And in run-of-the-mill, everyday finance, there simply isn't enough calculation stuffed in between the concrete monetary points that are recorded in the ledger.
> If you have a large number of costs to add up, for example, you should be able to add thousands of them, round to nearest 0.01, and get the right result.
> But for tax calculations, such as 10% of $21.15, 0.1 x 21.15 = 2.1149999999999998 in 64-bit IEEE floating point, and rounding the nearest 0.01 gives 2.11, not the 2.12 that we want.
This problem will be there even if we use integers for the currency amounts, but floating-point only for these fractional calculations.
Luckily for us Canadians, I'm pretty sure the Canada Customs and Revenue Agency won't care which way you call this rounding. They also don't collect or refund overall discrepancies of less than around two dollars in a single tax return.
I think I've been mostly rounding taxes down over the years, and tax credits up. E.g. if a tax credit is $235.981..., I make it 235.99.
The myth that has been foisted on programmers is that if you use floating-point for numbers, the actual ledgers won't balance, and sum totals of columns of figures will appear incorrect if verified by pencil-and-paper arithmetic. That will certainly be true if the math is done very carelessly; and it's true that it's easier to get it right with less care using integers.
A percentage calculation whose rounding is called the wrong direction will, in and of itself, not cause such a problem. E.g. if we split some sum of money into two complementary percentages, we can do it such that the two add up to the original.
You have to be careful not to do this as two independent percentages. Like, dont take 10% of 21.15 and then 90% of 21.15, individually round them to a penny, and then expect them to add up to 21.15. It has to be centround(21.15 - centround(.1 * 21.15)) to get the 90% residue.