Accenture helps predict the unpredictable

Listen to this article

00:00
00:00

Online auctions have been one of the internet’s great success stories in recent years, prospering where other businesses have failed as users flock to sites such as eBay to sell unwanted goods and pick up bargains.

In the US, consumer online auction sales are expected to top $65bn by 2010 and account for nearly one-fifth of all internet retail sales, says analyst Forrester Research. Growth is also expected in Europe where eBay already has 34m users and in Asia, where China saw online auctions double in popularity last year.

But while bidders are landing themselves cut-price cars, computers and other consumer goods, sellers run the risk that auctioned items could fall short of achieving their hoped-for price, says Rayid Ghani, a researcher at IT consultancy Accenture.

For the past two years Mr Ghani and a team of scientists at the company’s laboratories in Chicago have been working on a computer program that attempts to remove seller uncertainty by predicting the outcome of an auction.

“We have developed machine-learning techniques that use data from thousands of online auctions to predict various characteristics of an auction, including the end price the item reaches,” he says.

The company is now looking to attract the interest of insurance and online auction companies, who could use the predictive software to offer sellers an insurance policy that guarantees they receive their hoped-for price.

“Rather than derive an optimal strategy for offering insurance, our intent is to give the provider all the information required to price it,” says Mr Ghani.

Using Accenture’s own technology and Clementine data mining tools from software company SPSS, the computer program analyses characteristics of the seller, such as their feedback score, rating and for how long they have sold goods. It then uses this information alongside data about the product being sold, auction details and knowledge from similar sales in the past.

“Our algorithms output a price prediction for each item as well as a confidence score and a company can then use its actuarial expertise to price the risk and determine the insurance premium.” The idea is that the seller pays a fee to an insurer to guarantee an agreed sale price for the auction item. If it should it go for less then the seller would be reimbursed the difference.

“I would be happy to pay an extra few dollars if I knew that my camera would sell for the price I wanted,” he says.

From a two-month trial, which tested the predictive software on Pocket PC PDA sales, Accenture estimates insurers could have made profits of $1.95 per auction item and monthly revenues of $87,000 on this category.

“As information keeps coming in, the machine keeps learning and builds on the existing model. When someone submits a new auction then it predicts the new price based on the extra information,” says Mr Ghani.

But Forrester Research analyst Hellen Omwando, while seeing the potential advantages of the system, questions whether the predictive software will be able to expand to cover categories, such as rare antiques or the auctioning of services, including gardening or house cleaning.

“It’s an interesting move to provide a safety net for small to medium sized businesses and individual sellers,” she says. “It seems like a scenario where everyone wins, but very rarely is that the case with insurance. How can the algorithm account for antiques, future services or emerging technologies where there is little data?”

Accenture agrees that the software is currently inappropriate when it comes to predicting prices for rare goods, such as art or antiques, where it is impossible to analyse historic sales data.

Mr Ghani says abnormal changes to supply and demand can also have the occasional strange effect on a prediction, but this is swiftly corrected by the software. “We saw during our testing of PDAs that prices suddenly plummeted. At first, we didn’t know why. Only later, did we learn that Dell had a promotion where they gave a free PDA away with every server purchased, resulting in extra PDAs flooding the market on eBay,” he says.

It is also debatable whether a seller would buy online auction price insurance instead of setting a paid-for minimum bid price for the auction or a secret reserve, where bidders do not know the value of the floor, says Ms Omwando.

But Mr Ghani points to research by Rama Katkar from Stanford University’s Graduate School of Business, which says that secret reserve prices can deter bidders from participating in an auction and reduces the likelihood of a sale. He says findings from the research show that only 46 per cent of secret reserve auctions resulted in a sale, compared with 70 per cent of public minimum bids.

“But when you have an auction that is being insured, the bidder doesn’t know anything about it and, as it’s not like a minimum reserve price, it doesn’t have an effect on the bid,” says Mr Ghani.

Irrespective of whether insurance companies seize on the data mining tool, Mr Ghani says it has many other applications.

“A different application would be to use it to work out how best you can develop an auction. By looking at data on previous successful auctions you could ask: ‘How do I get the most money?’ and discover the right times and ways to sell.”

Copyright The Financial Times Limited 2017. All rights reserved. You may share using our article tools. Please don't copy articles from FT.com and redistribute by email or post to the web.