How the Fair Range
is worked out.
A Fair Range is a statistical estimate from comparable sales: a low, a middle and a high figure for what similar stones have sold for. This page sets out where the sales come from, how they are weighted, how sure each range is, and when we give none.
- Output
- Low · middle · high
- Confidence
- Graded A to D
- Comparables shown
- The 5 nearest
- Status
- Draft, not live
Three figures,
and how far to trust them.
Every range comes with its grade, the number of comparable sales behind it, and the five sales closest to the stone.
- P10 Low
- One in ten comparable sales went below it.
- P50 Middle
- Half went above it, half below.
- P90 High
- One in ten went above it.
The range
Per carat and for the whole stone, in US dollars, rounded to three significant figures. A range is not a quote to the dollar.
The grade
A to D, from how many close and recent sales stand behind the range, and how wide it came out.
The comparables
The five nearest completed sales, with their date, weight, source and price per carat, so you can check the range against the sales it came from.
From a sale
to a range.
The same steps, in the same order, for every stone.
- 01
Collect
Sales of stones with a report from GRS, SSEF, Gübelin, GIA, AGL or Lotus: auction results, our own completed sales, prices from a panel of dealers, and asking prices.
- 02
Describe
Each stone is coded the same way: variety, weight, colour, treatment, origin as the laboratory’s opinion, laboratory, number of reports and any phenomenon such as a star. The laboratory’s own wording is kept beside the codes.
- 03
Weigh
Each sale counts according to its source, as set out below. An asking price is discounted before it is used, because it is not a sale.
- 04
Fit
Three statistical models learn the low, middle and high price per carat from those sales, with a trend for time. Price per carat is never allowed to fall as weight rises.
- 05
Check
The range is adjusted so that about eight in ten of the most recent sales, held back from the models, fall inside it. Then we count the comparable sales close to your stone.
- 06
Publish or refuse
With enough comparables and nothing in the way, the range is published with its grade, its comparables and the model version. Otherwise we say why and show no figure.
Where the prices
come from.
Not every price says the same thing. A completed sale through Erebaur is the strongest evidence, an asking price the weakest.
| Source | What counts | Weight | How it is used |
|---|---|---|---|
| Erebaur completed sales | A sale that went through escrow between verified, unrelated parties. | 1.0 | Counts once the money has been released. A sale between related parties never counts. |
| Auction results | Realised prices at the major houses, with the laboratory details printed in the catalogue. | 0.8 | The buyer’s premium is taken off, so the figure is the hammer price. A lot sold as a jewel is marked, because its price includes the setting. |
| Dealer panel | Prices reported by a panel of dealers for stones they have sold. | 0.5 | Fills the gaps where auctions are thin, as in the smaller sizes. |
| Asking prices | What owners ask, on Erebaur or elsewhere. | 0.25 | Discounted by 15% in this draft, a figure to be set from how far asks and completed sales actually differ. Never shown as a comparable. |
Weights are relative: a completed sale through Erebaur counts four times as much as an asking price.
How sure
the range is.
A comparable sale is one close to your stone in weight, colour, treatment, origin and laboratory, and no more than four years old. The grade counts them and looks at how wide the range came out.
Grade A: 25 or more comparable sales and the high figure no more than about 1.6 times the low.
Grade B: 15 or more and the high figure no more than about twice the low.
Grade C: 8 or more and the high figure no more than about 2.7 times the low.
Grade D: 5 or more, with no limit on width. Read the range together with its comparables.
These thresholds are draft settings, to be reset against real sales before launch.
The segment
index.
Beside the ranges, the Prices page shows how each segment moves over time. This is how that index is worked out.
- What it measures
- The change in what comparable stones in one segment sell for, from quarter to quarter, with their weight, color, treatment, origin, laboratory and kind of sale held constant. The first quarter with enough sales is 100.
- Sales only
- Completed sales on Erebaur, auction results with the buyer’s premium taken off, and dealer panel prices, weighted 1.0, 0.8 and 0.5 as for the Fair Range. Asking prices are never in the index.
- How
- For each quarter, one regression of the log price per carat on the features of each stone and a marker for each quarter, over that quarter and up to four quarters before it. The step between the last two markers is the change, and the changes are chained.
- Minimum
- A quarter needs at least 10 sales, and a segment 40 sales across at least two such quarters. Below that we show “not enough sales”, never a number.
- Uncertainty
- Each figure comes with a 95% interval. It widens with fewer sales and with distance from the base quarter.
- Updates
- When a deal closes in a segment, its index is marked out of date and worked out again. A new quarter never changes the figures before it.
The index describes a segment of the market. It is not the value of any one stone.
Who pays us,
and what stops it showing.
The conflict
We earn a commission when a stone sells on Erebaur. A marketplace paid on sales has a reason to show ranges that flatter the asking price. That commission is an interest in the stone, so we never show a Fair Range on or beside a stone for sale here: it is shown in Price Check only. Saying so is the first safeguard. The rules beside this are the rest.
What limits it
- We never estimate a stone we have an interest in: one that we, or anyone working for us, own or part-own, hold on consignment or have lent against, or one for sale here, on which we earn a commission.
- The model never sees the asking price of the stone it is estimating.
- Every stone goes through the same model with the same settings. No range is adjusted by hand: a wrong one is withdrawn and the model is fixed.
- A seller cannot pay for a range, or for a better grade.
- Every range, and every refusal, is stored with the model version that produced it.
Four things
the Fair Range is not.
- It is notAn appraisal.
- InsteadAn appraisal is one expert’s opinion of one stone, written for a stated purpose such as insurance. The Fair Range is arithmetic over the sales of other stones.
- It is notWhat this stone is worth.
- InsteadIt is where similar stones have sold. A particular stone can sell outside it, for reasons no laboratory report records.
- It is notA market price that moves.
- InsteadA single coloured stone has no live price. The range changes only when the model is retrained, and each version is dated.
- It is notAdvice to buy or sell.
- InsteadIt is not investment advice, and it says nothing about what any stone will be worth later.
In the encyclopediaLaboratory reports and appraisals Types of value
Every range says
which model made it.
A version names the code and fingerprints the exact sales and settings it learned from, so two different models never share one. A range in Price Check can always be traced to the version that produced it.
fr-0.1.0Draft · October 2026
First version. Quantile gradient boosting on the logarithm of price per carat, weighted by source, calibrated on the most recent sales, with comparables and grades as described above. Tested on synthetic data only. No range has been published.
Format: fr-<code version>+<fingerprint>