THE ENDLESS OPPORTUNITY OF
INNOVATION AND MACHINE LEARNING
INTRODUCTION
Rob Alexander joined Capital One in 1998, three years after
the bank was founded. He served in a number of leadership
roles in the company’s credit card and banking businesses. In
2007, after directing the effort to overhaul the systems and
infrastructure underlying Capital One’s credit card business, he
became chief information officer.
WHAT ROLE IS MACHINE LEARNING
PLAYING IN YOUR BUSINESS MODEL?
Central to our AI strategy — and what motivates us — is the
recognition that technology is driving a revolution in just about
every industry around us. You don't have to look far to see evidence
of that. The winners in banking are going to be organizations that
are really great at technology — building software, creating digital
experiences, leveraging data and analytics, making machine
learning and AI deliver a great customer experience.
HOW DOES SUCH A BIG SHIFT TO
AUTOMATION START?
It starts with some of the obvious things you hear folks talking
about when it comes to machine learning — fraud detection, credit
risk and cybersecurity. You want to be able to approve as many
transactions as you can by making sure you only identify fraud
when it's really very likely to happen. You want to get better at
distinguishing credit risk and lending the right amount of money to
the right people at the right time. Cybersecurity is a place where
there is an enormous opportunity to apply machine learning and
AI. The threats are constantly evolving and very sophisticated. AI
increases your ability to protect customer data and to enhance the
security-related experiences of your customers.
Many of those decisions are being made by opaque, black-box
neural networks that don’t explain how an output was reached. Is
managing this hidden quality a factor in your approach to AI?
We're focused on building capabilities around what we call
explainable AI. We think it’s important to have models that aren’t
just black-box models but ones that enable us to understand why
deep learning and neural net models are making the decisions
ROB ALEXANDER
they’re making. We're also very conscious of the ethical issues
CHIEF INFORMATION OFFICER, associated with the application of machine learning. Financial
services can help people achieve their dreams, but when it’s done
CAPITAL ONE
poorly, an institution can get in the way of someone’s dreams. We
take that responsibility very seriously.
HOW IS MACHINE LEARNING HELPING YOU
DEEPEN THE CUSTOMER EXPERIENCE?
We’re able to see things that look unusual and send a customer
a notification in the moment. For example, we built a first-of-
its-kind spending tool called Second Look, which uses machine
learning to alert customers about unusual spending patterns
— double charges, a repeating charge that is higher than last
month, or a tip higher than a customer’s norm. Second Look has
saved customers millions of dollars in unwanted charges. This
is AI that tells our customers we’re looking out for them, we’ve
got their back. We believe banking should be more integrated
into customers’ lives and meet them in a way that is accurate,
relevant and timely. Our goal is to bring greater humanity and
simplicity to banking. Toward that end, we launched Eno and
our Alexa skill on Amazon Echo — our text-based intelligent
assistant and our voice experience, respectively.
WHAT ROLE DOES THE CLOUD PLAY IN
INNOVATION AT THE COMPANY?
As early as 2012 we had developers who were using the cloud and
evangelizing its use. In 2015, we declared ourselves a cloud-first
organization, meaning all new development will be built on the
cloud and our legacy applications will migrate to the cloud. The
move has been resoundingly embraced by our organization; it’s the
key to our ability to move quickly and attract great talent — and it
has differentiated us from what other financial services institutions
are doing. We control our environment better; we move with
speed and agility and flexibility. As the cloud infrastructure evolves,
we can evolve with it, and we think that’s really powerful.
AGILE IS A KEY INGREDIENT.
HOW DID YOU BECOME AGILE?
We realized in 2010 that we needed a technology delivery model
that was suited to how the world and consumers were changing
— a faster cycle time on developing new products and capabilities
that fit with the digital age. We accomplished that by having
people around the table working together, physically present,
with the business customer embedded in that process. You get
very granular connections of technology to the business when
you have an agile team that might consist of several engineers, a
scrum master and a product owner who is a businessperson. You
can see the power of iterating rapidly, through two-week sprints
with deliverables, and human-centered design incorporated into
the process. All of this leads to better, faster customer innovation.
