Analyzing The Results

Running the model will first generate columns for all the different variables selected, building bands where appropriate. The tool will then work out which columns are related and ignore columns that are extraneous.

For example, if you selected ' Age' and ' Date of Birth' columns then one of these will be ignored by the model.

It will then build a sample set of data out of the whole database based on the specified sample size. If the sample size is 5,000 then it will take that number of people from the target group and the same number of people from the prospect pool. These records will then be used to build the model using a technique that gives coverage of all the selected variables to ensure that significant attributes that are only lightly represented are given the precedence that they deserve.

Traditional random sampling methods may sample out too many of these significant attributes to make them statistically useful.

A progress window will be displayed as the job processes, and once complete, results will be available under the following drop-down headings:

Coefficients and Ignored Columns

Coefficients and Ignored Columns allow you to further analyze the results generated after processing a model using the Predict Inclusion tool.

Running the model first generates the derivable columns from all the selected predictor variables, and builds bands where appropriate, for example, Age bands. The tool then works out which columns are related and ignore columns that are extraneous. For example, if 'Age' and 'Date of Birth' columns are selected, one of them will be ignored by the model.

The Tool then builds a 2,500 row group for target (AND universe) domain data and another 2,500 row group for non-target (AND universe) domain data. The records are used to build the model using a technique that covers all the selected variables, to ensure that significant attributes that are only lightly represented are given the precedence they deserve. Traditional random sampling methods may sample out too many of these significant attributes to make them statistically useful.

A progress window is displayed as the job processes, and once complete, results are available under the Coefficients and Ignored Columns drop-down headings.

Coefficients

Coefficients show each of the selected variables ranked by their significance, along with their individual model score, probability value and ZScore. The coefficient data for a value shows if it has a positive or negative influence on the result and how strong that influence is. The ZScore indicates how different the sample mean is compared to population mean - either above or below - and is measured in standard deviations.

The phi coefficient is a measure of the degree of association between the target and predictor variables.

Ignored Columns

Below is a snapshot of the columns that were ignored for this iteration of the model with the reason they were ignored:

Use Ignored Columns to help you identify columns that could be removed, or perhaps indicate that not enough data was supplied.

Gains CHart and Confusion Matrix

Shows a graph with three colored lines displaying the results of the model as follows:

• Yellow: Shows the gain for the modelled sample of records.
• Grey: Shows the gain for a completely random sample of records.
• Green: Shows the proportional improvement between the modelled sample and the completely random sample.
• The red dot indicates the point at which the modelled sample is achieving the greatest gain over a completely random sample of records.

Hover your mouse over the red dot and a screen will be displayed that shows detailed information about this optimal intersection.

Gain at top

The % improvement between the modelled sample and the random sample at the optimal point on the graph.

Score

The actual model score for people at this point.

% of Target

The % proportion of people from the target group that have been included at this point.

In Target

The actual number of people from the target group that have been included at this point.

Proportion

The proportion of records that were in the segment compared to the total number of records that the model predicted would be in the segment. In the example above it would be 8,057 as a proportion of (8,057 + 38,720) which in this case is 16.7%.

Confusion

A matrix showing the overall success of the model by comparing the actual sample of people that were included in the model with the sample that the model predicted. The intersections are as follows:

• 15,220 people were in the actual segment and the modelled segment.
• 7,826 people we in the actual segment and not in the model segment.
• 52,351 people were not in the actual segment but were in the modelled segment.
• 184,477 were not in the actual segment or the modelled segment.

The key figure in this illustration are the 52,351 people that have not yet purchased the product, but which the model predicts are most likely to purchase it because they are most like the people in the target segment according to the modelled attributes.

Engine

An Engine column is created containing the propensity value for each target record. You can select this key group of people in a segment by selecting all people with an index greater than 0.465, and then adding an exclude statement of people that have already purchased the product.

• Predict Inclusion Tool
  • Creating a Predict Inclusion Report
  • Analyzing the Results
  • Further Results Analysis