If you're a glutton for punishment, here is some additional information as to how floating point calculations can vary (regardless of whether or not that ultimately is the problem we're seeing):
http://docs.sun.com/source/806-3568/ncg_goldberg.html Mike --- In [email protected], "Mike" <sfclimb...@...> wrote: > > I have the same thing running on a Pentium III vs. a Pentium 4. I am assuming > (as in blind guess) that it is related to floating point rounding between the > CPUs. > > All it takes is a very tiny difference to send cascading changes in later > results. The taking or leaving of a single trade will impact the fitness > function which in turn will dictate the parameters for the next out of sample > period. > > Mike > > --- In [email protected], "jacklweinberg" <jacklweinberg@> wrote: > > > > Question: > > How is it possible for a walkforward test to produce different results on > > two different computers? > > Background: > > 1. AB program, database, system, watchlist, walkforward settings etc. are > > on a usb key that is moved from one computer to another. > > 2. Before the walk forward test, the same walkforward settings file is > > loaded. The same system is executed. The results are also put into the same > > usb key. > > > > Why do two different computers produce different results? (and strangely > > enough, very different results!) > > >
