I have a simple feed forward neural network regression model that I'm training on customer data to predict their usage amount. The MAPE is above 50%. The data is heavily skewed and when I log transform the predictor set as well as response variable the MAPE comes down to 12%, however reverse transforming the predicted value to original scale blows up the MAPE again.
The predictor set is close to 80 variables which has a good representation of customer behavior. However, Pareto NBD model predicts at 34% MAPE with just 4 variables. The MAPE computed for Pareto NBD is at Decile level. Does training Neural network at decile level makes sense?
Highly appreciate inputs.