I'm training a XGBoost Regressor to predict price which has a highly right skewed distribution. (all prices are positive) I took log transformation on the target thinking it would help to 1) stable the variance and 2) introduce non-linearities. And the log-transformed target yields very good MAPE/SMAPE. However, when I transformed the prediction back by taking exponential, and calculated MAPE/SMAPE, it is even worse than non-transformed target (original price).
Could not wrap my head around this, would appreciate any help.