|
Model Build Process - Once the data is overlaid
■ A set of models will be created to predict responsiveness on a large sample of this data.
■ The types of models will include regression, statistical networks (advanced neural net-like), decision trees and more advanced algorithms.
■ A hold-out or validation set will be used to compare the models and to select the best model going forward.
■ An entire campaign may be used for validation as well, if possible.
Validation will take place in two phases, with an optional third phase
Validation: 25% of the data that is overlaid is held out and used for an initial validation set. This data will be used to validate the best models performance and ensure that its performance is in line with what is expected.
Back-test campaign: If possible, an entire campaign will also be held out from the model build process and will be scored separately in a 'what if scenario'. Using this data, it will be possible to show what the impact would have been, both in reduced costs as well as in lost revenue from those that were suggested not to mail.
Live test (optional): As the first and ongoing campaigns are mailed, a percentage of the prospects which the model has suggested NOT to be mailed could also be sent out. This will allow for ongoing validation that the prospects being eliminated by the model are producing a negative ROI. | 
