Flight attendant uniforms attract attention. From a primary association with sex appeal during the 1960-70s, to the diverse role they perform today, the flight attendant’s uniform sits front and centre in the advertising imagery of many airlines. However, relatively little is known about the ways in which consumer behaviour is influenced by airline uniforms.
How can you know if a customer is thinking about defecting from your products or services? Is it possible to predict this attrition, or even stop it in its tracks?
This was the question the marketing team at American Express’ Global Corporate Payments posed when looking to tackle at-risk customers who otherwise looked healthy and normal in its database.
“The customer team didn’t know why these people would churn – we didn’t have any insights, only hunches as to why those customers didn’t work out,” AMEX marketing development manager for Global Corporate Payments, Simon Taranto, told the audience at IBM’s Smarter Analytics conference in Sydney on 10 April.
“The other problem we had is not having enough horsepower to meet the challenge; there was no business case for investing to drive the bottom line.”
AMEX decided the best way was to adopt a predictive analytics approach, and chose IBM’s SPSS predictive analytics modelling software as the smarts for identifying the customers likely to drop off its list. To do this, the software generated a model utilising 150 data variables chosen by the AMEX team including customer charge volumes and values, merchant information, industry data, geographic location and more.
Taranto said these data sources were fed into SPSS to come up with multiple algorithms using combinations of numeric, categorical and cluster information, each of which were tested and cleansed.
The resulting model was run over a four-month period on actual customer data, identifying 24.1 per cent of those who had actually gone on to drop their AMEX account. “We did this without any marketing intervention because we wanted to see what customers actually do, versus us influencing their decision using any knowledge we had gained,” Taranto said.
In comparison with AMEX's traditional approach of picking a general customer sample of 100 to tackle, 100 customers chosen by the model improves identification of attrition risks by 8.4 times, generating a 740 per cent increase in better results, Taranto continued. The model is also able to rank attrition risk by industry segment with a 0.95 correlation to actual data.
AMEX’s attrition model uses refreshes itself by relearning the company’s customer base every 30 days based on updated behavioural data, along with transactional and variable information recorded over the previous 18 months. All of this is necessary to build up an adequate and current knowledge of what makes a customer tick, Taranto said.
It then delivers a list of potential customers at highest risk, which staff can use to tailor suitable communication and follow-up activities, such as direct or electronic mail or calls from the outbound team.
AMEX set up SPSS with a one-desktop licence and defined its payment forecast and payback over a two-year period. While it’s early days, Taranto claims it is on track to deliver forecast ROI sold into management by +/- 10 per cent.
While there are plenty of benefits to predictive analytics, Taranto and his team faced a wealth of challenges in getting its technology-led solution up and running. An initial hurdle was securing investment and getting approval for a technology solution not on AMEX’s approved supplier list. This took months and required substantial collaboration with the IT and finance teams.
Other challenges include disparate data sources, scale, project design, measurement of attrition, identifying the right variables, validating the 30-40 test models, ensuring the project sticks to core business objectives, and balancing the speed of deployment with the ability to execute on information generated.
Another interesting challenge was getting staff to respect information produced by the machine. “What we found was that people were trying to make sense of the machine learning in a human way and put attribution down to a key factor, which is not how these things work,” Taranto said.
The effort was worth it though, Taranto said, adding that AMEX plans to expand the customer attrition program across Asia-Pacific over the next 18 months. It is also investigating new analytics models to further improve and predict customer challenges such as speed of bringing on potential customers, analysing growth opportunities, and cross-selling/upselling existing customers.
Better managing how the marketing and customer teams respond to information from the model is also a focus. Just because the technology to predict is complicated, it doesn’t mean the customer contact plan needs to be.
“We initially found marketing became too overzealous and tried to make the way to deal with those customers identified overly complicated,” Taranto said. “It’s the speed to contact that is far more important than perfecting your solution for dealing with them.
“Simplicity in your solution is my main observation.”
AMEX’s key learnings from deploying predictive analytics:
- Set expectations early
- Validate models against repeated patterns
- Keep fastidious records on prototype model performance
- Engineer your alerts to be incremental
- Simplify model presentations to benefit communicate with non-technology executives
- Articulate the model’s effectiveness in terms of ‘outperformance’
- Follow the validated model; don’t second-guess the machine
- False alerts using the model are an account management conversation opportunity
- Translate the model’s performance into retention or impact by stakeholder
- Retrain the model frequently
- Avoid ‘obviousness’ by identifying and pre-cutting low attrition segments
- Anticipate your engagement lag with customers
- Differentiate between behavioural versus attitude-based predictors of attrition