We have previously discussed automated guaranteed - a way to buy advertising inventory in a non-RTB, but programmatic fashion. If you are going to slice and dice your inventory and sell it in a guaranteed fashion, it is essential to forecast your inventory correctly, which gets to the question of what is inventory forecasting, why is important, and how does it work?
Website owners want to maximize the revenue they can get by selling advertising. In the most simple case, a site is only a single type of page, with a single type of advertisement, with one visit per user per day. Here, they'd simply use historical data, modeled against day of week, holidays, adjustments for unusual traffic spikes or dips - and other factors deemed relevant - and simply predict how many impressions they'd likely see. When selling campaigns, websites often have fixed obligations to deliver a certain number of impressions. So they wouldn't sell guaranteed campaigns against exactly what they predicted - there would be some room for error. But because they're selling a fairly simple version of their inventory, it could be pretty close. The rest you'd need to sell as non-guaranteed, which generally goes for less. So the more confidently you can forecast, the more money you can make. If you're at risk of under-delivery as a publisher, you might have to a pay a make-good to the advertiser - or you might have to buy traffic from Facebook or Taboola/Outbrain/etc to hit your targets.
This gets significantly more complicated as you slice and dice your inventory in more different ways. For example, audience segmentation can fetch a premium (e.g. 18-24 males) as well as users that have expressed certain proclivities (e.g. science readers). Further, many sites - like Forbes - have first-impression-of-the-day interstitials, and different ads based on the type of device as well as the type of page (homepage v article page). Finally, many campaigns have frequency caps - such as daily or lifetime. Forecasting among all these different, overlapping groups is a particularly challenging question.
There are a few major vendors in the inventory forecasting space - generally the publisher adservers. This includes DFP, OpenX, and AppNexus/Yieldex. Each is a black box, meaning it won't tell you exactly what its algorithms are. The general idea, however, is that a publisher wants to know if it can sell a particular campaign - given what it has already committed to over a particular period - or whether it's already sold campaigns are on pace to deliver. Behind the scenes, a forecasting tool will run a simulation based on historical data. DFP states explicitly that it uses the traffic data from the prior 28 days to create a model. It also adjusts automatically for standard factors including day and date - meaning that if you run 28 simulations, one for each of the prior 28 days, or one 28 day simulation, you will get nearly the same totals. The simulation will also model factors including user sessions (how many pages various types of visitors visit, how often it's the homepage, etc), device, and relative prioritization of line items to see which would deliver and which would not. Publisher adservers have access to your data, as well as that of all the publishers on their platform - so they can tell how your traffic tends to deviate, and they can also easily identify global or cyclical changes and apply them in their models. Because none say exactly how they do it, it's hard to say for sure.
Ultimately, forecasting is an incredibly hard challenge. The more a publisher divides its inventory, the less reliable any forecasting is. So there will be a tradeoff - which is largely effected by publishers limiting sellable guaranteed inventory at some fraction of forecasted availability. Forecasting may become more accurate, though the fundamental math here is unlikely to be improved upon too significantly. For individual publishers, this is a genuine challenge. Even platforms like TripleLift struggle with this to some degree - our sales planning and account management teams regularly confer to ensure we can deliver on certain proposals for guaranteed delivery, and we will likely build out better tools to analyze our historical data for exactly this purpose.