So I'm working on an old Kaggle competition which requires you to predict the price of something, and the evaluation metric used is RMSLE. I found a tutorial for that data set, and the person in the video took the log of the prices before training the model, so that the model would predict the log of the actual price. Then, to see his RMSLE for the training set, he used RMSE on the predictions and the log-transformed prices (he got 0.31).
I trained the same model, but without taking the log before training the model, so the model would predict the actual price of the items. Then, I used RMSLE and I got 0.32.
So I was wondering, which one is usually more appropriate? Taking the log of the independent variable before training so the model would predict the log of the original independent variable, or not?