z**l 发帖数: 82 | 1 两个segments分别作model,KS for the 1st segement is 40, the KS for the 2nd
sengment is 30. Two models validate on the total population (segment 1 &
segment 2), the KS will be 50.
谁能统计的理论解释这个现象? |
A*******s 发帖数: 3942 | 2 possible. one case is that u segment the population on a powerful predictor
in the model. Then within each segment, that predictor has less variability
than in the whole population, and thus lower the predictive power of the
model.
【在 z**l 的大作中提到】 : 两个segments分别作model,KS for the 1st segement is 40, the KS for the 2nd : sengment is 30. Two models validate on the total population (segment 1 & : segment 2), the KS will be 50. : 谁能统计的理论解释这个现象?
|
d*****s 发帖数: 1407 | 3 segmentation is designed to improve the overall ranking performance, is not
it?
【在 z**l 的大作中提到】 : 两个segments分别作model,KS for the 1st segement is 40, the KS for the 2nd : sengment is 30. Two models validate on the total population (segment 1 & : segment 2), the KS will be 50. : 谁能统计的理论解释这个现象?
|
t********l 发帖数: 996 | 4 Build model for different segment will improve the predictive power on each
segment rather than build just one model for the overall population.
It is normal to see KS on combined segments is larger than KS on each
segment. Especially two segments are in different cycle bucket having
different default rate, in the model that predicts the default rate, the
combined KS will be larger than any one of the KS but that combined KS does
not make sense. |
z**l 发帖数: 82 | 5 两个segments分别作model,KS for the 1st segement is 40, the KS for the 2nd
sengment is 30. Two models validate on the total population (segment 1 &
segment 2), the KS will be 50.
谁能统计的理论解释这个现象? |
A*******s 发帖数: 3942 | 6 possible. one case is that u segment the population on a powerful predictor
in the model. Then within each segment, that predictor has less variability
than in the whole population, and thus lower the predictive power of the
model.
【在 z**l 的大作中提到】 : 两个segments分别作model,KS for the 1st segement is 40, the KS for the 2nd : sengment is 30. Two models validate on the total population (segment 1 & : segment 2), the KS will be 50. : 谁能统计的理论解释这个现象?
|
d*****s 发帖数: 1407 | 7 segmentation is designed to improve the overall ranking performance, is not
it?
【在 z**l 的大作中提到】 : 两个segments分别作model,KS for the 1st segement is 40, the KS for the 2nd : sengment is 30. Two models validate on the total population (segment 1 & : segment 2), the KS will be 50. : 谁能统计的理论解释这个现象?
|
t********l 发帖数: 996 | 8 Build model for different segment will improve the predictive power on each
segment rather than build just one model for the overall population.
It is normal to see KS on combined segments is larger than KS on each
segment. Especially two segments are in different cycle bucket having
different default rate, in the model that predicts the default rate, the
combined KS will be larger than any one of the KS but that combined KS does
not make sense. |
z**l 发帖数: 82 | 9 If the two segments have the total different distributions, we cannot build
one model on the total populations.Usually,the model built on the total
population cannot beat the models built on the different segments. |
K******Q 发帖数: 62 | 10 想请问下Two models validate on the total population (segment 1 & 2)是指two
models分别validate on the total population,还是two models selected vars
combine together to validate on the total pop? |