Even in today's hyperconnected world, the act of an individual performing an online search is not completely random. The users have an intent, however vague, which is a very important unknown factor. In this paper we show that the different categories of searches reach their peaks in terms of search volume at different times. Furthermore, we show that the said peaks are made much sharper in relief when the time of search is converted to the user local time, based on the IP address.

Of course, these findings have very important implications for advertisers who participate in Paid Search Advertising. That can be accomplished by different bids for time blocks for different types of queries. We demonstrate the improvement in advertiser RoI by following simple heuristics for campaign management.

Online business and paid search monetization systems are extremely complex due to the interdependence structure of advertisers, users and the platform. Since a deep characterization is not possible to use standard system identification methods, often it is the key performance metrics are what is measured. One such metric is the Cost per Click (CPC). While there are clearly seasonal patterns in the price paid per click by advertisers as a whole, the values are subject to large shocks due to world events and due to sophisticated bots.

While many techniques exist for modeling time series data, executives in online business are required to work at a very fast pace and make quick decisions based on what can be easily understood and explained. This rules out ARIMA models. Furthermore, the model should be able to adjust organically as new data comes in, without necessitating a retrain. It is also not possible to drop multiple outliers in the AR modeling paradigm as it leaves holes in the data.

We demonstrate how AR(1,2,7) model with correction factors for the day of the week performed very well in the out of time sample. However, we found that OLS was swayed too much by large values on, say, Labor day or Thanksgiving day which rendered it ineffective and very sensitive to shocks, which led to poor performance on out of time samples. Robust regression was able to overcome this shortcoming and performed admirably on data two years out from the time the model was built.