Poisson regression exposure/offset

[fsociety]

I am currently trying to model claim frequency in an actuary model with varying exposures per data point varying between 0 and 1. Usually, this is tackled by incorporating the exposure as an offset to a Poisson regression model. However, I am unsure how to actually approach this within xgboost, preferably using the Python API.

I have found little information on that topic, but following resources have been already helpful:

XGBoost - Poisson distribution with varying exposure / offset

Poisson Regression in xgboost Fails for Low Frequencies

However, I am wondering whether someone can help me out and give guidance on how to tackle this issue.

[hcho3]

Usually, this is tackled by incorporating the exposure as an offset to a Poisson regression model.

Can you provide a reference that describes this practice? What is “exposure” and “offset”?

[fsociety]

@hcho3 Sure, sorry. I think this link provides a good short summary:

Poisson regression |