y*****n
发帖数: 5016 | 1 In financial industry, we track the change of the slope of log odds against
score bands to measure the deterioration of seperation power. on the other
hand, a parallel shift of the line, or say, a change of intercept should
indicate a overall increase/decrease of quality across all score bands
thereby strategists may need to adjust the cut-offs.
my question is, to measure the change of intercept, if i put the intercept
difference between current and baseline in the numerator, then what should
be in the denominator? i don't think it is appropriate to use the baseline
intercept. maybe the baseline overal log odds is a good choice for
denominator? any thoughts? | A*******s
发帖数: 3942 | 2 definitely not directly use the intercept in a logit model, since intercept
has no physical meaning at all. And think about if the intercept is very
close to zero or is negative.
Go back to scorecard alignment practice--usually we use two metrics--target
odds at a reference score, and points double the odds--to build a linear
transformation between xbeta and score. The former stands for absolute risk
level of the score (intercept), and the latter is about relative risk (
slope). So the answer to your question is quire clear--just to compare the
actual odds at reference score vs. the target odds at reference score.
against
【在 y*****n 的大作中提到】
: In financial industry, we track the change of the slope of log odds against
: score bands to measure the deterioration of seperation power. on the other
: hand, a parallel shift of the line, or say, a change of intercept should
: indicate a overall increase/decrease of quality across all score bands
: thereby strategists may need to adjust the cut-offs.
: my question is, to measure the change of intercept, if i put the intercept
: difference between current and baseline in the numerator, then what should
: be in the denominator? i don't think it is appropriate to use the baseline
: intercept. maybe the baseline overal log odds is a good choice for
: denominator? any thoughts?
| y*****n
发帖数: 5016 | 3 the question is -- if you want to measure the change as a ratio/percentage (
between actual odds and the target odds ), then this ratio/percentage will
be lower at high score bands (where the log odds are higher) and higher at
low score bands, given a parallel shift of the Log odds line (not the actual
logistic regression of course)
intercept
target
risk
【在 A*******s 的大作中提到】
: definitely not directly use the intercept in a logit model, since intercept
: has no physical meaning at all. And think about if the intercept is very
: close to zero or is negative.
: Go back to scorecard alignment practice--usually we use two metrics--target
: odds at a reference score, and points double the odds--to build a linear
: transformation between xbeta and score. The former stands for absolute risk
: level of the score (intercept), and the latter is about relative risk (
: slope). So the answer to your question is quire clear--just to compare the
: actual odds at reference score vs. the target odds at reference score.
:
| s*********e
发帖数: 1051 | 4 basically, you need to check the shift of the base odds at the reference
point. i will send you the macro later.
(
actual
【在 y*****n 的大作中提到】
: the question is -- if you want to measure the change as a ratio/percentage (
: between actual odds and the target odds ), then this ratio/percentage will
: be lower at high score bands (where the log odds are higher) and higher at
: low score bands, given a parallel shift of the Log odds line (not the actual
: logistic regression of course)
:
: intercept
: target
: risk
| A*******s
发帖数: 3942 | 5 the practice in my last job is to compare actual odds vs. benchmark/target
odds at a fixed reference score, and so there is no such concern about
different score bands.
probably you should try odds ratio, which is actual odds over benchmark/
target odds, equivalent to comparing difference in log odds and irrelevant
to what score it is referred to.
(
actual
【在 y*****n 的大作中提到】
: the question is -- if you want to measure the change as a ratio/percentage (
: between actual odds and the target odds ), then this ratio/percentage will
: be lower at high score bands (where the log odds are higher) and higher at
: low score bands, given a parallel shift of the Log odds line (not the actual
: logistic regression of course)
:
: intercept
: target
: risk
| y*****n
发帖数: 5016 | 6 i think i will do it in a simplified way. i will create a measurement = (
difference in intercept between two log-odds-vs-score-band regression lines)
/(base period average log-odds across score bands). if it exceeds 20% for 3
months, then i will suggest an adjustment of cut-offs in strategies. of
course, this will only apply to the situation where there is no significant
change in the slope. otherwise this measurment is meaningless.
in other words, the numerator is the absolute change in log-odds arcoss
score bands, while the denominator is the average log-odds across score
bands. |
|
