The following tests are used for ranking:
· Kolmogorov-Smirnov (KS) Test: This test draws a cumulative bad distribution curve (in deciles or total bad values) and a cumulative good distribution curve (in deciles or total good values) against the descending ordered scores. The maximum vertical distance between these two curves is checked.
When the rank order test is done, the actual bad rate in each decile and the average of the risk score in each decile is computed. The average risk score in each decile is in descending order. The actual bad rate in each decile is computed, and if it is in descending order, then the model is said to be rank-ordered.
The expected result is that the highest number of bad values are concentrated in the top 1 or 2 deciles, that is, KS occurs in the 1st or 2nd deciles.
· Rank Order Test: In a rank order test, the actual bad distribution curve rate and the average of the risk score are computed in each decile. The average risk score in each decile is sorted in descending order. The actual bad rate in each decile is computed, and if the deciles are in descending order the model is said to be rank-ordered.
The expected result is that the rank order is confirmed.
· Lorenz Curve: Lorenz Curve is drawn between the cumulative percentage of bad values (in deciles or total bad values) and the cumulative percentage of all entities.
The expected result is that the curve should reach the ceiling early.
· Lift Curve: A lift is defined as the percentage of bad values in the deciles or the percentage of bad values in the whole population of all accounts. For a decile, if the lift is more than 1, it means that the bad ratio in the deciles is above the bad ratio in the whole population.
The expected result is that the lift is greater than 1 decile.