It seems wrong to consider 4 minutes late as on time, if the prediction is <5 minutes out. So I think the 90% in <5 minutes seems like an under-estimate.

The acceptable variance should be less than 20% of the predicted time. So if <5 minutes out, 1 minute difference makes a big difference and should be considered late. If it says 0 minutes but it takes 4, that’s a bigger problem than if it says 20 and it takes 24.