Prediction interval for return rate model was giving negative lower bounds for test data. Implemented a fix, but needs a more fundamental discussion across model types and interval types.
Current solution for rate model
calculate upper and lower bounds as p10 and p90 of the observed sample of rates
take a large sample of a normal distribution with those bounds
drop negative values from the sample
subsample the remaining values to get the user-specified sample size
The current solution was implemented in Feb 2021 as part of the return rate model update, to deal with negative lower bounds for some test data, but needs a broader discussion on this thread to verify/modify.
Issue
Prediction interval for return rate model was giving negative lower bounds for test data. Implemented a fix, but needs a more fundamental discussion across model types and interval types.
Current solution for rate model
Details in the wiki
Questions for Discussion
The current solution was implemented in Feb 2021 as part of the return rate model update, to deal with negative lower bounds for some test data, but needs a broader discussion on this thread to verify/modify.