of the specific class of problems it’s best suited for. Applications suitable for machine learning typically have the following characteristics:
- You want a prediction about something, not a definite answer.
- You can provide a comprehensive set of example data about the problem.
- You have a continuous stream of similar data to your sample set.
- You’re not trying to predict something that will be materially impacted by external data not included in your data stream.
Table 2, Examples of good and bad machine learning problems.
|Good machine learning problems||Bad machine learning problems|
|Predict the likelihood that a customer will default on loan.||Predict profits from the introduction of a completely new and revolutionary product.|
|Use face recognition to determine if a person is who they say they are.||Predict next year’s sales from past data, when an important, new competitor just entered the market.|
Existing, in-the-wild machine learning applications relevant to the financial services sector already include:
- fraud detection
- money laundering detection
- credit scoring
- sentiment analysis
- marketing personalisation
- product recommendations
- natural language processing
- optical character recognition
- biometric authentication
- face recognition
Rise of the bots
The business case for chatbots is clear: it can take a human an hour to respond to a customer query, but a chatbot can respond to an almost limitless number of people within seconds, 24 hours a day, for as long as the customer would like to chat.
Gartner predicts that, by 2020, autonomous software agents will participate in 5% of all economic transactions. Chatbots have the potential to automate almost 50% of services tasks currently performed by humans. Juniper Research forecasts that chatbots will be responsible for cost savings of over $8bn per annum by 2022.
Chatbots are only practical thanks to machine learning. Human languages don’t follow a set of formal rules, so programming conventional computer algorithms to understand human language is all but impossible. Because machine learning is taught rather than programmed, it can learn to understand what we mean, even if the words don’t exactly add up to that meaning. It can adapt to different language patterns, grammar and dialects, even as they change over time.
Virtual agents such as Siri, Alexa and Cortana are marching up the adoption curve, and they’ve arrived in the financial services sector too. RBS, NatWest and SEB Group all deploy bots based on IBM’s Watson technology – AI that handles call center traffic by responding to customer queries. Other examples of AI-powered financial services bots include:
Bank of China, China Construction Bank and China Merchants Bank have all deployed WeChat, a messaging and call platform with 650 million active users.
In March 2017, Amazon and Capital One announced that customers can now pay their bills by talking to a bot running on Alexa, Amazon’s intelligent personal assistant on its Echo device.
Mastercard Labs partnered with Kasisto, makers of “conversational AI platform” Kai, to create a bot for banks that will launch in 2018, providing a way for consumers to shop and transact in Facebook Messenger, and pay using MasterPass.
In the US, Facebook is rolling out a native payment solution that will allow third party merchants to accept PayPal payments in their Facebook Messenger bots. Customers will be able to make payments in Messenger, link their PayPal accounts to their Facebook accounts, and receive receipts via Messenger. As an early pilot of this capability, PayPal’s Braintree partnered with Facebook and Uber in December 2015, to allow users to hail and pay for an Uber ride from Messenger.
Bots aren’t just virtual: India’s City Union Bank is trialling Lakshmi, a robot for handling customer enquiries inside branches. If the robot proves popular, the company plans to install it in as many as 30 branches.
In some fields, machine learning is set to displace human labor. In cybersecurity, it will help to make up for a growing shortfall of skilled labour.
“There are not enough cyber specialists in organisations to deal with the number of threats today, and the imbalance will likely become much worse.”
Machine learning can analyse patterns at a speed that humans can’t match, and detect anomalies that conventional, rules-based software will miss. Compared to humans using deterministic software for cybersecurity, machine learning can:
- Sift reams of log scans and detector data in real-time.
- Provide early warning of potential threats.
- Detect security threats and attacks in real-time.
- Identify zero-day attacks.
- Immediately act to self-protect a system from data loss.
- Use behavioural biometrics to identify users.
- Find authenticated hackers.
AppSensorFS is ieDigital’s’ groundbreaking security ‘nervous system’ that adjusts the Interact application’s security posture in response to events occurring within the digital financial software platform. It deploys detectors within the application to monitor user behaviour and other events, sending an alert or automatically taking action if it identifies a potential security threat.
