Missing rounding modes?

[imirkin]

I was just playing with https://github.com/yaffle/BigDecimal to replace https://github.com/MikeMcl/big.js usage.

It seems like the proposed list in the README file, and the implementation above only provides floor/ceil, while big.js only provides up/down [in addition to half-up/half-down/etc]. The Java BigDecimal implementation of course has both sets.

I'd like to put in a good word for adding "up" and "down" to the roundingMode list. These are convenient for some financial applications, where negative/positive indicates whether one is long or short in a position, however the rounding is still expected to follow "positive" conventions, so rounding "up" would expect to make a number's magnitude larger. (Like let's say you're short 111234 shares, that's -111234. But then you want to see the number in thousands, so you want to see -112 rather than -111. And if you're long those same shares, you want to see 112 rather than 111. Or in some cases you do want the -111/111 in both cases. So up/down, not ceil/floor.)

In case there's any confusion, I'd expect:

let val = -5.1m;
expect(val.toFixed(0, "ceil")).toBe("-5");
expect(val.toFixed(0, "floor")).toBe("-6");
expect(val.toFixed(0, "up")).toBe("-6");
expect(val.toFixed(0, "down")).toBe("-5");

val = 5.1m;
expect(val.toFixed(0, "ceil")).toBe("6");
expect(val.toFixed(0, "floor")).toBe("5");
expect(val.toFixed(0, "up")).toBe("6");
expect(val.toFixed(0, "down")).toBe("5");

[jessealama]

My current thinking is that Decimal ought to provide the five rounding modes specified in IEEE 754:

• round ties to even (default, same as in the various Math functions)
• round ties up
• round down
• round up
• round to zero

The round method would support all five, as would all mathematical operations and the constructor. How does that sound?

[ptomato]

I would suggest that it should provide the same set of rounding modes as NumberFormat and Temporal (expand, trunc, ceil, floor, halfExpand, halfTrunc, halfCeil, halfFloor, halfEven.) Four of these correspond to Math functions:

• Math.round() - halfExpand (ties away from zero)
• Math.trunc() - trunc (towards zero)
• Math.ceil() - ceil (up)
• Math.floor() - floor (down) I'd suggest that the default should be halfExpand since that's what many people mean colloquially when they say "rounding".

[imirkin]

I'd like to second @ptomato's suggestion. IEEE 754 rounding makes sense for floating point contexts (maybe?), but for base-10 decimals, I think the Java BigDecimal is the gold standard -- other libraries tend to match it. You can find similar designs in decimal.js which is also used by many needing the functionality (or big.js as I do, which has less functionality, but by the same author). I'm not too fussed about the names, but there are reasons why these libraries support the rounding modes -- they're not hard to support, and each has its use-cases. Further, any logic which attempts to match what these do would greatly benefit from having this built-in.

I would also second the default mode being half-up (or halfExpand in @ptomato's comment). This is what e.g. decimal.js does. For the Java BigDecimal, a new MathContext defaults to HALF_UP if not supplied in the argument (although the DECIMAL32/64/128 variants are indeed HALF_EVEN to match IEEE 754R).

[jessealama]

Thanks! This is excellent feedback. It seems that there's a bit of a conflict here about which way to go:

• Follow IEEE 754 and Math's lead and use half-to-even as the default.
• Follow NumberFormat, Temporal, and various decimal JS libraries's lead and use half-expand as the default.

However, the signals point toward using half-expand as the default. The other options should at least exist, if one wishes to go against the default. (Supporting them is not difficult.)

[munrocket]

Crypto exchanges and banks will need banker rounding or in other words halfEven. Rationale behind this method is fairer and more accurate statistical representation.

Rounding mode in divide operation also needed? It will be in functional style or some state? For example new Decimal128('1').divide(new Decimal128('3')) will give infinite 0.3333333333333333333333334 with some rounding by default in the end.

[jessealama]

Rounding mode in divide operation also needed? It will be in functional style or some state? For example new Decimal128('1').divide(new Decimal128('3')) will give infinite 0.3333333333333333333333334 with some rounding by default in the end.

I'm not quite sure I understand. The intention is that all arithmetic operations take an optional argument (actually, an options bag) where one can specify the rounding mode. Perhaps I didn't document this correctly (or somehow missed divide)?

[jessealama]

Just to update this thread on the status quo, as of writing:

We'll have a single round method that allows the programmer to use one of the five IEEE 754 standard rounding modes (of which the default is "banker's rounding" AKA halfEven). So trunc, floor, ceil, and even half-expand will be available (not as separate methods, though).

I think that answers the original question?