Palisade Knowledge Base

HomeTechniques and Tips@RISK Distribution FittingDiscrepancy from Fits Performed by Other Software

4.3. Discrepancy from Fits Performed by Other Software

Applies to: @RISK 5.x and newer, Professional and Industrial Editions

When I fit my data in @RISK, I get a very different result from the ________ software. Maybe @RISK fails to converge at all, or maybe it converges on a fit but the parameters are very different. Is there some setting I need to change?

Probably there is. Specifically, if the process that generated the data has a natural lower bound, you should specify that lower bound on the Distributions to Fit tab of the fitting dialog.

Why is this necessary? Many software packages assume a lower bound of zero for distributions that don't have a left-hand tail. Other packages, including @RISK, take a more general approach and make the lower bound subject to fitting also, as a shift factor. This allows, for instance, a distribution shaped like a log-normal but offset to left or right, if that matches the data best. But sometimes that is actually too much freedom, and @RISK fails to converge on a fit. (In general, "convergence failed" means that the numerical process of homing in on an answer for the MLE got stuck in a loop and couldn't finish.)

When the data have a natural lower bound, and you specify that lower bound to @RISK, it can do a better job of fitting more efficiently. Specifying the lower bound may even make the difference between "convergence failed" and a successful fit, as for example in some Weibull distributions with shape parameter less than 1.

On the Distributions to Fit tab of the fitting dialog, "bounded but unknown" restricts the fit to distributions that don't have left-hand tails, but it doesn't affect the fitting algorithm for those distributions. But when you specify a specific lower bound, then @RISK uses that as a fixed shift factor, and the mathematics of doing the fit are simplified.

Last edited: 2015-06-01

This page was: Helpful | Not Helpful