Datamin deficiency

A look at what works and what doesn't in data mining

B S Nagesh (left), MD, Shoppers Stop &
Joseph Sedjwick, Head, BI, StanChart

Data-mining is a bit like prospecting for gold. If you hit pay-dirt, the rewards can be glittering. But more often than not, the exercise involves chipping away at the mine of information at a company's disposal for small, but useful gains.

The tool, companies are learning ever since it gained currency in the mid-1990s, can rarely be the touchstone for dazzling corporate strategy.

"Not all business issues can be solved using databases. Data mining, however, can definitely hold the key to specific business problems," says Ramnath Krishnamoorthy, director, modelling and analytics, ACNielsen ORG-MARG.

So how do companies extract the maximum from their efforts? Here are some experiences.

In 2001, Standard Chartered Bank was looking to convert more customers to ATM transactions. But it had only 70 ATMs (competition had 200 to 250 each). That meant StanChart customers usually had to use other banks' facilities, which involved a high transaction fee.

To find a solution, StanChart scrutinised its customer information database. A reading of usage patterns showed that the average customer conducted four ATM transactions a month. But the data also showed that few of the bank's customers actually used an ATM simply because of the high fees involved in accessing another banks' ATMs.

Working from the results of its data-mining exercise, StanChart launched AccessPlus, a savings account with a free debit card, which offered four free transactions at other ATMs. The initiative saw a 30 to 40 per cent increase in transactions at non-StanChart ATMs. The bank also claims that its incremental business jumped 25 to 30 per cent with new customers signing in for AccessPlus accounts.

StanChart's early experience changed the bank's approach to data mining. In the last financial year, the bank invested about $150,000 in its business intelligence function. From just five employees before 2000, it now has a full-fledged business intelligence unit with 20 professionals.

Yet, it is not as though data mining has become the lynchpin of decision-making at the bank. As Joseph Sedjwick, head, business intelligence, Standard Chartered Bank, says, "It would be unrealistic to assign macro-level objectives such as increase in market shares or sales to data mining."

In StanChart's case, data mining has been used to reduce costs of campaigns, increase response rate and bring about finer customer segmentation. The bank now claims a 15 to 20 per cent reduction in the cost of managing campaigns and a 25 to 30 per cent increase in response rates.

How does this happen? That's because the inferences and conclusions drawn from databases depend on limited customer information, which is then segmented by various criteria (income levels, liking, propensity-to-buy and so on). The company then puts this information to use "" the most common use being focused marketing promotions/campaigns targeted at select segments.

But the response rate of campaigns devised via data mining depends on variables that hard numbers can't predict accurately. And that's one of the major hitches with data combing, as retail chain Shoppers' Stop discovered when it launched a campaign in the festival season in 2003 (November to January).

The departmental store selected high-ticket shoppers (those who had previously shopped for over Rs 8,000 to Rs 10,000) from its database and sent direct mailers offering them a Mont Blanc pen worth Rs 8,000 free with purchases amounting to Rs 25,000.

"Since the festive season was on, we expected a good response, so we had ordered Mont Blanc pens accordingly," says B S Nagesh, managing director, Shoppers' Stop. As it turned out, the response was 25 per cent higher than what the store expected.

Why did this happen in spite of the structured database at the company's disposal? "Data mining helps you know what happened "" transactional information on what customers actually bought, how many customers shopped and so on. But it does not tell you what hasn't happened "" why the customers who didn't shop on a given day behaved that way," Nagesh points out.

The response rate could also be a function of the quality of data acquired. For instance, customer information sourced by the company may be inadequate.

This is a learning that financial services major HDFC has drawn. HDFC has separate databases for its home finance and retail products. Information from home loan and credit card customers is sourced from the application forms.

"Going by experience, we don't rely completely on the credit card database," explains an HDFC bank executive. This is because, largely, customers tend to give incomplete information when they apply for credit cards. Whereas in case of home loans, they are bound to give complete information to get loans sanctioned.

HDFC is trying to solve this problem by inducing its agents to demand filling up all details in forms. The lesson: the data source points determine the quality of data. The surer you are about the data you have, the better the decision-making process.

Also, data mining makes little sense if the database is lower than 1 lakh. The larger the size of data, the more the chances of finding the customer segments to target. "Data mining helps you find the top one or two deciles to target," says Ajay Kelkar, vice-president, head, marketing HDFC.

Once that's done, there's hardly any room for intuition. "Unlike in market research, where you search for potential customers, it's difficult to disregard customer information that you already have," points out Sedjwick of StanChart.

The larger question, however, is: how justified are spends on in-house data mining when companies already invest in market research? For one, bigger databases need to be managed with sophisticated software that costs Rs 20 lakh to Rs 2 crore. "More than the software employed, it's important to have people with both technical and analytical skill sets to run the operation," says Nagesh.

The payback period for investments made on data mining ranges from one to three years. And the returns on investment may not show in terms of overall market shares or turnover but in a stronger ability to make better business decisions and solve business problems.

But consultants point out that this can't happen if companies churn data without a clear sense of purpose. "You can extract more from data mining by working backwards from an identified problem," Arjun Erry, director, sales, of software consultancy SAS India says.

Another problem area, point out consultants, is that companies take time to implement marketing decisions only after a few months of data validation, say consultants.

Six months ago, Shoppers' streamlined its business intelligence function. Recently, it also shopped for high-end data mining tools and has installed Business Objects, a data mining software that, Nagesh says, " enables data mining in less than half the time."

But a challenge remains: to build an accurate forecasting model. However, with stock keeping units running into thousands and footfalls numbering 12 million per year, there are too many variables to reckon with. So Nagesh prefers to be realistic about the kind of objectives he expects data mining to achieve for Shoppers' Stop.

"Largely, I expect data mining to help us understand what can increase footfalls. Then, it should curtail marketing waste by helping us spot the likely customers to target for new collections or selective promotions," he says.

Till such time, questions like whether the lady who bought crystal ware twice this year will buy the new stock of crystal jewellery next, is anyone's guess.