That's actually not the case. Some statistics are measurements, others are instead designed to have predictive function. Net rating can be (and is) predictive, but not for the function you've attempt to apply it.
The problem with your usage of net rating to gauge a playoff matchup between the Bulls and Cavs is twofold, both an issue of linearity, net rtg's are essentially averages over 82 games, and only ~4 of which are even relevant; and co-linearity, or in essence, the fact that these ratings are composed of discretely different lineups that perform substantially differently under various conditions.
What you're wanting to do is determine the outcome of a specific matchup, that will likely have irregular minute distribution (compared to the regular season) for the best performing lineups for the respective teams.
This is an entirely different question than to ask, how would a particular team do, on average, playing 30 randomly distributed teams over the course of 82 games. If instead, you wanted to predict an entire season, then you could use the Net Rtg as a predictive tool. But again, it fails for the purposes of predicting individual matchups, especially in a playoff setting.
For what you want to do, you would need to simulate the set of potential possessions using individual player metrics and aggregate the sum total to potentially predict the box score. This is most commonly done by using Markov chains of discrete states (i.e. possessions). But in essence, it is a non-trivial process.
