Page 43 - Regression Guideline for AMC
P. 43
MODEL STABILITY
• Shrinkage is a term from predic6ve analy6cs and data mining (“big data”). It references the fact that predic6ve models always work best for the data set on which they were developed. If the regression weights developed on one data set are applied to a second data set composed of completely new proper6es (assuming approximately the same market area), predicted property values for the second set of proper6es will not be as accurate as for the first set of proper6es. The difference in accuracy between the “training” data and the applied data is the amount of shrinkage.
• Sta6s6cians have determined that shrinkage is related to the number of cases in the original or training data set and that the larger the number of cases in the original data set, the less shrinkage, whereby the original es6mates retain close to their same level of accuracy. Beyond about 40 cases per predictor, there is diminishing returns and li.le addi6onal reduc6on in shrinkage making this the recommended op6mum. Good performance, however, can be obtained with 20 cases per predictor.
43
