Line4 logoLine4

Search line4.ai

Interval arithmetic: how XBRL checks allow for rounding

Interval arithmetic treats each reported number as a range, set by its decimals: a value rounded to thousands could really be anything within 500 either side. A rule such as total equals the sum of its parts passes if the ranges overlap, so correctly rounded returns pass and real errors still fail. Line4 counted 34,344 of the 115,657 EBA 4.2 value rules using it, and 3,366 of 6,511 EIOPA 2.10.0 rules. Outside Europe, the SEC also works with ranges: its EDGAR XBRL Guide checks duplicate facts as closed intervals, so 500 with decimals -2 means 450 to 550.

Run the EBA or EIOPA rules on a sample filing, free in your browser →

Rounding is allowed for the same way the published rules ask. Nothing is uploaded.

Guide · 6 min read · Updated 21 November 2026 · By Jess Luper

Why exact equality fails good returns

Firms round. A bank that reports to the nearest thousand sends 118,000 for a true 117,600. Add up a column of rounded lines and the sum can miss the rounded total by a few thousand, even though every number is right.

A rule that demands exact equality would reject that return, and push firms to nudge numbers until the totals add up. The EBA asks them not to: its Filing Rules say the EBA XBRL validation rules use interval arithmetic for validation, and that “the validation checks will take into account the declared accuracy” (EBA XBRL Filing Rules v5.9, rule 2.18).

34,344of 115,657 value rules in the EBA 4.2 reporting frameworks compare numbers with interval arithmeticLine4 count of the EBA 4.2 frameworks package, November 2026 →
3,366of 6,511 value rules in EIOPA Solvency II 2.10.0 do the same, about halfLine4 count of the EIOPA 2.10.0 taxonomy, November 2026 →
69,355more EBA rules only test that one amount is not negative, which needs no rounding allowanceLine4 count of the same EBA package →

How a number becomes a range

The decimals setting on each fact says how accurate it is. A value v with decimals d stands for the range from v minus half of 10 to the power -d to v plus the same. Decimals -3 means within 500; decimals 2 means within half a cent; INF means exact.

The EBA Filing Rules give worked cases: 117822 at decimals -3 is “117322 to 118322”, and a profit margin of 0.093 at decimals 4 is “0.09295 to 0.09305”. XBRL International’s Calculations 1.1 uses the same closed range for rounded values, “[v-0.5*10^(0-d), v+0.5*10^(0-d)]” (section 5.2.3).

01AddAdd the low ends and add the high ends: [x1, x2] + [y1, y2] = [x1 + y1, x2 + y2] (Calculations 1.1, section 5.7). The range grows with each line added.
02CompareTwo values are treated as equal if their ranges overlap. Calculations 1.1 puts it as: the calculation is consistent “if CT overlaps with RT”, the calculated and reported totals (section 5.2).
03Greater or lessA greater-or-equal test passes unless the whole of one range sits below the other, so rounding alone can never make it fail.

An example: one total, rounded right and reported wrong

An EBA asset encumbrance rule, v23833_h, tests that one amount equals the sum of two others: iaf:numeric-equal(($a_0), (iaf:numeric-add($b_0,$c_0))). Here is how it treats two filings with every amount at decimals -3.

Two lines and a total, decimals -3
Line b
True 117,600, reported 118,000, range 117,500 to 118,500
Line c
True 236,600, reported 237,000, range 236,500 to 237,500
Sum of b and c
355,000 as written, range 354,000 to 356,000
Total a
True 354,200, reported 354,000, range 353,500 to 354,500
Interval arithmetic

The ranges 354,000 to 356,000 and 353,500 to 354,500 overlap, so the rule passes. Had the firm reported 352,000 (range 351,500 to 352,500), there is no overlap and the rule fails.

Exact equalityShow118,000 + 237,000 = 355,000, not 354,000. The rule fails a return in which every number is correctly rounded.

The rule text is from vr-v23833_h.xml in the EBA 4.2 frameworks package. The amounts are Line4’s illustration; the ranges follow EBA Filing Rules v5.9 and Calculations 1.1 section 5.2.3.

A rounded total that misses by a thousand can still be right. Interval arithmetic lets the check tell the difference.

What EBA and EIOPA rules actually do

The functions come from Eurofiling and carry the prefix iaf. All 7,370 EBA 4.2 rule files declare it; in EIOPA 2.10.0, the 3,366 rule files that use it do. Line4 read every value rule in both packages to see which ones call them.

Nearly every EBA rule that compares two reported amounts uses them. Of the 81,313 EBA rules that do not, 69,355 only check that one amount is not negative, and 6,130 compare an amount with a fixed number.

In one line
EBA 4.2 frameworks
34,344 of 115,657 value rules (3 in 10) use interval arithmetic
By framework
COREP 13,228 of 55,753; FINREP 8,916 of 21,383; IRRBB 2,931 of 4,092
EIOPA 2.10.0
3,366 of 6,511 value rules (about half), in 3,366 rule files
Functions defined in the packages
0: the validation software must supply them
In plain words

Most interval rules are a sum checked against a total, or one amount that must not exceed another. EIOPA leans more on subtraction (1,194 rules) and maximums (444).

