Floating Point and Financial Software

[Martin Minow]

A 64-bit floating point number (i.e., C double) should be suitable for
financial software under the following conditions:

-- Money must be represented in integral units (cents, not dollars and
cents).

-- The maximum value to be computed must be less than about 10^17. This
   includes intermediate values.

-- Addition, subtraction, and multiplication by an integer are the
   only operators.

Under the above conditions, there should be no loss of precision.

However, when division is required (as in currency conversion
or interest rate computation), one must be careful to control
round-off error. For example, a mortage payment schedule might
be computed using true (non-exact) floating-point arithmetic, with
the last or first payment adjusted to cover any residual error.
(You might want to re-read Donn Parker's book on computer crime,
paying special attention to the "salami" method of embezzelment
by accumulating round-off errors in a private account.)

Note that not all financial computation needs to be done with "to
the penny" accuracy: even our own dearly beloved IRS allows
(indeed, encourages) us to compute our tax declaration using
a whole-dollar round-off method.

Martin Minow
minow at apple.com