Artmetria Index Methodology

The art market is not a single transparent market. It is fragmented across auction houses, galleries, private sales, fairs, regions, categories, periods and reputational networks.

Artmetria starts from this complexity.

The Index is built to help users compare signals without pretending that every signal has the same reliability, depth or meaning.

A high score should not be read as a recommendation. A low score should not be read as a dismissal. Each score is a structured signal that must be interpreted within its coverage, data quality and category context.

The Index value is composed of two measures of recorded market activity in a rolling window: median sale price and transaction volume. It is read with:

• auction liquidity and price visibility across the houses and segments it covers;
• category, period and region context;
• coverage, confidence level and data completeness, carried with every value.

Provenance, literature and condition are recorded where a catalogue publishes them, and they are part of the framework. They are not yet among the inputs a score is computed from.

These signals are organised into a proprietary framework inspired by the way specialists assess artworks, markets and collecting contexts.

The purpose is not to replace expert judgement, but to make the first layers of analysis more legible.

The Artmetria Index value rebases to 100 on a per-vertical anchor date, then tracks recorded market activity: how prices move, and how much is transacted. The public site prints no value while the Index is being finished (see the Index page); inside Market a value never travels without its coverage and its band.

Market motion comes from the median sale price and transaction volume (on a log scale). Since 20 September 2026 those are the only two components. The qualitative reading layer used to enter as a third term; it was removed, and the section below says exactly why and what it cost.

Why median, not average: a single 100 M€ outlier should not drive a vertical's index. The median is robust to extreme values by construction.

Why log volume: raw transaction counts grow exponentially with ingestion coverage. The logarithm smooths volume bursts while preserving directional signal.

Why there was a quality term, and why there is not one now: a market can be cheap (low median) but composed of premium lots, or expensive but composed of entry-level pieces, and folding a qualitative reading into the value captures that mix. That reasoning still holds. What did not hold was the input — see the series break of 20 September 2026 below.

Baseline per vertical: each vertical anchors on its own densest credible historical window; niche sub-verticals with thin historical coverage anchor on a more recent window.

The two figures shown next to the value (↗ +X.X% 90d and ↗ +X.X% 1y) are real rolling deltas — not cumulative motion versus the vertical baseline.

For each delta, the engine computes the index value at the comparison date (90 or 365 days earlier) using its own 90-day window, then takes the percentage change. Both endpoints share the same baseline so the delta isolates the directional motion.

Deltas are suppressed below a minimum threshold of sold lots in the reference window — at that point the rolling comparison is noise.

Three primitives keep the displayed value honest about its own sampling error and composition.

The quality reading is an unweighted mean. Each scored lot in the window counts once, whatever it sold for. A lot with no score is left out of the calculation rather than entered as a zero. Since 20 September 2026 this reading is computed and kept, and enters nothing that is published.

Series breaks. On 14 September 2026 the quality factor stopped being weighted by estimates; on 20 September 2026 the quality term was removed, because the scores it rested on did not read the works. The value is now price (0.60) and volume (0.40). Values before each date are not comparable with those after.

Whether a quality term returns, and in which form, is not decided.

Confidence interval on the value. Each snapshot carries a 95 % confidence interval (the ± X.X shown next to the value), derived from an asymptotic formula on the median. The interval is wider on small samples and tightens as coverage grows.

Adaptive window. When the rolling 90-day window holds too few sold lots — the empirical floor below which median + volume become unstable — the engine automatically widens the window and states the change beside the value.

Coverage, here, is not a claim about how much of the art market Artmetria sees. It is a measurement of how much recorded evidence sits behind one published figure.

Each Index reading is computed over a rolling window of recorded sales — 90 days by default — and expressed against a reference window anchored in 2008. The number of sales that fall inside that window is what the coverage state describes. It is counted per market, not for the platform as a whole, so two readings shown side by side may rest on very different amounts of evidence.

Coverage being sufficient does not make a reading correct. It means the reading was computed on enough records to be worth arguing with.

Every Index reading carries one of three states, decided by how many recorded sales fall inside its window. The thresholds are fixed, and they are the same ones the platform applies internally:

• Canonical — 1,000 recorded sales or more in the window. The reading rests on a dense record.
• Partial — at least 100 sales, but fewer than 1,000. The reading is shown and should be treated as directional; precision improves as the record deepens.
• Insufficient — fewer than 100 sales, or an empty baseline (no sold lot, or no median price). No Index value is displayed at all.

The last state is the one that matters most. Where the record is too thin to read, Artmetria does not publish a weaker number with a warning attached — it publishes nothing, and says why. A missing value here is a measurement, not an outage.

A canonical state does not mean perfect coverage. It means the segment is sufficiently recorded to be used responsibly.

Provisional value — no longer produced. This label meant the baseline had price and volume signal but no quality scores yet. Since 20 September 2026 the value carries no quality term at all, so the label has nothing left to qualify. It still appears on snapshots taken before that date, where it describes the formula in force at the time.

Separately from its coverage state, each reading carries a confidence band derived from the same sample size:

• High — 1,000 recorded sales or more.
• Medium — 200 to 999.
• Low — 50 to 199. Read as indicative.
• Thin — fewer than 50. The window is widened before the value is shown at all.

A 95% confidence interval is shown alongside the value itself, inside Market, derived from the dispersion of prices in the window rather than assumed. A wide interval on a thin window is the honest output, and it is shown rather than rounded away.

Two adjustments exist because a sale calendar is uneven, and neither is silent.

