Reweighting
Also known as: Post-stratification · Sampling weights · Inverse probability weighting · Sample reweighting
Adjusting a sample back to the proportions of the population when the groups were drawn at an artificial ratio, so that estimated frequencies and lift reflect reality.
Legal basis
Wikipedia — Inverse probability weighting
In a test it is common to draw more cases with the event than they truly represent — otherwise the risk group is too small to measure. That creates a sample with an artificial proportion of events: good for estimating the score's power, but misleading for estimating frequencies, because the sample holds far more events than the portfolio does.
Reweighting corrects this by giving each group back the weight it has in the real population — the totals of the eligible base. With the right weights, lift and frequencies come to reflect the portfolio, not the sample. It is the same idea as post-stratification and inverse-probability weighting in surveys. The Gini, being pure ordering, does not depend on it.
Frequently asked questions
Why reweight a test sample?
Because test samples usually have an artificial proportion of events (oversampling of the risk group). Reweighting gives each group back the weight it has in the real population, so that estimated frequencies and lift reflect the portfolio rather than the sample.