Forecast with Prophet - MAPE is too high

[albert_alaluf]

Hello all,

I'm not sure if this is the correct place to ask this question. I need a forecast for the coming 2 weeks and I use Prophet, ARIMA and ETS to compare which one is doing better? 

I spent several hours and I ended with ~30% MAPE errors. ARIMA and ETS even not close, much worse. I tried all the possible options such as auto model, manual, changing the parameters, 90-10 or 80-20 train/test ratio, the historical value which I can start from 2014 or later. After I tried all these combinations, my success was ~30%. It is weird that any sample datasets I get <10% MAPE error, but my data as a real data, no success. 

I cannot share the data, however, I like to hear your suggestions.

NOTE: The dataset has categories. Because each category has its own behavior, I run the model per category. I'm not sure this is worked well, but one model for all, I may lose the seasonality.

Avi