The crystal ball of web traffic
© corbisrffancy - Fotolia.com
Anticipating a website audience has become essential in terms of selling advertising space at the right price. The new tool developed by Inria, in partnership with the TBS Group, makes it possible to fine-tune traffic forecasts and reduce the margin of error.
What if website audiences could be forecast like the weather? This is the dream for many websites whose biggest income is derived from advertising revenue. Inria joined forces with TBS Group to take up the challenge. This engineering company, specialised in communications, has signed a 18-months contract, finished by now, with Jérémie Mary, a lecturer at the Lille 3 University and researcher within the Sequel project team at the Inria Lille – Nord Europe centre (a joint venture between Centrale Lille and the Lille 3 University*). The aim of this partnership is to supply online companies, particularly the media sector, with a tool designed to automatically anticipate web traffic over several weeks."For a site, this information is essential. Predicting traffic is a crucial element of the decision-making process when it comes to monetising visits to premium sections of a website – a home page, for example, where billing is often carried out on a cost per thousand impressions basis**," explains Jérémie Mary.
For a site predicting traffic is a crucial element of the decision-making process when it comes to monetising visits to premium sections of a websit.
Taking topical events into account
Advertising platforms (Ad Servers), where advertisers and websites meet, have been predicting audiences for many years. However they restrict themselves, for the most part, to reference weeks and do not consider events that may have impacted traffic. While this method has proven fairly reliable in the short term, it has two drawbacks: "On the one hand, it does not consider trends, and on the other hand, it does not take holiday periods or the weather into account. And yet these events have a significant effect on web traffic. When it's raining, we'd rather go online than to take a walk!" he says.
The model which was developed reduces prediction errors by 20% over four to six weeks in comparison to Ad Servers.
Assisted by Inria engineer, Romain Laby, he has developed a prediction model combining time series statistics and wavelet analysis. Not only does it research and automatically identify multiple cycles within the data (new year, holidays, day of the week), it also takes into account one-time events likely to lead to increased traffic (bad weather) or an audience peak (Fukushima disaster, Dominique Strauss-Kahn case, etc.). But that is not all. The tool allows the user to add one-time or recurrent events specific to their business. One example could be the Fashion Week dates for a magazine site aimed at women. As these events always take place on roughly the same dates, the system would be able to anticipate peaks in traffic at these precise times. "This model reduces prediction errors by 20% over four to six weeks in comparison to Ad Servers," adds Jérémie Mary. A precious saving for websites. Frequent poor estimations of traffic can, in fact, lead to a loss of income that may ultimately weaken the company.
*at Joint Research Unit 8146 CNRS-Centrale Lille-Lille1, LAGIS (Control Engineering, Computer Engineering and Signal Laboratory) and Joint Research Unit 8022 CNRS-Lille1-Lille 3-Inria, LIFL.
** The cost per thousand impressions (CPM) corresponds to the cost per 1000 views of pages containing advertising (in the form of banners, columns, etc.). The advertiser pays a certain amount to the site every time a page containing one of their advertisements clocks up 1000 views.