How a bank reduced CPA 37% by optimizing paid media spend using real-time predictive modelling.

The bank integrated three preferred advertising vendors, including two local vendors who cover Facebook and Google DoubleClick Networks, and a third who manages Google Keyword Search and Yahoo Advertising. The Bank adopted their standard of advertisement tracking code for both campaigns and promotions, thus reducing deployment time and mapping/conversion effort in the analysis.

Using Celebrus data collection, one year of historical customer data was used from the website to compile start, visitor, device, page, and form interactions. These data points were used for prediction variables. The data was augmented with loan application data as a prediction target, reconciled with a unique identifier (in this case, mobile phone number which was included in both data sets).

In the final phase of testing, two additional sub-models were built: one for visitors being driven by vendor advertising, and the other for direct website visitors. In-session scoring was implemented using goals tied to score rules, with each goal only counted once per session.

The Celebrus Personalization Connector was used to build triggers that integrated with the advertising vendors’ systems when the score was reached. Trigger rules were configured for low, medium, or high score based on session value metrics. Score-based segment information was passed to advertising platforms using scripts within the actions, enabling the bank to use a different advertising strategy for each segment.

For the low score segment, the bank didn’t spend any budget on retargeting. For the medium score segment, they spent less cost per thousand impressions (CPM) and ran the ads for a shorter duration. For the high score segment, they spent more on CPM and increased deployment duration. The advertising agencies handled the creative elements of the projects in line with the bank’s instructions, derived from the model decisions.