Banks Deploy Artificial Intelligence to Deepen Understanding of Customers

Register now

As BBVA Compass was tinkering with its retail banking pricing strategies last fall, robots were scraping the web behind the scenes.

Like its peers, the Birmingham, Ala. bank needed to adapt to caps on interchange fees. It had to decide which benefits its retail customers would keep and which would get the ax.

On a hunch, the bank decided to drop its $25 checking anniversary bonus, while keeping its debit rewards and other features.

The computers confirmed that people were upset, but prodded executives to tell customer service reps and tellers to explain that BBVA Compass had decided to keep its other perks.

Those robots, really snippets of code dipping and diving into paragraphs and sentences across the internet, gave the bank what focus groups and pollsters would have taken months to confirm.

The reports the bank receives from the machines give BBVA daily insights into consumer reaction to the bank and its competitors.

"We look at our own [bank]. We look at our major competitors," says John Wessman, BBVA Compass' executive vice president and chief marketing officer. "You tend to get a little bit different perspective from consumers. They will be much more open and sharing on their Facebook accounts than they would be if they called a call center."

The technology BBVA used is an outgrowth of sentiment analysis — the same type of study researchers conduct when polling an audience. A computer program, instead of asking people questions, surveys the Web on a large scale.

Technology giants such as IBM and SAS offer clients software packages that comb the Web for clues to consumer sentiment. Those programs parse language and attempt to understand the meaning of words, pairs of words or phrases, in context.

This has become ever more important as banks move their customers out of branches and toward the Web. Those customers are less likely to walk into a brick-and-mortar building or chat directly with a teller.

And regulators such as the newly created Consumer Financial Protection Bureau have already said they plan on crowdsourcing complaints. That means a bank needs to monitor its presence on the Web even more closely.

Still, the technology is more of an ongoing study than a science.

Many vendors, and even more computer scientists and academics, are working to perfect methods of gauging feelings and emotions across the Internet without hiring a pollster.

The practice has only recently filtered into the financial services industry, and the largest players in the past several years have just grasped the benefits it could have for business.

In the end, the technology could be as valuable to banks as face-to-face interactions.

The Study of Sentiment

Linguists and programmers have been developing their brand of sentiment analysis since the beginning of the last decade.

The science incorporates natural language processing, which examines sentences and paragraphs, mimicking the way a researcher might speak to a member of a focus group. It uses filters similar to, but more sophisticated than, programs that spot spam emails.

It's artificial intelligence. Not the kind you've seen in sci-fi movies. It's more about machines learning from data and experience sifting through large quantities of data to inform business management and problem-solving.

The simplest method of sentiment analysis measures good words against bad.

A more complex model has a computer working to figure out phrases and groups of words associated with positive or negative feelings in sentences that researchers choose.

"If I have 100 tweets, and 50 of them are positive and 50 of them are negative, what is the difference?" says Eugene Wu, a database graduate student in MIT's Computer Science and Artificial Intelligence Laboratory. "The computer will distinguish them and look for a word or pair of tweets that are more positive. [The computer] can start saying these words are closely associated with positive feelings."

Methods computer scientists use to gauge sentiment differ according to the geographies they're polling, as well as the cultures and businesses those algorithms are trying to cover.

The technology will be someday be able to tell with certainty whether a person is a buyer or seller, for instance.

Researchers also try to figure out the subjects of individual conversations that happen on Twitter or in email messages.

It's called topic modeling.

"The question is not if people are tweeting positively or negatively, but can we discover automatically the nature of the underlying topic that is under discussion," says Philip Resnik, a professor of linguistics at the University of Maryland, who also works at the university's Institute for Advanced Computer Studies and is lead scientist at Converseon, a social media agency.

One early application for sentiment analysis was academics trying to gauge the positions of movie reviews, says Apoorv Agarwal, a fourth-year doctoral student at Columbia University. Agarwal last year co-wrote an academic paper on sentiment analysis of Twitter data.

The reviews, because of their rating systems, provide a guide for computers to measure against.

Academics later set out to analyze different views taken on a subject in longer online articles. Further analysis would identify the different moods found within a story.

Related software tools are being used to target influencers and determine whose opinion is given more weight on the web.

Financial services is just one industry interested in the technology. Clandestine federal agencies — meaning yes, the CIA — and news organizations are, too.

For instance, The Wall Street Journal uses sentiment analysis to help its editors create its Sentiment Tracker feature for its Weekend Review section, picking out tweets and Facebook updates about popular topics.

An elementary first step for many banks is the basic sentiment analysis embedded in social media monitoring tools.

