Using the Right Kind of Data (2,213 words)

Koslowsky maintains that a test mailing or selection from a file must yield at least 1,000 records for analysis.

"And," Lynde adds, "there must be at least some history attached to a record before we will score them in our modeling process. We typically look for at least two transactions."

With all the data assembled, the model builder can analyze components equally or weight them individually to place emphasis on the most predictive data.