Generally, the further you go away from a reference frame, the worse the prediction. However, there are a couple of situations we can consider.
Let's consider the case that you want to match two good quality sources that have been color timed in a different way. It is possible that each frame was color timed differently, however usually the color timing within a shot or even in a whole scene are very similar.
Here's an example for the Raiders bluray and the WOWOW. First the model is calibrated on this frame:
Bluray:
WOWOW:
Bluray matched to WOWOW:
First, I predicted the corrections for in the same scene, within the same shot:
Bluray:
WOWOW:
Predicted correction with color correction model from the above reference frame:
Then I predicted the corrections for in the same scene, but not within the same shot, where the color timing differences are expected to be larger:
Bluray:
WOWOW:
Predicted correction with color correction model from the above reference frame:
The prediction is still very close. The color of the whiskey is actually more consistent with the reference in the prediction, than in the actual WOWOW frame. Certainly far better than could have been done manually without a good reference, but we should be careful using frames from a different shot.