I'm a type 1 diabetic and data scientist. Estimating the causal effect of a unit of insulin or food on blood sugar is an absolute crap shoot. Consider that there's a +/-20% margin of error on the reported carbohydrates on nutrition facts. We might consider this irreducible error that just cannot be modelled (Maybe you could get a calorimeter, estimate the distribution of errors, and reduce that error somewhat). Therefore, even if we created a model that explained all explainable variance, we still have a 20% margin of error. If a meal has enough carbohydrates, a 20% overestimate of insulin requirements would lead to an insulin overdose that would kill you if the resulting low blood sugar is not dealt with. In other words, the irreducible variance is so large that a "perfect" model would regularly suggest lethal insulin doses.

My "solution" is to eat low-carb/keto as a "variance reduction" strategy. Still, removing carbs also introduces gluconeogenesis (the production of glucose from protein) as a factor to consider. The synthesis of protein to glucose also occurs on a much time different time horizon than the consumption of carbs themselves which has implications for insulin dosing and insulin type.

I could go on! But long story short, modelling blood glucose is bloody hard.