The gamma distribution is only defined for positive numbers. If you observe negative values then, as @nirgrahamuk was suggesting, the distribution can't be gamma. My guess is that a negative "cost" is fundamentally different from a positive cost.

Hi, as the other posters stated, you can only use Gamma with positive values. if the data contains negatives, these are errors and you should delete them. You may want to split the dataset into zero values and positive values so that you can model the positives with Gamma and estimate the no claim versus claim split some other way.