Best data type to store Financial/Monetary Values?

binsoii

Hi Thanks for pointing that out. What solution should i use then? I'm still confused. Sorry, i really need help about this. Because we would be developing an Accounting System next if this POS would work out. the thing is, i don't want to proceed with a complex solution if there is a common solution developers are using to handle this matter. You said just be realistic about the system, are you suggesting to just use double? But i will need to create reports down to at least 6 decimal places. Am i overcomplicating this? if so, can you point out how to handle this based on your experience? Thanks!

Regards!

kshegunov

I was summoned.
I'm looking at you @VRonin ;)

@binsoii

Forget inefficiency, it has no bearing here.

In my mind you have to decide between a fixed and floating point first. For finances I'd always go with fixed point if possible (especially if it's something banking related, a very touchy bunch those bank fellows ;)), as the precision isn't subject to scaling, in contrast to floating point representation.

If you decide to use floating point however, then you need to decide how to proceed - with standard double or with a multiprecision. Ideally you'd have a type that has a lot of bits for the mantissa and somewhat few for the exponent. Unfortunately we don't live in a perfect world and double was designed to keep general purpose data, so the width of the mantissa and the exponent are fixed. I don't know if boost::multiprecision allows changing this, but in any case it's worth researching. Also when doing any operations on the numbers you have to be absolutely sure you don't carry out the calculation error into the significant digits, so it's imperative to choose appropriate algorithms.

cfdev

qreal is the best type I think.

VRonin

@cfdev said in Best data type to store Financial/Monetary Values?:

qreal

http://doc.qt.io/qt-5/qtglobal.html#qreal-typedef qreal=double or, rarely, float

Chris Hennes

@binsoii said in Best data type to store Financial/Monetary Values?:

But i will need to create reports down to at least 6 decimal places.

If you really have a need to generate six decimal places regardless of the magnitude of the number you cannot use a floating point number for it, as @kshegunov mentions. By construction you have a system with a fixed decimal restriction. To be honest, my advice to you is to stay far away from developing this type of software until you are intimately familiar with the details of how computers store and operate on numbers.

kshegunov

It's probably feasible to get away with a 34.30 fixed point signed int64, which would give about 7-8 significant decimal digits after the dot with range of a few billion. However it gets really tricky when you need to provide even some basic functions - logarithms, (square) root(s), powers for such a representation.

To be honest, my advice to you is to stay far away from developing this type of software until you are intimately familiar with the details of how computers store and operate on numbers.

That's somewhat funny (in an ironic way), because this is probably one of the most complex problems one can encounter in programming, so it's really hard to be intimately familiar. :)

Chris Hennes

@kshegunov said in Best data type to store Financial/Monetary Values?:

this is probably one of the most complex problems one can encounter in programming, so it's really hard to be intimately familiar. :)

Absolutely true! I consider myself quite well-versed in floating point numbers (at least, of the IEEE 754 flavor), and I wouldn't trust myself to work on a financial system with these requirements.

Chris Kawa

Don't use floating point types for financial data and don't re-invent a wheel implementing fixed point functions.
There are libraries for this. For example: fixed_point - CppCon talk: John McFarlane “fixed_point"
There are more if you don't like this one, just google it. But don't try to do it yourself. You'll just waste time and re-make the same mistakes people already did and fixed.

binsoii

@Chris-Kawa Thanks! i'm not trying to re-invent the wheel by the way. i just want to use the existing ones already available and easy to use, just like the System.Decimal in C# .Net sitting there ready to be taken. Why there isn't just like this already available in Qt!? :( How could this amazing framework don't have a simple thing to address this. and i really don't like the idea of using third party libraries just for this case. (I don't even know where to start integrating the example you gave, i'm really new to Qt)

Cheers!

binsoii

@Chris-Hennes we already have an existing system, using Django and Angular to accomplish this. we used Decimal in postgresql and accounting.js library in the front end.

VRonin

@Chris-Hennes said in Best data type to store Financial/Monetary Values?:

To be honest, my advice to you is to stay far away from developing this type of software until you are intimately familiar with the details of how computers store and operate on numbers.

Strongly disagree. I don't know 10% of what's necessary to understand end to end encrypted communication but I use OpenSSL as I trust them to know what's necessary. Division of cognitive labor is what made humanity great. A reliable enough library (like boost) is all you need.

@binsoii said in Best data type to store Financial/Monetary Values?:

Why there isn't just like this already available in Qt!?

Qt integrates a lot more than you'd expect by a framework born to be a UI. Sometimes even the functionality integrated in Qt is not at par with what's available externally (see FTP support for an example).

@binsoii said in Best data type to store Financial/Monetary Values?:

and i really don't like the idea of using third party libraries just for this case

My suggestion involves using a library (boost) that is much closer to the standard than Qt is. I would not consider boost code quality inferior to Qt's in any aspect. <exageration>I don't think there's a C++ programmer that doesn't have boost ready to be used on his/her machine</exageration>.

Bottom Line:
if you are concerned with precision use boost::multiprecision::mpf_float_100 100 decimals (base10) precision is more that I can imagine anyone needing (you can still increase it arbitrarily btw, using boost::multiprecision::number<gmp_float<N> > where N is the number of base10 decimals precision you want)

Chris Hennes

@VRonin said in Best data type to store Financial/Monetary Values?:

Strongly disagree. I don't know 10% of what's necessary to understand end to end encrypted communication but I use OpenSSL as I trust them to know what's necessary. Division of cognitive labor is what made humanity great. A reliable enough library (like boost) is all you need.