"Monitoring is just how you've been mentioned and how many times. … Sentiment analysis is going to go beyond that," says Seth Grimes, a Washington-based consultant who founded the biannual Sentiment Analysis Symposium, which attracts business users and technologists from around the country. (

Some banks are using social media monitoring to find employees who might be giving away a little too much information over social networks. That could potentially reel in a bank executive unknowingly committing an SEC violation, says David Wallace, SAS' global financial services marketing manager. SAS offers analytics services that apply linguistic rules and statistical methods that tease out insight from unstructured text, such as survey data, email, call center logs and loan applications as well as social media streams.

Big-Bank Projects
A majority of the bigger banks already employ sentiment analysis technology, says Boxley Llewellyn, IBM's retail banking director. IBM provides sentiment analysis services to BBVA Compass' Spanish parent company. "What's new is we used to look at our enterprise data, now we look at the enterprise data, the stream data and the social data, and all the banks are going to look at all three," Llewellyn says. "That's when the eyes light up when we talk to banking and business executives."

BBVA Compass is using technology from NM Incite, a joint venture of Nielsen and McKinsey & Co.

The software picks out posts where the author expresses positive, negative and neutral feelings and gives BBVA the ability to react.

BuzzMetrics, the NM Incite tool that BBVA Compass uses, can show how social media users comment on the bank.

The software has sped up BBVA Compass' reaction to other trends, as well. Over the past two months, the bank has started to consider raising the cash back rewards on its credit cards because one of its larger rivals was receiving positive sentiment on its benefits.

Citigroup has engineered a content platform called CitiVelocity that boasts a credit sentiment monitor. It works off of Thomson Reuters' Machine Readable News service, which provides news stories in computer-readable format.

Citi executives say the tool is purely used for researching companies on the credit-default swap index in the U.S. and Europe.

CitiVelocity has been live and free to use for Citi's institutional clients since November.

The monitor looks like a hot-or-not rating. The most popular companies appear at the top of one row. The most unpopular fall in another.

A client can click on any company name and see a chart of how positively or negatively that company is viewed over a period of time.

A company receives a score of between minus-1.00 and 1.00 — 1 being completely positive, minus-1 being completely negative.

"The interesting thing about CitiVelocity is it brings attention to the names and the companies that are being discussed," says Ron Papka, Citi's global head of client analytics and market data distribution. "The thing that I find most incredible is that you can take a stream of language and a stream of words and it gives you a metric."

Meanwhile, American Express has been using social media monitoring since 2009, prodded by the 'financial tsunami' going on at the time, says Christopher J. Frank, Amex's vice president of its global market insights group.

"You look at consumer confidence intervals. You look at the overall mood. And we wanted to make sure that we really understood that, and not just from the American Express point of view, but what is on the minds of our customers," says Frank.

Today that technology, which is provided by Visible Technologies, has crept into everything Amex does. It informs the company's marketing strategy, and gives it direction on what rewards it presents to its customers, among other business practices.

"When you look at a series of the programs that we have released, like Link, Like, Love; Serve and Go Social, we are now taking this information, and not only looking to take the pulse, but being proactive," says Frank. "Link, Like, Love" is Amex's Facebook app that presents its users deals and offers based on their likes, interests and social connections. Serve is Amex's digital wallet. Amex's Go Social program helps merchants create social and mobile offers and receive detailed reporting on those offers.

Some large banks consider their work to be trade secrets. Representatives from TD Bank, Wells Fargo and Bank of America declined requests for interviews on this topic. "How we track and analyze sentiment is still evolving and to some degree a competitive advantage that we don't necessarily want to make public," Wells Fargo spokesman Matt Wadley said by email.

Where Sentiment Analysis Falls Short
For all the work that's been done with sentiment, some practitioners believe it's too soon to take people out of the equation.

"To get the most out of social listening tools, I personally believe the best sentiment analysis requires human review on every post," says Frank Eliason, Citi's senior vice president of social media. "Often the numbers are minuscule because the tool cannot accurately determine" an author's feelings.

Typically, Eliason says, you see barely-negative-barely-positive readings with the remainder of the sentiment of the message unknown or neutral.

"But as you dig into that you will see numbers drift dramatically when you use human intelligence," he says. "One company once offered sentiment analysis of banking, with certain banks clear winners. As I dug into it, the tool included topics not even related to the banks, very much skewing the results."

This has become even more important since banks got a big dose of bad vibes during and after the Occupy protests in cities across the country. The movement reportedly led people to switch from bigger banks to smaller community banks and credit unions.

Overall, sentiment analysis technology is still in early stages.

"The devil is in the details, every specific problem has different properties, so I don't think we are going to find a generic solution," says Resnik of the University of Maryland. "But what is happening right now is we are starting to build metaphorical tool kits that allow us to get up to our elbows in data, and figure out what the problems are and apply answers."

For reprint and licensing requests for this article, click here.