I don't know what you are working on but if you use fractions for what you are doing it would be interesting to hear about how you use them and your results. Because fractions are kept as an integer numerator and an integer denominator, they probably take up more memory than floats but less than ScaledDecimals. And because both the numerator and denominator can be large integers, fractions can be very precise. But because they are implemented in both software and hardware, then can be slow. However, many numerical functions seem to have fractions or divisions built into them. If fractions are used and the divisions not preformed until they absolutely have to be, then use of fractions could be faster than expected and may be faster than ScaledDecimals or floats. I have no proof of this but if you do anything in this area, I would love to hear about it.
I think all this is premature optimization for me :) as I'm only building an early prototype (I'm doing a start of Dempster Shafer Theory [1] implementation (actually Transferable Belief Model)... and it's won't reach a big size for a while. It allows to have an imprecise, incomplete even uncertain value for a proposition (sort of multi-valued attribute with confidence...). I use it to get expert opinion on values, it's a known technique for different captor data fusion, but in my case, it doesn't demand too much performance as the combination is not that important (compared to sensor data fusion) ;)
I don't think that this is a case of premature optimisation, so much as not considering what kinds of objects you are really working with. I used to do a lot of numerical modeling in mathematical biology, and I found that to get consistent results I had to create a Rational number class in C++ (linked to FORTRAN via a C wrapper function, which led to horrible debugging due to C++ name mangling). It seems to me that the name Fraction is a bit misleading, as 'fraction' implies division (which was horrendously slow on the hardware I used to use), whilst 'rational numbers' are indeed exact (and countable). I tend to avoid floats unless I need to use transcendental functions.
If you are looking at something that is divided into parts then rational numbers will always give better results, and as they can be implemented purely in Integer arithmetic they are rarely slow unless you are dealing with weird fractions where the numerator or denominator is greater than the machine Integer, I guess in Squeak that's if either is greater than 31 bits.
The big question is speed more important than correctness? I would always tend to go the correctness route, then look at optimisation as long as the errors introduced by the floating point processor didn't compromise correctness beyond acceptable measures.
In population ecology or epidemiological models using rational numbers avoided the annoying fractional instances (the 0.456 infected creature, or the 0.23 surviving bilbies for example) that could lead to totally meaningless results. The worst case I saw was a model of bilby population, where the model showed a sufficient population density for survival in relation to predators and resources, but the model actually was super optimistic due to its having fractional bilbies living close enough together to breed, analysis showed that the model was out by 2 orders of magnitude, and if it had been used in a real world situation the cute little critters would have disappeared from the proposed refuge within 5 years.