SPEAKING OF TALENT, WHAT DREW
YOU TO CAPITAL ONE?
I was drawn to the people and the culture, with its core strategy
around data and analytics — the scientific method of testing
and learning as a way to build a better business model. After
school, I spent four years in the Air Force, where I worked on the
development of missile systems. One project was upgrading our
command-and-control platforms for our nation’s ICBM force. You
had to understand the requirements and how to upgrade what
was 1960s technology. Then I went back to business school and
worked as a consultant. But I really wanted to join an organization
where I could be part of a team and have an impact.
WHAT IS THE KEY ELEMENT FOR A BUSINESS
AIMING TO MOVE FIRMLY INTO THE DIGITAL
AND MACHINE LEARNING AGE?
There’s an almost endless opportunity to leverage more data and
to make more-informed decisions, with better context, and give
each customer a differentiated experience. In many cases, you can
even rethink and redesign the experience based on intelligence
in ways that were not anticipated. The biggest challenge for any
institution trying to drive digital transformation is getting to the
point where you have a critical mass of digital and engineering
talent that allows you to attract more of that talent. To do that you
need to think about your technology operating model — the way
you build technology that’s inclusive of the platforms you build,
the technologies you use, the processes involved, the speed at
which you work, the productivity tools you allow associates to
use. All of these things wrap around creating an attractive place
where talented people want to work. You also need to think about
your underlying data ecosystem. You can hire data scientists, and
they can come up with a cool model, but unless you have a whole
data ecosystem upon which you can work, you can't really take
advantage of the talent and technology. We are very much about
building that data foundation that will support that leap into the
kind of next era of machine learning and AI.
HEDGING RISK WITH MACHINE LEARNING
INTRODUCTION
Peter M. Phillips is the president and CEO of the PathWise
Solutions Group at Aon Benfield, the reinsurance and capital
advisory division of U.K.-based Aon Corporation. His 25 years
of experience covers capital markets and, since 2001, variable
annuity hedging. He led the creation of the PathWise platform,
a computing solution for the insurance industry that accelerates
simulation-based risk management.
WHAT ARE THE CHALLENGES
FACING THE INSURANCE INDUSTRY,
SPECIFICALLY IN THE ANNUITIES MARKET?
Variable annuities are among the most complex securities or
products in the marketplace today. It’s just one challenge after
another; under every single rock, there’s a nasty surprise. You might
have monthly time steps for 30 years — that’s 360 cash flows to
model just for one scenario. You can have up to 10 different stock
indices to model. You have interest rates, with a whole curve to
model, and volatility for rates and equities in the mix. You’re dealing
with a really tough problem. At the center of managing variable
annuities lies data. You have huge data sets. And some insurers
have 3 million to 4 million individual policyholders. You’re talking
about enormous data warehousing and computing challenges if
you have to do calculations on an individual policyholder level and
store every cash flow inside a simulation.
HOW DOES AN INSURER MANAGE
SUCH COMPLEXITY — AND WHERE DO
TECHNOLOGIES SUCH AS MACHINE
LEARNING FIT INTO THE PICTURE?
The answer is, historically, companies have struggled. The industry
itself is undergoing change — and always will be. But at the heart
of it lies technology. With annuities you have complex derivative
mathematics and payoffs. You have to know contingent claim theory
[where the payout is determined by the occurrence of uncertain future
events]. You have to know statistics. You have to know computer
science. Machine learning is going to really help the industry in claims
management and fraud detection, the table stakes for insurance
PETER M. PHILLIPS companies, and it will help specifically with variable annuities. I see
it extending into underwriting and client support areas. Machine
PRESIDENT AND CEO, learning is also well-positioned to enhance the client experience and
AON BENFIELD to help reduce support and overall infrastructure costs.