The window can widen. When a 90-day window holds too few recorded sales to read honestly, the observation window is extended to 180 days rather than publishing a value computed on almost nothing. Thin sub-segments — a single period, a single region — reach that condition routinely. The reading records that its window was widened.

Coverage itself can move. A change between two periods can come from the market or from how much of it was recorded. When the number of sales in the current window differs by more than 30% from the comparison window, the reading is flagged for coverage drift — measured both against 90 days earlier and against the same period a year before. The flag does not correct the figure. It tells the reader that the variation may be describing the record rather than the market, which is a different statement and should not be read as the same one.

The blended global index averages markets that don't behave alike. Three regional sub-indices isolate the motion :

• Americas — sales held in US salesrooms (New York, Los Angeles, Philadelphia, Chicago, Dallas, Boston, Miami) + Canada.
• Europe — UK rooms (London, Edinburgh), continental Europe (Paris, Monaco, Geneva, Zurich, Cologne, Munich, Berlin, Vienna, Milan, Rome).
• Asia — Hong Kong, Tokyo, Beijing, Shanghai, Seoul, Taipei, Mumbai, Singapore.

Regional readings are not published: the current version assigns a sale to its house's country.

Aggregators count as one name. Some entries are aggregators that group several houses under one name.

Coverage caveat. The regional mix reflects only the houses currently indexed, and their weight changes as records are corrected: on 25 September 2026 one house's prices, which had stored the buyer's total instead of the hammer, were withdrawn from the record pending re-ingestion. The place-of-sale mapping is more complete on recent sales than on the historical archive, so the geographic readings are thinner before 2024. These are reasons the regional readings stay unpublished until the next index version, not footnotes to a published one.

Artmetria uses a C1–C10 framework to organise the qualitative reading of an object across ten named dimensions: artist identity, period, quality of execution, subject, composition, technique and support, format, condition, provenance, and bibliography and critical resonance.

It is one instrument inside a larger layer, not the whole of Artmetria's intelligence and not a market model: market data describes market behaviour and enters the Index through its market layer, never through the criteria. What the framework reads is the object as its record documents it.

The C1–C10 grid is the operative reading framework behind the qualitative layer: a credible way to break a work into interpretable components rather than compressing it into a single opaque number. Which of the ten a given reading scores depends on what the catalogue recorded; where the record is too thin, the reading abstains rather than fills in. The framework is public and the weighting is not, and the configuration of criteria a published reading draws on is itself versioned: a change to it carries its own version and date, as the series break above does.

C1–C10 should be understood as a reading grid. It helps break down an artwork or market signal into interpretable components, rather than compressing everything into a single opaque number.

Ten criteria, named. You can hold a work against them yourself, and disagree with a reading on a specific ground rather than in general.

1. C1 — Artist identity
2. C2 — Period / dating
3. C3 — Quality of execution
4. C4 — Subject / iconography
5. C5 — Composition
6. C6 — Technique / support
7. C7 — Format / scale
8. C8 — Condition
9. C9 — Provenance
10. C10 — Bibliography & critical resonance

Named is not the same as scored. Which of the ten a given reading scores depends on what the catalogue published about that object: where the evidence is missing the criterion is left unscored rather than filled in, and below a minimum of scored criteria the Index publishes no composite at all. Condition, provenance and bibliography are recorded where a catalogue publishes them, and they are not yet among the inputs a score is computed from.

Artmetria distinguishes between market-level signals and object-level signals.

Market-level signals may relate to activity, liquidity, auction history, regional presence, sale frequency, price visibility or institutional/market attention.

Object-level signals may relate to the characteristics of a specific artwork, such as medium, format, period, subject, condition, provenance, exhibition history or literature references when available.

This distinction matters because a strong artist market does not automatically make every artwork equally strong, and a compelling object may exist within a market segment that remains thin, emerging or difficult to read.

Every analytical signal depends on the quality and depth of the underlying data.

When Artmetria has stronger coverage, the interpretation can be more stable. When coverage is partial, the interpretation must remain more cautious.

A score is accompanied by confidence indicators, data-quality notes, coverage warnings or source-context explanations, as shown above.

This is a core part of the methodology: uncertainty should be made visible rather than hidden.

An Artmetria score is not a certificate, appraisal, authentication, investment recommendation or guarantee of future value.

It is an analytical signal produced from available information and methodological assumptions.

Scores may help users compare, question, prioritise or contextualise. They should not be used alone to make acquisition, sale, legal, tax or financial decisions.

Artmetria is designed to support judgement, not replace it.

Art markets are affected by taste, scarcity, provenance, macroeconomic conditions, institutional narratives, collector behaviour, geography, liquidity and timing.

For this reason, Artmetria does not present its Index as a price-prediction engine.

The purpose is to make signals more readable, not to claim certainty about future performance.

The platform favours structured interpretation over speculative forecasting.

The methodology is designed to evolve.

As Artmetria expands coverage, improves normalisation, refines scoring layers and adds more contextual signals, the Index can become more precise and more useful.

This evolution does not mean that earlier outputs were “wrong”; it means that the platform is built as a progressive analytical system.

What is published today should therefore be read as the first public expression of a method, not as a finished authority.

Artmetria is intended for educational, analytical and research-oriented use.

Users should treat the Index as one layer among others: direct viewing, expert advice, provenance research, condition review, legal checks, tax considerations and independent judgement all remain essential.

The Index helps ask better questions. It does not remove the need to ask them.