Returns on a bank selling campaign based on scoring list
Gianluca Bo, Flavio Bonifacio, Metis sas

Introduction
What is the real impact of a marketing campaign? Is it really useful applying statistical methods to guide selling activities? The case history presented here shows an example of a successful combination of data mining techniques and company goals. Two years ago in Dublin we presented a paper entitled "An ordered list of customers for a new Bank: scoring with EM3.0". We are now able to present the returns of the campaign based on that scoring list. The first section briefly illustrates the project, the methodology used and the results obtained. The second part analyses the impact of the scoring list on "sales".

The Project
Two years ago we were asked to pick out the best potential clients for a new Bank from an insurance company´s list of clients.
Our first goal was to identify the way a good insurance client could become a good client for a bank.
In order to answer this query we adopted two different methods. We qualified the two procedures as the deterministic approach and the probabilistic approach.

A deterministic approach
Let´s postulate that the best clients for the bank are insurance clients inclined to acquire saving products instead of, or over and above, the traditional insurance products: those that have lower accident rates, those that pay high premiums and those having no delays in payment terms.
These are the best clients for an insurance company and we assumed that they were also good potential clients for the bank, which means that this constituted our target population.
We referred to this process of client segmentation as the deterministic approach because the criteria are totally a priori defined.

A probabilistic approach
Our first step was to exctract the target population, however it was reasonable to think that people who had not yet bought a life insurance product, but who were somehow similar to those who had already bought one, could also be optimal for the bank.
Taking the portion of population defined as "good clients" as a target variable and the relevant aspects as independent variables of a model, we have the proper frame for the identification of the best clients for the bank.
The logistic regression gives us the scored list of best bank clients in terms of probability of buying a savings product.
We qualified this procedure a probabilistic approach.

We have implemented these two approaches in the Enterprise Miner 3.0 Project based on the assessment tool and the regression tool. Figure 1 illustrates the Process Flow Diagram. As a result of joining together these two processes we obtained a score list in which clients were divided into nine categories. Seven derive from the different premium classes of the life insurance owners, two from the different probability levels assigned by the logistic model.

We first tested the procedures on a sample of ten company sales points (March, 2000) and then extended the experiment to all the company sales network including in the score list 15% (138.640) of the insurance company´s clients (June, 2000).

The Results
By December 2001 18.114 clients had opened a current account at the new bank.
The first evidence is that the percentage of bank clients is higher in the group of insurance clients included in the score list than in the group not extracted by the scoring process.

According to the data presented in the table, the probability of opening a current account among scored clients is three times (4,0/1,3) the probability calculated among non-scored clients.
Furthermore it is possible to observe that the probability of opening a current account varies among the clients according to the scoring class.

The redemption observed among clients singled out by the deterministic approach varies between 11,7% and 1,9%, which is always higher than the total redemption (1,6%).
Good results were also obtained through the probabilistic model.
The redemption is 3,1% for class ­1 (higher probability) and 1,5 for class ­2 (high probability: 1,3 is the percentage of redemption of not scored clients).

Conclusion
Although bench marking was conducted afterwards and not planned beforehand (we don´t know how thoroughly our score lists were used by the agents) the statistical outcome proved trustworthy: our lists singled out the right persons. These were the individuals to be contacted by the agents. In the future we plan to use the bank´s new clients who opened a current account as the target of a new logit model. This new model will be compared with our present model and utilised to substitute the probabilistic portion of the old scores.