I hate to do this for the second week in a row, but again you cannot use RMSE for data which is bounded between [ 0, \infinity) as you are assuming the conditional distribution of your data allows for values which are negative. This issue is simpler than last week though insofar as our target should have been log-transformed- though other transforms or loss functions are also valid. I think Zindi should maybe consult deeper it's Data Science Specialists or researchers/people from industry before releasing competitions as a violation of these assumptions can bias insights gained from such an exercise. I would love to get people's thoughts on this, I know I got a good deal of support last week on this exact issue. I would love to also get suggestions on what people thought may have been more appropriate.