The model made by analysts. This does not have anything to do with fair provability in order to determine bias. Bias is determined initially at what the first set of odds are way before the map is picked and still do not change until rounds are played in the game. Therefore the way this model is calibrated in CSGO is prone to exploits since there is a clear asymmetry in information between this model and the betters who trust their decision. If the game is 14-10, but the winning side have a shit economy that if they get beat they double eco. The odds will still be 1.1 for current winning side, +3 for current losing side (hypothetical odds). But it will only take the losing team 2 significant rounds to come back and if they do the team that was winning then has to eco for last round and the odds see a big variance in changes. That's why it's so exploitable.
*The point I'm making is that when a game has multiple pivotal points that can change odd predictions there is no way a model nor analyst can quantify them in real-time or circumstantially. It will take something far way complex to determine this, but if you have information that this team plays well late mid-round and they are consistent with their performance with it then it's a bet I would be more than willing to take.