Actually it does have a weird impact. The influence seems to depend on the off/def ratio.
Let's take a winger that you want train defense. He practices defense atty and surely becomes better, but only up to a certain percentage of the primary atty. That percentage varies with every player and I found no dependane on player qualities. It can be 50% for one, 10% for another. If you keep training him defense beyound that percentage he will sudenly become worse. I tested that by counting scoring situations, creative pass situations and other things to give every player a performance rating, for overall 53 forwards. Then I put those numbers into a graph with defense on the x-axis, performance on the y-axix. The result was pretty much a sine wave.
Now I doubt they wanted the engine to behave sine wavily.

And maybe there are other hidden influences we don't know about. Energy loss is a big factor, too. But still intersting to see such graphs. It's a motivator to keep testing new things.