How do AI-driven art valuations actually work?
Companies like LiveArt and Limna draw on a wide range of data – from previous sales to ‘cultural momentum’ – to predict the price of art with increasing sophistication. By Jo Lawson-Tancred
The opacity of the art market has long been considered a key reason why it has resisted the same data-driven analyses that have revolutionised other sectors. The art dealers who orchestrate this mysterious ecosystem of private sales may have been quite happy for the market to remain a world unto its own, but times are inevitably changing with the rapid rise of online marketplaces and ever more sophisticated AI.
Leading the charge in this new sphere is LiveArt, an ‘AI-powered data platform’ co-founded by Adam Chinn, former chief operating officer at Sotheby’s, John Auerbach, former head of e-commerce at Sotheby’s, and tech entrepreneur Boris Pevzner, who founded the digital collection management system Collectrium (sold to Christie’s in 2015). Positioning itself as a bastion of Web3 principles, LiveArt is eagerly disrupting the traditional gallery system by facilitating anonymous peer-to-peer online trading of artworks priced from around $20,000 to $3 million.
The platform’s shameless treatment of fine art as an asset class may ruffle feathers, but it has enticed would-be traders with a growing package of market analysis tools. For each artist, an almost perplexing array of indices and price history can be conjured in moments, presumably to help seasoned stock traders feel more at home. One service that sticks out is the LiveArt Estimate, which offers buyers and sellers an AI prediction of the approximate value range of an artwork. ‘Two people who don’t know each other want to transact an object that is unique,’ explains Pevzner to Critical Edge. ‘What is its current market value? What is the range that they should be confident negotiating around?’ In this Web3 context, AI approximates the trusted expertise of a dealer, or adviser, within the old centralised structure.
Some aspects of the art market will always be unpredictable to any form of intelligence, human or artificial. Pevzner is emphatic that the tool is just an indicator and not built to predict real auction prices – still a slippery objective owing to the indecipherable role of marketing or ‘hype’ in propelling some lots far beyond their estimates. ‘What’s most important is that our customers are satisfied with the predictions,’ Pevzner says of the company’s ultimate test. ‘If they end up doing a deal close to our suggestion then that’s a success.’
Unsurprisingly, Pevzner explains that this process works best for work priced in the data-rich mid-range. The AI’s estimate is calculated using a long list of data points that fall roughly into the following categories: artwork attributes such as size; artist attributes such as exhibition frequency; sales attributes such as how many times the work has been sold and where; and statistical features, or previously unknowable patterns in the data that have been detected and used by the model. Finally, external economic events also have a small part to play. ‘Even though people say that the art market is not correlated with the equities market, we find that there is some correlation,’ says Pevzner. ‘It doesn’t matter what the correlation is; the machine learning model figures that out once we feed various market indices into it.’
For the top 300 or so ‘high liquidity’ artists – those most frequently traded – the auction record alone is rich enough to generate apparently reliable estimates, so the model prioritises these figures. ‘Nothing beats sales data in terms of predictions,’ confirms Pevzner. ‘It’s only if you don’t have sales data that you have to rely on other stuff.’ One such case is the category of emerging artists, which accounts for about a third of sales on the platform. Here, quantitative data is necessarily buoyed by the measure of ‘cultural momentum’ – qualitative attributes relating to the artist such as exhibition history ranked by gallery venue, critical standing, and social media engagement.
A similar promise of being able to generate price predictions for emerging artists has been made by the phone app Limna. Limna charts cultural momentum according to extensive exhibition data collected by co-founder Marek Claaseen over the years since he founded the online database Artfacts.net in 2001. So far, it’s hard for an outsider to assess the accuracy of either company’s claims, but if cultural momentum does prove to be a strong indicator of market performance then they may both be getting a head start in the race to amass and make use of this data.
Understanding the patterns of artists’ trajectories from emerging to renowned means we can compare those who appear to be on similar paths and potentially predict where an early-career artist is likely to end up. LiveArt’s clusters of ‘similar’ artists tend to be determined by the model according to customer interest, disregarding traditional art-historical categorisations. ‘When we let the algorithm see what people actually buy together, there has been some divergence from the art-historical view – and this divergence is really what traders are after, right?’ says Pevzner of these newly emerging statistical measures of similarity. ‘What else is trending in a similar way? That is more useful than academic data.’
Correctly identifying and investing in tomorrow’s rising stars would be akin to winning the jackpot. There may not be much empirical evidence of these artists’ potential – they’ve likely not had the chance to build a reputation on the secondary market yet – but if platforms like LiveArt and Limna could win users’ trust in their predictions they would presumably be able to excite lucrative speculation.
Mapping out the future for still emerging artists might, however, feel a little presumptuous or even limiting. There’s the risk that their destinies are prematurely foretold by the data, as the model accelerates the popularity of some artists versus others due to factors over which they have little to no control. Think, for example, of the network scientist Albert-László Barabási’s The Art Network (2018), a graph showing how early exhibition venues determine the success of an artist’s career. But Pevzner insists that the pricing for emerging artists will, by the time they appear on the platform, already have been decided by the gallery system, with its own complex web of reasonings and more socially influenced dynamics. This may be true, but surely doesn’t rule out the possibility that data-driven analysis will have an outsized impact on the course of an artist’s career, with their initial gallery representation serving as a key data point for the model’s predictions.
Another key advantage for online marketplaces dabbling in AI is their access to primary sales data, formerly the preserve of tight-lipped gallerists. The short supply of this private intel has long stalled attempts to develop more sophisticated machine learning models, but companies like LiveArt may now be gaining a competitive edge by harvesting anonymous primary sales data from the transactions hosted on their sites. In a similar vein, Limna encourages users to enter into its app the prices they have been quoted in conversation with dealers, presumably in confidence. Over the coming years this kind of data should significantly improve overall accuracy of predictions, with the most significant advancements seen for emerging artists.
Historically, the elusive logic of pricing art has been strategically deployed by the old-fashioned art market. AI-generated valuations promise a more empirical underpinning, which will put at ease the growing number of online investors who are snapping up fine art as the latest fashionable asset. So far, these algorithms have been of little consequence to the wider art ecosystem but, as they get more sophisticated, it may be time to start paying attention.
Jo Lawson-Tancred is a freelance art writer living in London.