Exponential Problem with Natural Consolidation

[conray]

Whats wrong with TM1 displaying values like (0,000000000000)?
I have a natural consolidation [CD Amt], calculated using [Std Amt] - [Act Amt].
By calculating it manually, i get a value of 0.
But the natural consolidation gives a (0,0000000000) value as shown in the attachment.
I have checked that there is no rounding issues and the values added up indeed is 0.

May i know what else could be the problem here?

Natural Consolidation Snapshot.xlsx

Excel file with snapshot data

(9.39 KiB) Downloaded 852 times

Cubeview screenshot

Natural Consolidation with exponential.gif (44.02 KiB) Viewed 17763 times

[ioscat]

http://www-01.ibm.com/support/docview.w ... wg1PM87649

Zero is not always a zero

[cgaunt]

Whilst I agree with Isocat that zero is not always zero, you should be able to control this affect in the formatting properties for the element in question.

[ioscat]

And you should be very careful in using this non-zero values

[Duncan P]

For those that don't have access to the IBM APAR system the link ioscat includes has this to say

Calculations may result in a very small number, but not zero. If this value is used as the denominator with the 'save divide' operator "\", the division will be done resulting in a very large number, instead of treating the very small calculated number as zero.

An example of this is when you calculate the reciprocal of one minus ten tenths :-

Code: Select all

1 \ ( 1 - ( 0.1 + 0.1 + 0.1 + 0.1 + 0.1 + 0.1 + 0.1 + 0.1 + 0.1 + 0.1 ) )

which gives 9007199254740900 and not 0.

Like ioscat said, you need to be careful.

[rmackenzie]

An example of this is when you calculate the reciprocal of one minus ten tenths :-

Code: Select all
1 \ ( 1 - ( 0.1 + 0.1 + 0.1 + 0.1 + 0.1 + 0.1 + 0.1 + 0.1 + 0.1 + 0.1 ) )

which gives 9007199254740900 and not 0.

Duncan, that's interesting - if you substitute these rules then you get the correct answer of 0 per usage of 'safe divide':

Code: Select all

1 \ ( 1 - ( 0.5 + 0.5 ) );

Code: Select all

1 \ ( 1 - ( 0.25 + 0.25 + 0.25 + 0.25 ) );

Code: Select all

1 \ ( 1 - ( 0.2 + 0.2 + 0.2 + 0.2 + 0.2 ) );

Code: Select all

1 \ ( 1 - ( 0.125 + 0.125 + 0.125 + 0.125 + 0.125 + 0.125 + 0.125 + 0.125 ) );

But then try this which gives -4503599627370400:

Code: Select all

1 \ ( 1 - ( 0.05 + 0.05 + 0.05 + 0.05 + 0.05 + 0.05 + 0.05 + 0.05 + 0.05 + 0.05 + 0.05 + 0.05 + 0.05 + 0.05 + 0.05 + 0.05 + 0.05 + 0.05 + 0.05 + 0.05) );

What's up with tenths and twentieths that works fine with halves, quarters, fifths and eighths? Looks like some sort of 'issue exists where denominator is 10 or greater' type of issue, perhaps?

This strikes me as a bug, wouldn't you agree? Plenty of customers that I've worked with think this problem should go away.

[Andy Key]

I would go for it being a limitation of IEEE-754 and the conversion of decimal to binary rather than a bug.
Do the same calculations (switching the \ to /) in Excel and you'll get the same results.
Now try doing the same with 64ths i.e. a number that can be represented in a precise manner in both decimal and binary (and is short enough to fit in to IEEE-754)

Code: Select all

1 \ ( 1 - ( 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625 + 0.015625  + 0.015625 + 0.015625 + 0.015625 + 0.015625 ) )

in Excel you get a #Div/0! error and you get a full fat zero in TM1.

[rmackenzie]

I would go for it being a limitation of IEEE-754 and the conversion of decimal to binary rather than a bug.
Do the same calculations (switching the \ to /) in Excel and you'll get the same results.

Seems to work ok on the calculator on the Windows desktop

Seriously, I've seen TM1 give some flakey rounding errors on standard consolidations. When the balance sheet stops balancing, it becomes a problem for people for whom an explanation involving IEEE-754 just doesn't cut it. On that point, I'm familiar with some of the issues in binary floating point arithmetic but are we saying that there's no solution, or advance, on a problem that's inherent to a standard dating back to, what, 1985??

[Duncan P]

The problem is that the only way to get the speed is to use a representation on which the standard arithmetic operations can be done in hardware. IEEE754-2008 defines a number of decimal floating point representations but they are only supported in a small range of hardware (e.g. IBM Power6 and z9) and not on the standard Intel stuff. Hence we are stuck with the problems of rounding recurring binary representations of 1/5.

[lotsaram]

head_is_spinning

[rmackenzie]

Hey Duncan, thanks for the response - always welcome and appreciated. Just to play devil's advocate for a moment - wouldn't it be an interesting avenue to explore for TM1 to offer the same 'Precision as displayed' option that is referred to in this Microsoft support note? That's kind of what the OP was referring to in the third post of the thread and not entirely worlds apart from the existing SpreadingPrecision parameter available today in the tm1s.cfg ??

[rmackenzie]

head_is_spinning

She'll be right

[conray]

So in summary, it is not possible to get a zero, unless using the "safe divide" method in the rule itself?

[ioscat]

Possibly yes and possibly no
Depends on

[conray]

Possibly yes and possibly no
Depends on

depends on what?
It really is annoying to get something like this with only 2 decimal point involve:

Data

ScreenHunter_04 Jun. 12 12.38.gif (10.96 KiB) Viewed 17499 times

look at all the (0.00)...

[rmackenzie]

So in summary, it is not possible to get a zero, unless using the "safe divide" method in the rule itself?

No, the safe divide is no guarantee of getting absolute zero. The summary of the problem is that human numbering system (decimal) and computer numbering system (binary) don't always place nicely. Sometimes there are extremely small rounding errors meaning that the zero value is actually an extremely small number. This leads to your zero suppression problem. Maybe what we need is some sort of config parameter like:

Code: Select all

ResolveToNullAtNthDecimalPlace=2

[Tilo]

IBM states that this would be fixed in TM1 10.2 but it is not entirely if even or IBM means a Version above 10.2.0.

Create a formula as follows:
['a'] = N: ['b'];
where values for 'a' are entered directly and afterwards there is plain consolidation.
Then put 'a' and 'b' into a consolidation 'c'
where 'a' has the weight 1 and
b has the weight -1
then very often the result is c<>0 (e.g. 0,000000000000019387).
(at least in 10.2.0 and with our model).
This causes zero suppression Problems, calculation Problems (especially when using multiplication or backslash division, overfeeding

This does not happen (in our model) if you can create 'b' as consolidation of 'a' with the weight 1.

[paulsimon]

Hi

There is a TM1 Rule function ROUNDP

Won't that do what you need?

If all else fails use something like IF( ABS( [ cell ref ] ) < 0.0001 , 0 , [cell ref] )

Regards

Paul Simon

[tm1ist]

When using a binary representation for numbers, it is expected to have very slight variations in numbers. It is because we use the decimal system and the binary system cannot represent all of the decimal numbers. It is vice versa, decimal numbers cannot represent some numbers in different bases as well. For instance, 1/3 is 0.33333333...33(which is never ending) in decimal system and when you multiply that 0.3333....33 with 3, you cannot get exactly 1. Long story short, as others explained, it is due to binary number system.

For financial applications, it is suggested to use decimal system because our monetary system is decimal and we don't have amounts like 1/3 USD which decimal system cannot represent. However, computer interpretation of decimal numbers takes longer time and memory, because the computer hardware works with 0 and 1(which is binary). This is not the case just for TM1 but also for any other software. In a programming language, when you compare the performance of a code piece when you use variable type double(which stores numbers as binary representation) or variable type decimal(stores as decimal representation), you will see that it will be faster when use variable type double. However, with double type you will have number precision issues like you do now with TM1. Also note that this precision issue is not just about 0 but for any number that is derived from some calculations of some noninteger values.

Therefore, what TM1 does is sacrificing monetary(base 10) number precision for faster performance. And, it gives you the chance to use a rounding function when you require the numbers to be exactly zero.

And the reason why 0.5, 0.25 etc can be represented in binary and 0.1, 0.2 cannot is about the exponents of 2. 2 exponent -1 is 1/2(0.5), 2 exponent -2 is 1/4(0.25) etc.