Tweedie Deviance Error Metric

bhctuary

I request adding the Tweedie Deviance Error metric with automated "p" selection.
This metric is very important for skewed (non-symmetric data) such as for insurance claims.

In general, I'd like to have a custom error option.

Thank you,

Bert

Pit

I agree, it would also be great to have the option to add a custom metric.

admin

Hi Bert and Pit,

v3.3 now has a custom metric option — you can define your own error metric formula using mean(), sum(), median(), maxval(), minval() with per-row actual and predicted variables, or combine built-in metrics like rms, mae, etc.

For Tweedie deviance specifically, for p=1.5 (common for insurance claims):

mean(2*(-4*sqrt(actual) + 2*actual/sqrt(predicted) + 2*sqrt(predicted)))

For general p (not 0 or 1), the per-observation deviance is:

2 * (actual^(2-p) / ((1-p)(2-p)) - actual * predicted^(1-p) / (1-p) + predicted^(2-p) / (2-p))

Just substitute your chosen p value into that formula and wrap it in mean(...). You can set it in the UI under "Custom metric" or via CLI with --custom-metric.