Applying ACTMAN: Statistical Scoring Models

Scoring is another method used for first-degree targeting. It determines the likelihood that a given prospect will become a first-time buyer by using statistical models to analyze the relationships between various customer activities and demographics.

The first step in scoring is to specify a dependent variable, usually "purchase/no purchase." Statistical analysis then quantifies the relationship between this dependent variable and a series of independent or explanatory variables. The outcome is a set of coefficients for the independent variables. The magnitude and sign of the coefficients reveals the relationship between the independent variables and the dependent purchase-decision variable.

Next, analysts enter prospect data for the independent variables into the statistical model. The resulting value for each prospect is referred to as its score, which ranks the likelihood that the prospect will make a purchase. Analysts then convert these scores into actual purchasing probabilities. An economic cutoff value marks the lowest score for probable prospects, and the firm targets only those customers who have a score above that cutoff point.

The specific steps for scoring are as follows.

	Step 1:	Send out a test mailing or use historical data from a marketing campaign to determine who purchases or does not purchase the product or service.
	Step 2:	Use a regression (or logit) model to correlate the independent (explanatory) variables to the dependent variable (e.g., "purchase/no purchase").
	Step 3:	Score prospects by entering their data for the independent variables into the statistical model.
	Step 4:	Group prospects into deciles based on their scores.
	Step 5:	Estimate the probability (likelihood) that prospects will purchase the product or service.

An example will help clarify this scoring method. Suppose a retailer wants to target customers for its one-hour film processing services. The retailer identifies three independent variables to determine the probability that a customer will buy these services: "months since last purchase," "total dollars spent," and "amount spent on film." ("Purchase/no purchase" is the dependent variable.) A regression model with these variables is run using data from selected customers. (This sample should include those who already use the one-hour processing and those who do not.) The resulting model's coefficients, or weights, for each independent variable appear in Table ACTMAN-2. The table's third column, marked variable value, contains a single customer's specific values for each independent variable. Multiplying the weight by the variable value gives a score for that variable; summing the scores of all variables (including the score for the "constant" value) produces the customer's total composite score.

Once the total score of each customer has been computed, the entire sample can be divided into deciles, with decile 1 containing the customers with the highest scores and decile 10 containing those with the lowest scores. The purchase probability for each decile then can be calculated based on the purchase rate of the customers in the decile. Table ACTMAN-3 contains the results of this analysis.

Scoring models have significant advantages. They are relatively easy to use, and they score each individual customer. More complex scoring models, such as Tobit models, can use the financial value of the customer, rather than "purchase, no purchase," as the dependent variable.