Quants and quaffs

THE term “alternative assets” encompasses all manner of offbeat investments. Philatelists delight in rare stamps; petrol-heads in classic cars; oenophiles in that most liquid of assets, fine wine. The wine futures market, though, is pretty inefficient. Prices hinge on tastings of stuff that is still in the barrel, long before it reaches its fullest bloom.

This en primeur pricing is set on the basis of experts’ palates, not through the equations of quants. Tristan Fletcher, of University College, London, is among those who wish to change that, by applying artificial intelligence to the matter.

Previous attempts to tame the fickle wine market with mathematics have relied on linear regression models. These take the untidy spray of data points about a given vintage—the particulars of the weather that year, the vineyard’s history of medallion-winning and so on—and use them to draw the straight line that has, over the course of time, most closely approximated the price. Pick the point on this line where a particular vintage lies, and out comes a price prediction.

Such efforts have produced mixed results, however, and Dr Fletcher thought he could do better. Instead of regression, he applied a form of artificial intelligence, well known among prognosticators of other asset classes, called machine learning. This is able to ferret out correlations (perhaps a great many of them, some weak or transient) that standard regression models gloss over. Rather than a simple straight line, the result is a price curve that snakes through the data, thus yielding, if the particulars of the calculation have been set up properly, stronger predictions than regressions can manage.

Dr Fletcher and his colleagues started with wines in the Liv-ex 100 (a kind of fine-wine FTSE) and looked only at the historical data on prices. They first ran an “autocorrelation test”—a way to quantify how calm or unruly prices had been in the past. As they report this week in the Journal of Wine Economics, they found two distinct groups. Half of the wines they looked at seemed to fluctuate in price over short periods, settling towards the mean return quickly. The other half trended up and down more wildly, over longer periods.

The team then ran two types of machine-learning algorithm on each group. For wines in the calmer group (in which the market was, presumably, behaving more efficiently), these algorithms outperformed the regression method by only a little. For those with more untamed price histories, though—ie, the ones on which most money could, in principle, be made—machine learning roundly won. That the wines should fall into two groups, for which the technique’s benefits differ, is curious. It is, of course, a matter for further study—perhaps over a bottle of Château Palmer.