You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
' There’s currently a bad bug though in how chains of calculations are dealt with. As it stands, in a chain of calculations X -> Y -> Z, it appears that random numbers are sampled for X to calculate distributions in Y, but then to calculate the distribution of Z the values of Y are resampled. The effect is that correlations are ignored entirely. What should happen is that random numbers are sampled from X and used to calculate the whole chain of values (Y and Z). Here is a quick proof of the error: http://getguesstimate.com/models/2218. For two data cells I’ve labelled A and B (e.g. X = {A,B}), there are separate cells for C=A/(A+B) and D=B/(A+B) (so Y = {C,D}), and finally a cell that is the sum of the two fractions E=C+D. This should be identically 1, because A/(A+B) + B/(A+B) = 1. Instead, an error range is found for E because the values for C and D are resampled ignoring their anti-correlation. Doing the entire computation in one step, shown in the cell labelled F, works correctly because the values sampled for X={A,B} are used directly in the computation of F=A/(A+B) + C/(C+D). '
The text was updated successfully, but these errors were encountered:
From a intercom request:
' There’s currently a bad bug though in how chains of calculations are dealt with. As it stands, in a chain of calculations X -> Y -> Z, it appears that random numbers are sampled for X to calculate distributions in Y, but then to calculate the distribution of Z the values of Y are resampled. The effect is that correlations are ignored entirely. What should happen is that random numbers are sampled from X and used to calculate the whole chain of values (Y and Z). Here is a quick proof of the error: http://getguesstimate.com/models/2218. For two data cells I’ve labelled A and B (e.g. X = {A,B}), there are separate cells for C=A/(A+B) and D=B/(A+B) (so Y = {C,D}), and finally a cell that is the sum of the two fractions E=C+D. This should be identically 1, because A/(A+B) + B/(A+B) = 1. Instead, an error range is found for E because the values for C and D are resampled ignoring their anti-correlation. Doing the entire computation in one step, shown in the cell labelled F, works correctly because the values sampled for X={A,B} are used directly in the computation of F=A/(A+B) + C/(C+D). '
The text was updated successfully, but these errors were encountered: