How LiveArt's AI actually works.
We believe the art market deserves AI that is both powerful and transparent. This page describes how LiveArt Estimate™ and the broader AI stack are built — honestly, including what they don't do.
Auction records · 1986–today
Features per artwork
Specialized models · one estimate
Validation · year N → N+1
Three questions. Three models.
LiveArt Estimate™ answers one question — what is this artwork worth right now? — by separating it into three distinct sub-questions, each answered by a different model. The three components compose into a single price with a confidence range.
01
What is this type of work worth, in the abstract?
02
How does this specific type of work drift in value over time?
03
How is the art market as a whole doing?
◆ COMPONENT 3 · MARKET
Market trend
Repeat-sales regressionTracks pure market cycles by analyzing works that have sold more than once. Differencing pairs of sales isolates market movement from artwork-specific drift.
◆ COMPONENT 2 · ARTWORK
Artwork-specific trend
Gradient-boosted decision treesCaptures how specific work categories — by artist, medium, period, size — drift relative to the overall market. Tree-based models capture interaction effects a single regression would miss.
◆ COMPONENT 1 · BASE
Base price
XGBoost regressionThe intrinsic value of an artwork independent of market timing. Trained on 100+ features per artwork: artist, medium, size, period, provenance, and more.
One artwork. One estimate.
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.
Why three models, not one.
The obvious question is why not use one big model. The answer comes down to three properties of the art market that a single model handles poorly.
01
Data is sparse per artwork
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.
02
Quality is unobserved
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.
03
Interactions dominate
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.
◇ Signal leak — where one model's output contaminates another's training data — is the most common failure mode in art-market AI. Separating questions enforces clean inputs.
A complete price history,
not just a point estimate.
Because LAE is built on a time-aware architecture — the three components each carry time information — the model can produce an estimated price for any artwork at any point in its history. That makes portfolio analytics possible.
◆ PRO FORMA RETURNS
Returns on hypothetical holding periods. Pick any two dates, get a return — at the artwork, artist, or portfolio level.
◆ INDEX BENCHMARKS
Compare against LiveArt indices, blue-chip cohorts, or traditional benchmarks (S&P 500, bonds, gold). Same currency, same period.
◆ PORTFOLIO ANALYTICS
Sharpe ratios, drawdowns, correlation matrices, volatility — derived from continuous price series, not just confirmed sales.
What the model doesn't claim to do.
Every model has limits. We publish ours so consumers of LAE can use it appropriately.
01
LAE works best for liquid artists.
Most accurate for artists with sustained auction activity — typically the top 500–1,000 artists by transaction volume. Below that, confidence ranges widen accordingly.
02
Emerging artists are hard.
Markets in rapid expansion show lagging predictions. Historical data alone is a weak predictor of current value when an artist's market is reshaping in real time.
03
Primary market is not in scope.
Gallery and private sale prices are not in the training data. For living artists where the primary market dominates, LAE reflects auction signal only.
04
Auction noise is partially filtered.
Manipulated sales, guarantees, and buy-ins introduce noise that no model perfectly removes. We filter what we can and surface confidence ranges as a reliability indicator.
05
LAE is a starting point, not a final answer.
The model augments — it does not replace the specialist, the appraiser, or the advisor. Confidence ranges exist precisely because a single number is rarely the right answer.
What else the AI does.
LAE is the headline output. Other components of the AI stack run on the same data foundation.
Price momentum
Repeat-sales-filtered 12-month signal at the artist or category level.
Artist embeddings
64-dimensional vectors enabling similarity comparison and clustering across 350K+ artists.
Similarity vectors
Comparable-artwork retrieval based on visual and metadata features — the workhorse behind cataloguing and search.
Market signals
Real-time structured signals from auction calendars, results, and corrections.
Image recognition
Cataloguing workflows: artist attribution, medium detection, edition matching from photos.
Historical LAE
Time-machine estimates for portfolio reconstruction, attribution analysis, and academic research.
How we know the model is working.
Validation is the part most AI vendors quietly skip. Here's the approach.
We train on data through year N and test on year N+1. The sequence: train on 2022, test on 2023. Train on 2023, test on 2024. Train on 2024, test on 2025. This prevents the model from interpolating between known points — a common form of cheating in time-series ML.
01
Stratified error reporting
Mean absolute error reported by artist tier, price bucket, medium, and region — not just a global headline number.
02
Calibrated confidence intervals
We check that an ±8% range actually contains roughly 80% of realized prices. Miscalibrated intervals are worse than wide ones.
03
Versioned models, published changelogs
Each model carries a version. Material changes ship with a changelog noting what shifted and why. Enterprise clients receive segment-level performance reports.
The principles behind the methodology.
01
Transparency over mystique.
Published methodology. Visible confidence ranges. No black boxes for prestige.
02
Augment experts, don't replace them.
The model supports specialist judgment. It is a starting point, not a verdict.
03
Purpose-built models, not one giant model.
Each question gets the model that fits it. We resist the urge to throw everything into a single architecture.
04
Continuous validation.
The market changes. The model is retrained, re-tested, and reported on a regular cadence.
Questions about the methodology?
Talk to the team.
Our engineering team is available to discuss architecture, validation approach, and model performance in detail. Schedule a session for your quants or research desk.
TALK TO ENGINEERING →