The first version of AppSensorFS used a deterministic approach to detect potentially hostile user behaviour across 60 detection points. If the type, sequence and frequency of user interactions matched a known pattern, AppSensorFS would log an alert and change the security posture of the platform.
Machine learning is now supplementing this deterministic approach with a non-deterministic capability that can catch atypical or previously unseen hostile behaviour. A joint team of researchers from ieDigital and Queens University Belfast monitored how real users used Interact, and tested the ability of different machine learning techniques (both supervised and unsupervised) to create an accurate representation of normal user behaviour.
Armed with the most successful algorithm from the testing, AppSensorFS has the capability to identify anything that lies outside of that classification of normal user behaviour, assign a level of risk to it, and take action accordingly.
- Detects hostile behaviour that deterministic models don’t or can’t.
- Makes the classification of behaviour as malicious or benign more efficient.
- Can generate insights that can be formalised as deterministic rules.
- Adapts in response to changes in threats.
At first glance, the process of collecting money from people who are in arrears might look like a straightforward, conventional process: get a list of debtors and phone numbers and start calling them at awkward hours. This approach doesn’t work very well in our digital self-service world. It tends to be one approach for all debtors, delivering a poor customer experience.
Research shows that machine learning models are better able to segment delinquent borrowers, and can identify where proactive advice can help keep customers out of debt. A machine learning model developed by the Massachusetts Institute of Technology has shown itself to be “surprisingly accurate in forecasting of credit events three to 12 months in advance”. The model outperforms conventional techniques, and delivers cost savings of between 6-23% of total losses. The researchers also found that machine learning approaches were more adaptive and “are able to pick up the dynamics of changing credit cycles, as well as the absolute levels of default rates”.
“Companies using machine learning have been able to reduce their bad debt provision by 35-40%.”McKinsey
Machine learning seems destined to shake up collection strategies, too. It’s used to select the most appropriate collections strategy for a given customer, enabling a tailored approach to collections that delivers higher performance than the traditional one-size-fits-all approach.
“Most of the Accounts Receivable collection actions nowadays are still manual, generic and expensive … It seldom takes into account customer specifics, neither has any prioritising strategies.”Predicting and Improving Invoice-to-Cash Collection Through Machine Learning
Collections is an emotional business. The stress of falling into arrears tends to accentuate the importance of what, when, and how you speak to customers. Taking the right approach can have a significant effect on the outcome. It’s an environment ripe for a test-driven, personalized approach that maximizes collection performance and customer experience outcomes.
In machine learning, the system selects the best strategy based on continuous correlation of parameter inputs to outcomes. It predicts the right approach for communicating (email, SMS or call), time-of-day, and many other collections approach variables based on the outcomes produced by such methods with similar customers.
Table 3, Comparison of strategy selection methodologies.
|A/B testing||Multivariate testing||Machine learning|
|Easy to set up||Complex to set up||Complex to set up|
|Only tests two variations||Tests multiple variations||Tests multiple variations|
|Cannot adapt||Cannot adapt||Adapts continuously|
|Takes a long time||Requires a large sample size||Can run perpetually|
|Takes a long time|
Machine learning is a sophisticated, general-purpose technology that’s having a real-world impact. Although long awaited, AI is now becoming practical within a whole strata of knowledge work that sits beyond the reach of conventional software. Our knowledge economy stands poised for automation, just as automation has fundamentally changed our industrial economy.
Financial services firms can leverage machine learning to cut costs and deliver superior services to their customers. Those firms that fail to take advantage of this emerging technology will fall behind.
ieDigital is driving forward with a generation of products that can make machine learning an integral part of your business. Our recent developments include detecting cybersecurity threats, smart loan origination, virtual assistants, and optimising collections. Machine learning is now solving some of financial services’ toughest business problems.
Want to know more?
ieDigital can help your organisation deploy a digital self-service platform that gives you the foundations to foster a relationship with your customer and, in time, extend and flex your value proposition.
We can also work with you to better understand your customers’ needs and to prototype, test, and try out new concepts and ideas with them.
For more information, please click here to request a chat or demo.