Most used functions, EBA 4.2 (rules using each)Showiaf:numeric-equal 19,075 iaf:numeric-add 18,993 iaf:numeric-greater-equal-than 11,114 iaf:numeric-less-equal-than 5,823 iaf:sum 3,945 iaf:abs 2,934 iaf:numeric-multiply 2,873 Other 7 functions 794

Line4 count, November 2026, of every va:valueAssertion test in the val folders of EBA_XBRL_4.2_Reporting_Frameworks_4.2.0.0 and the EIOPA 2.10.0 taxonomy. A rule is counted once per function it uses, so the function counts add up to more than the rule count.

Interval arithmetic and Calculations 1.1

XBRL International’s Calculations 1.1, a Recommendation since 22 February 2023, brought the same idea to calculation linkbases. It has two modes, “round-to-nearest” and “truncation”, and processors must support both (section 5.2.1). Truncated values use a half-open range, [v, v+10^(0-d)) for a positive v (section 5.2.4).

A failed calculation raises calc11e:inconsistentCalculationUsingRounding or calc11e:inconsistentCalculationUsingTruncation, and these errors “do not render an XBRL report invalid” (section 3.1). EBA and EIOPA write their totals as XBRL Formula rules with iaf functions instead, so for them the rule’s severity decides.

What it means for supervisors

Write every comparison of two reported amounts with interval functions, not plain = or >=. Otherwise correctly rounded returns fail and firms learn to adjust their numbers.

Set minimum decimals, then trust them. The range is only as honest as the decimals a firm declares, which is why the EBA sets minimums in rule 2.18 and forbids truncating to thousands while claiming decimals -3.

Test your software on the edges. A sum of many lines builds a wide range, so check that your engine adds ranges the published way, and that a total just outside the range still fails.

Where AI fits, and where people decide

AI can explain a failed total in plain words: how far apart the two ranges were, and whether a rounding slip or a wrong line is the likelier cause.

The arithmetic is deterministic. Software works out the ranges the same way every time and keeps the result with the filing; people decide whether a rule should allow rounding at all.

Put it to workBefore your next validation release
  1. 1Does every rule comparing two reported amounts use interval functions?
  2. 2Does your engine supply the iaf functions itself, since the packages do not define them?
  3. 3Does it build each range from the fact’s declared decimals, and treat INF as exact?
  4. 4Does a wrong total just outside the range still fail in your tests?
  5. 5Do you check that firms do not declare more accuracy than they report?

Where Line4 fits

Line4 is SupTech software for financial regulators. It collects returns in XBRL, Excel and CSV by portal, API or MCP, validates them against your taxonomies and rules at the point of submission (see XBRL software for regulators), and puts AI to work reading them, with supervisors making every decision. If you work with a standard we have not yet run in production, we will say so plainly and show you how we would approach it.

Questions

What is interval arithmetic in XBRL?

A way of checking rounded numbers. Each value is treated as a range set by its decimals, such as 118,000 at decimals -3 meaning 117,500 to 118,500, and two values count as equal if their ranges overlap.

Why do EBA validation rules use interval arithmetic?

So that correctly rounded returns pass. The EBA Filing Rules say its validation rules use interval arithmetic and that the checks take the declared accuracy into account, so firms should not adjust numbers to make totals add up.

How is the range worked out from decimals?

A value v with decimals d covers v minus half of 10 to the power -d to v plus the same. Decimals -3 gives plus or minus 500, decimals 4 plus or minus 0.00005, and INF means the value is exact.

What are the iaf functions?

Eurofiling interval arithmetic functions used in EBA and EIOPA rules, such as iaf:numeric-equal, iaf:numeric-add and iaf:sum. They are not defined in the taxonomy packages, so the validation software supplies them.

How many EBA validation rules use interval arithmetic?

Line4 counted 34,344 of the 115,657 value rules in the EBA 4.2 reporting frameworks, about 3 in 10. Most of the rest only check that an amount is not negative, which needs no rounding allowance.

Does EIOPA use interval arithmetic too?

Yes. Line4 counted 3,366 of 6,511 value rules in the EIOPA Solvency II 2.10.0 taxonomy, about half, with iaf:numeric-equal the most common function.

How does Calculations 1.1 handle rounding?

It turns each rounded value into a closed range and judges a calculation consistent if the calculated and reported ranges overlap. It also has a truncation mode with half-open ranges; processors must support both.

How do regulators outside Europe deal with rounding?

The SEC also works with ranges. Its EDGAR XBRL Guide (November 2026) checks duplicate facts as closed intervals: a value of 500 with decimals -2 stands for anything from 450 to 550, and duplicates that do not overlap are an error. A calculation that does not add up does not make an EDGAR filing invalid, the guide says, while an EBA validation rule set as an error can stop one. Line4 applies each regulator’s own treatment, so the same rounding is judged the way that regulator judges it.

Share this guideShare on LinkedIn
Was this useful?

For supervisors

New guides and Insights, by email

One short email when we publish something worth your time, from Jess Luper. No spam, unsubscribe any time.