Spatial models can be integrated into purchasing decisions for real-time online ad exchanges to improve performance. The country is divided into 52,000 neighborhoods, or Digital Zips, used for geo targeting where ads are bought based on desired neighborhoodattributes. Neighborhoods behave in predictable ways that can be exploited using observed performances.
The idea is to use spatial techniques to predict the optimal manner in which to successfully distribute future ad space. Such techniques can be included in the current model to further explain the residual error and optimize campaign performance.
Figure 1 shows the predicted surface. Notice that there is a single local maximum. Depending on the campaign, ads will be served the residuals between the predicted and observed data. A spatial analysis in Figure 2 shows that the residuals have autocorrelation that is not being utilized.
A second regression is performed, this time incorporating the information from the residuals of the previous
regression. This is the spatial model in Figure 3. This spatial model looks much like a balanced synthesis of the initial regression and its residuals. Figure 4 shows the new residuals of the spatial model. The key is that they are much smaller in magnitude and much closer to being completely random. The next step is to act on this knowledge. Those points that were previously omitted from the campaign are reconsidered. If the spatially predicted success outperforms the current average of the campaign, those points are added to the campaign’s digital zips. This method is performed multiple times, creating a ‘learning algorithm’ that spatially adjusts based on observed performance.
Mitchel Gorecki is a graduating senior who majored in Biomedical Engineering, Mechanical Engineering and Economics with a concentration in finance. He plans on working at MaxPoint, an innovative online advertising company, while deciding whether to later pursue a Phd or Md. The project he is working on uses spatial statistics and a learning algorithm to optimize the distribution of online adverting. Hopefully it will evolve into a tool that enables online advertisers to easily improve their campaigns.