gbm with tweedie offset

[Giorgio Spedicato]

I am running a gbm with tweedie marginal distribution to predict loss cost.... I suspect that the gbm does not correclty hand the offset parameter: 

Using different exposures as offset does not alter the predicted outcome. In the example below I tried both log and un logaritmed exposures and i optained the same (unreliable) result...

I suggest the documentation to provide further details on how offset and link functions are handled in gbm

#first model: 

gbm <- h2o.gbm(x = c(predictors.numeric,predictors.categorical), 

  y = "losscost",

                learn_rate = 0.01,

                max_depth = 4,

                min_rows = 5,

                distribution="tweedie",

                offset_column = 'exposure',

                training_frame = trainSplit,

                validation_frame = validSplit,

                stopping_metric="AUTO")

gbm2 <- h2o.gbm(x = c(predictors.numeric,predictors.categorical), 

  y = "losscost",

                learn_rate = 0.01,

                max_depth = 4,

                min_rows = 5,

                distribution="tweedie",

                offset_column = 'log_exposure',

                training_frame = trainSplit,

                validation_frame = validSplit,

                stopping_metric="AUTO")

[mr.li...@gmail.com]

Giorgio,

It sounds like you're approaching this from the generalized linear model perspective. But keep in mind that GBM is an ensemble tree method, so the traditional log(loss/exposure) approach used in GLM's may not be as applicable here. Instead, I recommend you use pure premium as your target variable with no offset. 

                    ~ Li