MUCH IS BEING SAID ABOUT THE ABILITY
OF NEURAL NETWORKS TO MAKE ACCURATE
PREDICTIONS AND HAVE A BIG IMPACT IN THE
INSURANCE INDUSTRY.
At PathWise, we have this whole ecosystem that’s neural-
network ready, or machine ready. Neural networks have a key
role to play in several areas. One is predicting policyholder
behavior. Retirement products have embedded financial
guarantees in them; predicting how someone will behave has
a key impact on the value of these securities, just like modeling
prepayments does with residential mortgages. Big data and
machine learning married together can really provide insight
into what’s happening at the policyholder level. It’s like there’s a
cave and we haven’t even put a flashlight in it.
IS THIS WHERE THE COMPUTING
POWER OF THE CLOUD COMES IN?
We do not do 100 simulations or 1,000 or 100,000 — or a
million. We do trillions of simulations. If I want 100 GPUs
[graphics processing units] for 48 hours, I can get it. If you
had to do this inside a company, it could take six months for
the equipment to show up, then another three months to put
it all together. You would need an army of people watching it.
Stuff is always breaking, you’re patching every single day. You
can die from patching. To stand up that type of infrastructure
and maintain it properly is a huge expense and very difficult to
do properly. The dirty secret of our business is that insurance
companies are spending tens of millions of dollars on IT —
and they’re not getting a lot of value from it.
ARE YOU’RE SEEING VALUE IN
TIME AND COST SAVINGS?
I can say to a client, “You want GPUs, let’s go get them.”
We’re talking about minutes instead of months. Adding more
storage is easy. The advent of GPUs and the cloud has really
helped our business and our clients grow. You can figure the
cost to run just one GPU, and there are huge savings because
the hardware is purpose-built for running Monte Carlo
simulations in parallel, versus serially with CPUs. You could
look at situations where, in terms of dollar performance,
we’re just way ahead of our competitors.
WHAT HAVE YOU DONE TO STAY
AHEAD OF THOSE COMPETITORS?
Back in 2008, we were innovating using the Sony PlayStation chip
because, if you programmed it, you could do a lot of calculations
very quickly. We were successful, but unfortunately, they stopped
production of the chip. That’s when we started programming on
GPUs — and we’ve not stopped since that day. We picked up
important lessons on the way. We don’t want to bet the farm on
someone producing specialized chips for some other purpose,
only to have them stop production. We also inventoried all our
frustrations with existing systems that were out there. One was
black-box models. You can’t see every step of every calculation.
You’ll see parts of the calculation, but you won’t see every step
and order of operation. I was determined to put an end to that
by building a paradigm where you can see everything. In variable
annuities, you have billions of dollars of securities put to work every
day in hedging programs, so it’s super important to have trust and
faith in the numbers that are coming out of your system.
HOW DO YOU GAIN TRUST IN THE NUMBERS?
One way to get there is to be able to follow all the data and
calculations from beginning to end. That was one problem we had
to deal with. Another was the systems. In the old days you needed
five different systems and someone running around with a clipboard
to manage this complex business and data handoffs. Systems were
really strong in modeling or projecting insurance products but awful
at modeling things like swaps or assets. Some systems just modeled
the insurance products but couldn’t do scenario generation. I wanted
a system that could do all of those things together, where you could
put assets in it, liabilities in it, scenario generators in it, and see your
risk at time zero and through time because you had all these things
working together in one system. Data is more than just controls; it
has to be optimized for storage and computing.
WHAT DOES THE FUTURE
HOLD FOR MACHINE LEARNING?
People don’t realize that we’re in the age of accelerated computing; I
don’t think they realize that machine learning is the future. It’s like the
industrial revolution. The show’s just beginning. Whole industries will
be disrupted. Our society, the fabric of it, will change. And AI is here,
right now. It’s going to present a sea change for insurers in how they
conduct their business, how they service their business, how they
launch products and manage risk. It will all change in a profound way.