Turns out that the first thing a financial analyst does to compare two implementations of a bond securitization model is to look at the pennies. If those tie out, then you're good. But if your software predicts money off by a dime in a billion dollar model deal, the analyst WILL notice and WILL try to get to the bottom of it.
Yeah, floating point roundoff error seems like it should be immaterial. However it can matter a whole lot.