In most cases you and I are in complete agreement on this point. The problem with finance is that the question isn't just reliability, it's regulation. In particular, does the regulatory environment this software is going to be used in specify when and how the rounding must take place? There are usually very, very specific requirements on software that handles financial transactions.

binsoii

@VRonin

if you are concerned with precision use boost::multiprecision::mpf_float_100 100 decimals (base10) precision is more that I can imagine anyone needing (you can still increase it arbitrarily btw, using boost::multiprecision::number<gmp_float<N> > where N is the number of base10 decimals precision you want)

Thanks a lot! boost::precision seems to be the best choice for my need. i will implement this and give an update once its done. PS (I don't where to start on how to use integrate/install this, i guess i have a lot studying to do. :D)

VRonin

@binsoii said in Best data type to store Financial/Monetary Values?:

I don't where to start on how to use integrate/install this, i guess i have a lot studying to do.

mpf_float_* only works on GNU compilers and depends on an external library.

You can use cpp_dec_float_100 it's twice as slow as mpf_float but all you have to do is download boost, add the path of booth in the INCLUDE += part of the .pro file and add #include <boost/multiprecision/cpp_dec_float.hpp> at the top of your source file

kshegunov

You do understand the problem here is not with the floating point precision, but it's with the floating point itself, right?
Floating point operations are exact to at most 1 epsilon which is the maximum relative difference between the two numbers. RELATIVE. I can even repeat it if it will help.

Suppose you are working on a 10-base floating point computer and you have 2 digits for mantissa and 1 digit for exponent. The absolute difference between two numbers with zero exponent is 0.01 by construction. Now, what is the absolute difference between two numbers with exponent 2? Well, it is 0.01 * 100 = 1. So you trade off your absolute precision for dynamic range. For most intents and purposes this is just perfectly fine!
When you need to do accounting, however, there are different regulations in place, and this trade-off is unacceptable! Your ABSOLUTE precision, can't be less than a specific amount, so that's why people use fixed-point, because the epsilon there is both relative and absolute measurement of the accuracy of an operation/number.

VRonin

@kshegunov

I'm not trying to be a smarta55, I'm honestly asking out of ignorance

Can You show me an example of operation that can compromise say the 50th decimal of a boost::multiprecision::cpp_dec_float_100 ?

kshegunov

@VRonin said in Best data type to store Financial/Monetary Values?:

You show me an example of operation that can compromise say the 50th decimal of a boost::multiprecision::cpp_dec_float_100 ?

Of course not. :)
That is, unless you're about to keep numbers in the magnitude of about 10 to the 40th power. We have no physical, tangible quantity that spans those ranges. Or to put some context in, as I'm a physicist after all:

the approximate diameter of the milky way (our galaxy) is in the magnitude 10^18 km, and has about 10^11 stars
the total number of atoms in the universe is estimated to about 10^80
we know physical constants with limited precision, but let's take one of the best known - the fine-structure constant (it's the exact number behind the atomic clock), so we know that with a certainty of about 10^-9.

Now look back and tell me, do you need 50 decimal places really? The point is there's already established way to represent numbers for this specific purpose, and there are regulations in place (I mean, really, they've written down in law how rounding should take place). Also note 50 decimal places would translate to about 150 bits (~18 bytes) for the mantissa alone ...!

As Chris said, it's been done, it's not new, it's known. So why would you venture into the depths of arbitrary precision arithmetic (i.e. brute force the solution) to combat a problem that's been solved way back?

jronald

Why not use cent as money unit and use int64?

boost::multiprecision seems a little complex, personally I don't like the style of boost/STL, reason:

syntax is not clear, even worse when considering the implementation
speed is not extremely fast, for example, recently I've checked comparison of string formatting mechanisms between C and C++, the C style is more easy to use and 2x faster than C++ style. BTW, Qt and C# adopt the C way.

tham

@jronald said in Best data type to store Financial/Monetary Values?:

Why not use cent as money unit and use int64?

boost::multiprecision seems a little complex, personally I don't like the style of boost/STL, reason:

syntax is not clear, even worse when considering the implementation

I think this is based on personal taste and skills, boost and many part of the stl leverage lots of generic programming and TMP, if you are not familiar with basic of generic programming, it is natural for you to find the syntax is weird.

About implementation, those library are not for average c++ programmer to maintain but for those programmers who are smart and love c++. As a season c++ programmer, I will expect they know and familiar with stl, basic generic programming and TMP.

Unless for me, api and the idea of stl is brilliant, it is master piece of a genius. Quality of boost are very high too.

speed is not extremely fast, for example, recently I've checked comparison of string formatting mechanisms between C and C++, the C style is more easy to use and 2x faster than C++ style. BTW, Qt and C# adopt the C way.

This depends on the purpose of the design philosophy, if you are talking about std::stringstream or boost::format, usually they are slower or much slower than c library(especially boost format), after all their main purpose are not blazing fast. This do not mean they are bad or poor, but they have different design purpose.

Do c api provide you type safety?Do c api manage your memory dynamic?Do c api provide you extension flexibility as std::stream provided? It is like compare apple with orange

Evidence

1 : std::sort is much faster than qsort of c, and I believe none of the c library can provide fast, light weight yet extensible algorithms api like stl provide(std::sort, std::transform, std::iota, std::set_difference etc), c do not have the expressive power of c++ provided. c++ can design almost any api c allowed to do, but there are many c++ api cannot be done by c
2 : boost spirit is very fast, even faster than c functions if you use it right and compile your codes on modern compiler

Both of the examples leverage generic programming and TMP, but their performance are great or superb, even better than standard c library.

If speed is what matter most, we are still writing machine codes today.

jronald

@tham
qsort is slower, but it can rewritten while keeping its interface unchanged.
Compare Net Library and STL/boost, the interface of C# is better than that of STL/boost, though C# is not good at low level things.