The interquartile range, demystified.

01.What the range actually is

Strip away the terminology and the mechanics are simple. You have a set of results from uncontrolled comparables — operating margins, markups, whatever your profit level indicator happens to be. Sort them from lowest to highest. The 25th percentile is the value below which a quarter of the results fall. The 75th percentile is the value below which three quarters fall. The interquartile range is everything in between: the middle half of your comparable set, with the top and bottom quarters discarded.

The discarding is the entire purpose. Outliers in a comparable set are rarely outliers because the market is genuinely that wide. They are outliers because a company took a restructuring charge, or had an unusually good year, or is not as comparable as your screening process believed. Cutting the tails is a crude but effective way of limiting the damage those companies do to your conclusion.

This is worth stating plainly, because it is the part practitioners skip: the interquartile range exists because your comparables are imperfect. In the U.S. regulations, the ordering runs in exactly that direction. Under §1.482-1(e)(2)(iii)(B), if all comparability factors are met and there are no material differences between the tested party and the comparables — or any differences have adequate adjustments — the entire range of results is the arm's length range. It is only when comparability is less than perfect, which is nearly always, that §1.482-1(e)(2)(iii)(C) permits the range to be narrowed statistically to the interquartile range.

The IQR is a reliability haircut. It is what you use when you cannot defend the full range. Practitioners have inverted this over three decades of practice until the IQR is treated as the gold standard rather than the fallback, and while that inversion is now universal enough to be harmless in most audits, it should shape how much authority you grant the number.

"The interquartile range is a statistical adjustment for imperfect comparables, not a certificate of arm's length pricing."

The endpoints are not what Excel gives you by default

Here is the technical point that catches even experienced teams. There is no single mathematical definition of a percentile. There are several accepted conventions, they produce different answers on the same data, and the difference is largest exactly where it matters most — small comparable sets.

The U.S. regulations prescribe their own convention at §1.482-1(e)(2)(iii)(C). In substance: the 25th percentile is the lowest result such that at least 25 percent of the results sit at or below it, and where exactly 25 percent sit at or below it, you take the average of that result and the next one up. The 75th percentile follows the same logic in reverse. Read the provision directly rather than taking my summary for it — it is short, and it is the operative text.

Now compare that to what your model is probably doing. Excel's QUARTILE and QUARTILE.INC functions use linear interpolation on an (n−1) basis. QUARTILE.EXC uses an (n+1) basis. Neither is the regulatory convention. On a set of eight comparables, these three approaches can produce endpoints that differ by 40 to 80 basis points — enough to move a tested party from inside the range to outside it, or the reverse.

I am not suggesting that examiners routinely recompute quartiles by hand and challenge the difference. Most do not. But if your result sits within a hundred basis points of an endpoint, the convention you used is a live issue, and "that's what Excel returned" is not an answer you want to give in an information document request response.

02.Three things the range does not measure

It does not measure statistical confidence. An interquartile range looks like a confidence interval and is read like one, but it is not. A confidence interval reflects sampling uncertainty and narrows as your sample grows. The IQR reflects dispersion among the companies you happened to select, and it will happily narrow because you found eight tightly clustered comparables — or widen because one of your twelve had a bad year. There is no standard error here, no significance, no probability statement. A range built on seven comparables and a range built on seventy look identical on the page and carry very different evidentiary weight.

It does not measure comparability. The range is downstream of every screening decision made before it. If the search used the wrong industry codes, retained companies with material intangibles the tested party does not own, or rejected a loss-maker on grounds that would not have been applied to a profitable company, the range inherits all of it and shows no trace. A perfectly computed range over a poorly constructed comparable set is a precise answer to the wrong question.

It does not measure whether your method is right. This is the one that costs companies real money in controversy. A tested party can sit comfortably at the 55th percentile of a TNMM range and still lose the audit, because the examiner's objection was never about the margin. It was about which entity was characterized as the tested party, or whether TNMM should have been used at all, or whether the intangible-owning entity is entitled to the residual it is claiming. The range does not speak to any of that. It answers a narrow question — is this level of profitability observed among comparable independent companies — and stays silent on everything upstream of it.

03.Why "inside the range" is not a conclusion

Assume the range is correctly computed on a defensible comparable set. Your result falls inside it. Here is why that still is not the end of the analysis.

Position within the range carries information. Being at the 26th percentile every year for four consecutive years is technically compliant and analytically conspicuous. Independent companies fluctuate. A tested party that lands just inside the lower boundary with metronomic consistency looks like a party being managed to a boundary, and examiners read it that way. If that is your pattern, you want the functional analysis to explain why the low end is the economically right place for this entity to sit — routine functions, no meaningful risk assumption, guaranteed cost recovery — rather than leaving the examiner to supply their own explanation.

The comparison has to be like-for-like. A three-year weighted average range and a single-year tested party result are not comparable figures, and studies mix these more often than you would expect. Confirm that the averaging convention, the time period, and the PLI definition are identical on both sides. A tested party margin computed on a different cost base than the comparables' is a defect no range can absorb.

Falling outside has a defined consequence. Under §1.482-1(e)(3), where the result falls outside the arm's length range, the adjustment is ordinarily to the median of the comparable results — not to the nearest endpoint. The OECD position at paragraph 3.62 is similar in effect, pointing to the median or another appropriate measure of central tendency where the results are equally reliable. This matters for planning: the cost of being outside the range is measured from the median, so the distance between your result and the median is your real exposure, not the distance to the boundary.

Other authorities may not accept your range. A range that satisfies the U.S. regulations may be narrowed, recomputed, or set aside entirely by a counterparty jurisdiction with its own percentile rules, its own view on multi-year averaging, or a preference for local comparables. Two-sided exposure does not respect a one-sided range.

04.Five checks before you rely on a range

Run these against any study in front of you. None takes more than a few minutes.

• Recompute the endpoints. Take the comparable results from the appendix, apply the convention required in your jurisdiction, and confirm you reproduce the study's published range. If you cannot, ask the preparer which convention they used and why.
• Count the comparables. Note the number feeding the final range, not the number that entered the search. Fewer than six should prompt a conversation about whether the range is doing enough work to be relied on, and about what the local rules say when the set is thin.
• Check the tails you discarded. Look at the companies in the top and bottom quartiles. If the discarded companies are functionally closer to your tested party than several of the retained ones, your screening has a problem the statistics are concealing.
• Confirm both sides use the same measure. Same PLI, same time period, same averaging convention, same treatment of extraordinary items, for both the tested party and the comparables.
• Measure the distance to the median, not the boundary. That distance is your adjustment exposure if the range is ever rejected. Know it before an examiner calculates it for you.

05.A final word

The interquartile range earns its place in transfer pricing. It is a reasonable, administrable convention for handling the fact that no set of independent companies is ever a clean match for a controlled entity, and both the U.S. regulations and the OECD Guidelines are right to accommodate it.

What it cannot do is carry the weight of the analysis by itself. The range is the last three lines of a study, and its reliability is entirely determined by the two hundred pages preceding it — the method selection, the functional analysis, the search strategy, the acceptance and rejection decisions. When a range is challenged, the challenge almost never concerns the arithmetic. It concerns the inputs.

So when the next study lands on your desk and the summary tells you the tested party is comfortably within the range, treat that as the beginning of the review rather than the conclusion. Ask which convention produced the endpoints. Ask how many comparables survived. Ask what the range would look like if the two most marginal companies had been screened out. If the answers hold, you have something worth relying on.

If nobody can answer them, you do not have a range. You have a number.