I don't like to comment on other testing threads but this is plain misinformation. A test should be conducted in such a way that as much randomness as possible is removed whilst retaining the core mechanics of the game. When the attributes are changed in such a way that they no longer resemble the actual gameplay, then the results from those tests are null and void. In my day job we discard gigabytes of data when a test fails to adhere to the core mechanics of the software being tested.
I don't know if you actually read my thread, but I have setup the West Ham team in such a way that there is no bias towards any tactic. All positions have natural foot players. This was done by mirroring players feet to the other foot. ie Payet 12 left foot, 20 right foot was changed to 20 left foot, 12 right foot, so he would perform equally as well in a left winger or right inside forward role.
Also all the best managers in the world have been assigned to the league. I have checked every single managers tactical tendencies and made sure that some teams would always attack and some would mostly defend. This way it doesn't matter if the test team is a top team or an underdog, as the AI will play the same way. I have also made sure the test team faces a variety of popular formations and not the same 4-2-3-1 in every match.
It is impossible to remove all the randomness from the game without sacrificing the core mechanics, however I am confident that if a test does well in my DB, then it will perform as well or better in real gameplay.
Also I don't know how you are playing the match as you have not provided this information. From my own testing and Will's testing, holiday and instant result does not correlate in any way shape or form to the actual game. Holiday mode uses a very simple algorithmn to compute the result and instant result is pretty much the same.
For further clarification please see my thread.