Is it possible to produce an own model accuracy measure that takes 100% - MAPE?
If MAPE is 5 % for example, the model accuracy would be 95%? Or is that statistically incorrect?
It's a simple transformation of the MAPE, so it's as "statistically correct" as using the MAPE in the first place.
Note that the MAPE may exceed 100%, so your accuracy may become negative. This can be disconcerting, and lead to strange behavior by forecast consumers, like truncating accuracy at 0%, which is not a good idea.
You may find this thread helpful: What are the shortcomings of the Mean Absolute Percentage Error (MAPE)?