Mining Customer Trends from Your Data

Selling a car in the classified ads is a frustrating process. You fill out a form, trying your best to pick the words that attract a buyer's attention, send it off to what you hope is an appropriate magazine, and wait. If only the magazine could tell you how and where it would best sell, the whole thing would be a lot less painful.

Auto Trader's recent Cognos-based data mining project, developed in parallel with a national database, aims to do just that. The first stage has provided the salesforce with information on what type of cars were being advertised in its magazines, while the next phase will take it a step further and look at what online users want to buy.

But as manager Peter Comber warned, the temptation was to provide answers to every customer query under the sun. His insistence that information it provided should be linked to sales has provided solid returns for the project, and avoided sending it careering down any paths it should avoid.

"We've had to make sure we've got the basic questions right. There's no limit to the things you can learn - so you need to find out what delivers value to the business."

Data mining is just one element in a development which has radically changed Auto Trader's business and brought its constituent parts closer together. The process began with the development in 1996 of the UK's first picture-based classified advertising site, originally with just two of the company's businesses - luxury cars and motorbikes. The full site followed the next year.

A value-add project, with such offers as space on the site along with a print ad in the magazine, ran in parallel with Web site development, and was facilitated by the need for a national database to bring together information from a number of disparate sources.

Comber says the company was very fragmented, with 13 nearly independent regional editions. "We had 18 different versions of our production systems. Two centres were not even recording historical data."

In setting up the national database, Auto Trader was faced with many obstacles, including a narrow timeframe and a prohibitively large volume of data - 160,000 vehicle advertisements, roughly the number of entries in the Yellow Pages, appear in its pages every week.

A Warehouse In Order

Nearly two years later, Auto Trader had collected and placed all that data into a single Oracle data warehouse and the company, together with partner Simpson Associates, set about the task of analyzing it. Cognos was chosen over other vendors for its ease of use and presentation. "What we found almost immediately was that people who know nothing about databases and cannot understand statistics could see information," says Comber. "The graphical presentation of the information was a massive selling point, because it softens it up and helps you to get to what you want to know much faster."

The poor quality of data supplied puts Auto Trader in a unique situation. "Whatever [customers] write we have to publish," says Comber. "If the customer does not say what the car is, we have to publish that. In addition, they don't have consistent ways of describing them - you could have a list of 15 descriptions for one car."

Cognos' Transformer functionality allowed Auto Trader to get round that, graphically describing the data mapping, so a number of descriptions can be related to a single car. In addition, the graphical interface allows non-technical people to design the data mapping, says Comber. "The people who know about the Ford Escort XR2 are not the same people who can program the database. They can then do the build using the information from the others and very quickly start to resolve the differences."

A Question of Characteristics

Once the database had been built and the tool chosen, the next task was to see what questions it needed to ask. Here there were a bewildering array of possibilities and once again an overwhelming pool of data in a variety of states. But Comber realized there was room to maneuver.

"Our initial problem was that with as many as 40 different characteristics in each ad - the name, model, where, and when it was for sale - we could fish out anything. Some centres had data on price, mileage, colour and so on well structured, while at others the data was just kept in flat files. But we had so much data we had room for error - as long as the errors were consistent, we could just knock them out and we would still have a huge data set to play with."

Comber and his team took a blank sheet of paper to the initial pilot group - agency sales, which deals with the large dealership clients. They wanted to know the type of car, price and region, but also had more vague questions, some of which Comber admits could not be answered. Some, for example, wanted to know how much was new and how much repeat business, but it was not possible to record whether the car was a new sale or a duplicate.

Mainly sales wanted to get a feel for the content of the magazine. "Because it's printed quickly and there is a lot of information, it's on fairly cheap paper, and there's a perception that was not helping them." Although this assumption was incorrect, sales needed concrete information to back it up.

The link between providing information and bringing in revenue is of course largely assumed, but Comber is careful to note this was the whole ethos of the project. "Initially it was important that information had a commercial value [in the sales process]. When the salespeople were back with their customers they could use the information from the national database to secure a sale. If they've got sales information and they can attach it, it will help bring additional revenue in."

All the mining is done within PowerPlay Transformer and Impromptu Reports and up to a dozen sales support staff are the users, rather than business analysts. Comber says it involved minimal training. "Once it's up and running, we have more routine queries. It just points you to the information and suggests further questions."

PowerPlay will also be used for the next phase, still in its early stage, but crucially switching the focus from what people have to sell to what they want to buy.

To this end, Auto Trader started to investigate log-ins from the Web site and the search queries users entered. When you submit a search, all the information you entered on the search form - the make, model, post code - the best way of replicating the regional nature of the magazine in the Web site is recorded, and written into a flat file. The file is imported into a separate Oracle database for analysis.

"The reporting solution is highly successful in saying what they are buying. What is the most sought after car? A Ford Escort, VW Golf or a BMW 3 Series? If you go to a car dealer, they know instinctively which car on their forecourt will sell first. But now we're able to start putting numbers into it, to say that you're twice as likely to sell this car as that one."

Of course, while the analysis tells Auto Trader and its sellers which cars Web surfers want to buy, an entirely different type of buyer might look in the magazine. But Comber says this is not a problem as the motor trade, having struggled to establish itself on the Web, has a universal acceptance that it is the way forward. "We accept that there's not enough emphasis on the different audiences. But the audience on the Web is changing, it's far broader now and more representative. When it started the audience was male, young and affluent, but since then with free ISPs and Web cafes, it has changed."

So what is the most popular car at Auto Trader's Web site? Put it this way, if you're a green Golf owner, you shouldn't have much trouble when it comes time to part with your little motor.