Methodology

For any artwork at any point in time, the final LAE is the sum of these three components. The result is a current estimated price with a confidence range — and, because the components are time-aware, an estimate for any historical point.

Most artworks transact once or not at all. A model that needs many observations per artwork has nothing to learn from. Pairs analysis on repeat sales sidesteps this.

Artist significance, retrospective inclusion, condition nuance — these matter to price but aren't in the data. Differencing techniques cancel out these latent factors when comparing repeat sales.

A Basquiat painting and a Basquiat drawing don't trend together. A linear model can't capture that. Tree-based models can — but only if the question they're solving is narrow enough.

Where we report year-by-year returns — such as an artist’s Annual Returns — we use the Compound Annual Return (CAR): the annualized, compounded rate of return realized across repeat sales of the same work. CAR is the same calculation as the compound annual growth rate (CAGR); we say “return” because it is measured from actual buy–sell pairs. Because it compounds rather than averaging each year’s change, it neither overstates nor understates performance across volatile years.

For an artist — or, for a held work, that work's artist cohort — the score reads a five-year window of auction activity together with the artist's trailing momentum: sell-through rate, buy-in rate, the split of lots selling above / within / below estimate, the sample size (lots offered), and 12-month price momentum. At the holding level it also reads compound annual return and the position's weight in the portfolio. Missing inputs lower the confidence rather than breaking the score.

The demand-and-liquidity axis is a weighted blend of whatever signals are present, re-normalised by the weights that actually fire: sell-through (weight 0.5, mapped 50%→90%), share sold above estimate (0.4, mapped 20%→60%), a below-estimate penalty (0.2), and a buy-in penalty (0.3, mapped 10%→40%). The question it answers: can this work be sold, and is demand beating the trade's own pricing?

Twelve-month momentum sets the trend: rising above +5%, falling below −5%, otherwise flat. A separate “peaking” flag fires above +25% — the recent-run-up-equals-correction-risk signal from the art-finance literature.

Market strength is crossed with trend, deliberately, so that “sell” and “watch” are not the same recommendation:

• Weak market (below 45) → Watch — illiquid; exit is hard, so monitor rather than act.
• Strong market (65+), peaking or falling → Sell — realise value while the work is still liquid.
• Strong market, rising → Accumulate.
• Strong market, flat → Hold.
• Moderate market (45–64): falling → Watch (don't sell into weakness); peaking → Sell; otherwise → Hold.

The crucial distinction: a strong but falling market says Sell (exit while you can); a moderate, falling market says Watch (selling now would be into weakness). That is why a deep, liquid name with flat momentum reads Hold, while a thinner one with falling momentum reads Watch.

The 0–100 composite used to rank holdings is an equal blend of market strength and a momentum-derived trend score (50 at flat, 100 at +20%, 0 at −20%), less a penalty for positions above 25% of portfolio value. Confidence is High when market data and momentum are both present over at least 30 lots, Medium above 10 lots, otherwise Low.

Every rule that fires emits a structured reason — a code, a severity, and the value that triggered it — rendered in plain language and sorted so cautions appear first. Because the thresholds are fixed and published, the “why” is always reconstructable from the inputs; nothing is generated by a language model. An optional AI-written narrative may be layered on top in future, but never in place of the rules.

The thresholds are a defensible first pass, not yet back-tested against realised sale outcomes. The artist-level read uses the artist's whole market; the per-work and per-holding versions narrow to the work's cohort and add return and concentration. The Disposition Score is a market read for professionals, not personalised investment advice.

Mean absolute error reported by artist tier, price bucket, medium, and region — not just a global headline number.

We check that an ±8% range actually contains roughly 80% of realized prices. Miscalibrated intervals are worse than wide ones.

Each model carries a version. Material changes ship with a changelog noting what shifted and why. Enterprise clients receive segment-level performance reports.