Can we calculate PSI value if there is no training data?
if we can please suggest me the steps how can we do it.
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
2
Yes. PSI uses actual and expected datasets. These can be defined however you want. Expected could be training, hold-out, last quarter, last week, artificial, etc. Think of expected as the benchmark that the current (actual) is being compared to. Defining that is up to you and the problem you are trying to solve.
-
$\begingroup$ Thanks for the reply. but my model got trained 5 years back we don't have data with which model has built. I am having data for the last 12 months when I am trying to compare the data with last 12months all my PSI values are stable and good. can you help me what could be done in this situation if we have no data $\endgroup$ Jul 29, 2020 at 11:00
-
$\begingroup$ Then that might be the best you can do. Declare 1 month or 1 qtr or 1 week or 2 qtrs or whatever is right for your business problem as the benchmark. Then show how the most recent is similar or different from the benchmark. The idea is to have something meaningful and stable to compare against. Meaningful could be 3 qtrs ago this was our model performance, variable levels, ... and we will go from here. Document the decision and the rationale. $\endgroup$– CraigJul 29, 2020 at 11:53