4 ms·

Great article, and a wonderful explanation as to why one should avoid (at all costs) doing financial calculations using IEEE-754 floating point math. > ... whe

by dvt 6y ago

Great article, and a wonderful explanation as to why one should avoid (at all costs) doing financial calculations using IEEE-754 floating point math. > ... where each \epsilon is very small, on the other of \epsilon_machine. Not sure if author is reading replies here, but I think "other" should be "order."

jhj 6y ago
IEEE 754 also provides for decimal radix floating point, just that there is no real hardware support for it beyond some IBM and Fujitsu chips, and what you can put together on FPGAs.
cesaref 6y ago
... except that every financial trading system i've ever seen is full of floats (and i've seen a lot).
- bigiain 6y ago
  Many years back I worked on a project involved in heavily government regulated gambling (horse race betting). The rules were every calculation including all intermediate steps had to be done to at least 1/10,000th of a cent precision, and rounding was only allowed once at the very final step. (Which, since my part of that project was a Javascript/jQueryMobile mobile webapp, I managed to push all that to the backend team saying "I'm doing calculations using Javascript, I don't even have proper integers to work with here - you'll need to send me all financial figures as strings, and do all the calculations at your end according to the rules. I can accurately display strings, I will not look a judge in the eye and try to claim I can accurately calculate _anything_ monetary in Javascript.")
KMag 6y ago
That depends on the type of the financial calculations you're doing. Certainly accounting and billing shouldn't be done in binary floating point formats. Financial modelling and forecasting calculations aren't exact calculations, and the intermediate library functions for binary floating point may suffer less rounding error than practical alternatives. Also, computational efficiency comes into play once you're at the scale of spending millions of dollars per year on hardware and electricity for financial modelling. Also, since IEEE-754-2008, there are decimal IEEE-754 floating point numbers, which are useful for some types of financial calculations.

by dvt 6y ago

jhj 6y ago

IEEE 754 also provides for decimal radix floating point, just that there is no real hardware support for it beyond some IBM and Fujitsu chips, and what you can put together on FPGAs.

cesaref 6y ago

... except that every financial trading system i've ever seen is full of floats (and i've seen a lot).

bigiain 6y ago
Many years back I worked on a project involved in heavily government regulated gambling (horse race betting). The rules were every calculation including all intermediate steps had to be done to at least 1/10,000th of a cent precision, and rounding was only allowed once at the very final step. (Which, since my part of that project was a Javascript/jQueryMobile mobile webapp, I managed to push all that to the backend team saying "I'm doing calculations using Javascript, I don't even have proper integers to work with here - you'll need to send me all financial figures as strings, and do all the calculations at your end according to the rules. I can accurately display strings, I will not look a judge in the eye and try to claim I can accurately calculate _anything_ monetary in Javascript.")

KMag 6y ago

That depends on the type of the financial calculations you're doing. Certainly accounting and billing shouldn't be done in binary floating point formats. Financial modelling and forecasting calculations aren't exact calculations, and the intermediate library functions for binary floating point may suffer less rounding error than practical alternatives. Also, computational efficiency comes into play once you're at the scale of spending millions of dollars per year on hardware and electricity for financial modelling. Also, since IEEE-754-2008, there are decimal IEEE-754 floating point numbers, which are useful for some types of financial calculations.