Will enable us to use all of Caret's functionality for ensembling, model comparison, etc.

The data set contains a factor variable (a-i) that indicates the value of the insured car WHEN NEW. Combining this class with the car age should allow us to classify the depreciated car value. This is a little tricky, for example, classic cars may not depreciate at all after a certain time period.

The naive model is relatively effective for the competition. Even the best competitors are beating it by only a few points. This indicates that in most cases customers purchase plans for which they have previously (and most recently) received a quote. For these customers, the naive model might be the best model.

Alternatively, there are other customers who will choose to purchase an unquoted plan. For these customers an alternative shopping or customer-based model should be used to improve accuracy.

This could be accomplished by ensembling the naive model with another model and have the ensemble model decide when each model should be applied.

Reimplement the naive model as a custom Caret model. More information at http://caret.r-forge.r-project.org/custom_models.html.

The naive model (competition benchmark) is relatively effective. We could take advantage of this by perhaps ensembling the naive model with other shopping and customer models. In effect, we would be deciding for which customers we should simply use the naive model and for which customers a more complex model is more effective.

Caret already provides good infrastructure for comparing models, etc. By reimplementing the naive model as a custom caret model it should make ensembling easier.

Can user-based collaborative filtering help predict a customer's purchase?

Item -> Plan Options
User -> Customer
User Attributes -> State, Location, Car Age, etc, etc
Ratings -> The number of times a user shopped for a particular option. A user's actual purchase might be weighted to indicate a much higher rating.