Multi-Year Averaging in TNMM: Simple, Weighted and Period Methods
Single-year vs multi-year data in TNMM: when the OECD requires looking beyond one year, how simple and weighted averages are computed, and when averaging masks a trend.
A single financial year is a sample of size one. For a stable, mature, tested-party whose business is steady, one year is usually enough — and it is the year the Local File reports. But the OECD is explicit that a single year can be misleading, and where it is, the analysis must extend to multiple years. The questions then are: how many years, how to combine them, and how to keep the tested party and the pool on the same clock.
When multi-year data is required
The situations where one year does not tell the story:
| Situation | Why one year misleads |
|---|---|
| Start-up or loss years | The tested party (or comparables) have not reached steady state; the single-year PLI is a ramp, not a return |
| Seasonal businesses | The year-end position depends on where in the cycle the balance sheet sits |
| Cyclic industries | A peak or trough year is not the cycle; the cycle is the data |
| Restructuring or one-off events | A reorganization, disposal or acquisition distorts the year’s cost base and revenue |
| Volatile PLI across years | The indicator swings materially year to year for the tested party — the variance itself says one year is not representative |
The rule of thumb in the Guidelines: where a single year is unreliable for the tested party, use a multi-year period (commonly three) for both the tested party and the comparables — a tested party averaged over three years against a one-year pool is a mismatch the TPO will catch.
How the average is computed
The unit of averaging is the PLI per year, not the underlying totals. For each company and each year, compute the PLI on the study’s definitions; then combine the years:
| Method | Computation | Use when |
|---|---|---|
| Simple average | (PLI₁ + PLI₂ + PLI₃) ÷ 3 | The years are economically comparable — stable scale, no year dominates |
| Weighted average | Σ(PLIᵢ × weightᵢ), weights = revenue (or the PLI’s denominator base) in year i | Scale moved materially across the period — the bigger year carries more weight |
Two mechanical traps:
- Average the ratios, not the ratios of the sums. For a cost-based PLI, averaging OP/OC by year is the standard presentation; computing one ratio over three-year totals (ΣOP ÷ ΣOC) is a different number — defensible only if it is the declared method, applied identically to tested party and pool, and stated in the file.
- Loss years have no ratio. A company in loss in one of the three years has no OP/OC for that year. The treatment (exclude the year for that company, use the two years, or exclude the company) is a decision with a reason — see loss-making comparables.
Consistency: the tested party’s clock
The multi-year choice binds the whole study:
- Same period, all parties. Three years for the tested party means three years for every comparable, same fiscal periods (or the documented calendar-vs-fiscal mapping).
- Same method, all parties. Simple average for the tested party means simple average for the pool — a weighted pool against a simple tested party (or vice versa) is an unexplained variance with the tested party’s name on it.
- The working capital adjustment follows the period. WC ratios averaged over the same years, the same tax rate treatment, applied to the averaged PLI.
When averaging masks a trend
Averaging is a smoothing instrument, and a smooth line can hide the thing the examination cares about:
- The tested party declining, the pool stable. A three-year average puts the tested party’s weak current year next to its stronger past years — the average may sit inside the range while the current year sits outside. The TPO’s question is the current year’s economics; the file should address the trend explicitly (why the decline: a contract loss, a price war, a one-off) rather than let the average carry it.
- The pool drifting. If the comparables’ PLIs moved across the period (industry margin compression), the period-average range and the latest-year range are different ranges. Where they diverge materially, the file states which range it tests the tested party against and why.
- A one-off in the period. A year with an extraordinary event, averaged in silently, distorts the distribution. The event is either adjusted out (with the adjustment documented) or the year is excluded (with the reason).
The defence to “your average hides a bad current year” is not a better average — it is the trend analysis: year-by-year PLI for the tested party and the pool, the range per year, and the explanation of the movement. A file that shows the trend and explains it is a file that has used the multi-year data; a file that buries the trend under the average has misused it.
The documentation
The Local File records: the reason for the multi-year choice (the triggering situation), the period and its mapping for tested party and pool, the averaging method and its weights, the treatment of loss years and one-offs, and the year-by-year table alongside the average. The period decision is one of the examinable choices in the range layer — see the formula guide for where it sits in the three-layer machine.
See also
Run the screens as a study, not a spreadsheet
Quartyl applies the method, PLI and screening steps above as a pipeline — and keeps a documented reason for every exclusion.
Related docs
IQR vs Full Range: Choosing the Arm's Length Range
OECD's interquartile range versus the full range: quartile math step-by-step, when each is defensible, the mid-point argument, and how to document the choice in India.
Read docTransfer Pricing Benchmarking: Methodology, Data & Worked Study (2026)
The end-to-end benchmarking study: scoping, search design, quantitative and qualitative screening, adjustments, the arm's length range and refresh — with a worked Indian case.
Read doc