Because there are just too many confounding variables, most of which we can't even count never mind account for.
In a metabolic ward we can certainly get closer to some form of accuracy, but at home it would take impractical levels of controlled repetition to try and average things out.
Perhaps this is purely a case of expectation management - it depends on how accurate and over what time period you want your predictions to work. As a T1, if I can get insulin right for a meal with a given nominal(!) carb count, keep the rise lower than say 5mmol/l above current BG and be back close (to within say a couple of mmol/l) after a couple of hours (depending on the food) then that works for me - I can then correct as needed.
I'm not aiming for nor expecting to predict my BG level 10h hence down to one decimal place, but what I do want to avoid is large excursions by predicting the most likely trajectory for my BG so I can then tweak later to make up for any of the many confounding factors I didn't take into account at the start.
And there are certainly lots of confounding factors, but with time the things that generate the largest effects can be identified (carbs and their type, insulin, exercise, and stress/illness to name the ones that come to mind), at least with some degree of certainty. Enough certainty that I'm willing to take the insulin, or not and be pretty sure that it will have a similar effect to the last time, or at the very least the behaviour will be constrained into a set of BG trajectories that I can then deal with further down the line. I don't need a very long time horizon because things will change, I'll eat something, do some exercise, etc., so long term predictions aren't all that useful except overnight, which is in many ways the easiest to analyse as there are probably fewer confounding factors: probably no insulin nor food involved and it's fairly repeatable.
The good thing about the factors I listed is that they are generally fairly obvious and produce effects that are not overly transient so one has both the opportunity to know they are potentially having an effect, and to see a change in BG response over enough time to be reasonably certain that the two are correlated at least. Stress can be a tricky one mind you.
What sorts of things are you looking at on the metabolic ward and what level of accuracy are you trying to achieve assuming you're predicting something in a time-based fashion. It sounds like you're trying to achieve an accuracy which is far higher than that I want/need.
P.S. The other thing to add, which is not at all the same as developing a theory/model in any other field that I know of, is that I really don't care very much about it being general - I only care that it works for me. Once that part is working acceptably, then I might start sharing what I thinks is happening just in case it helps others with their own models. This is where couching theories in the right terms is important, the usual "but we're all different" mantra, etc.