This story is over 5 years old.


​Artificial Intelligence’s Latest Trick: Predicting the Price of Wine

A former hedge fund algo trader is trying to bring machine learning techniques from research to the real world.

Tools from the field of artificial intelligence research have made their way into all kinds of useful (and not-so-useful) real-world applications, like solving centuries-old science problems or designing Mario levels. Here's a proposition for another, which could arguably fall into either camp depending on whether you happen to be into trading niche investments: predicting the price of fine wines.

Tristan Fletcher is an AI researcher at University College London and Imperial College who cofounded a company called Invinio with the aim of advising people on managing their wine portfolios. Because some people buy wine as an investment, not just a drink.


"I used to be an algo trader in a hedge fund," Fletcher told me in a phone call to explain where his interest started. "Because I was working there I wasn't allowed to speculate in anything conventional as part of my contract, but I wanted to do some trading in other things and I found out that wine was something that could be traded on exchanges."

Fletcher said he'd previously tried to use techniques from artificial intelligence to predict—"unfortunately not with great success"—which direction financial markets would move.

Despite that experience, he was still keen to bring AI into practical applications elsewhere. In addition to his work in research and with Invinio, he is head of machine learning at FinTech startup Thought Machine.

"Using machine learning techniques in the field of fine wine price prediction is completely unprecedented."

The wine-predicting work is detailed in a paper Fletcher co-authored with other UCL researchers (and with the assistance of Invinio), published in the rather niche Journal of Wine Economics. The researchers found that the machine learning algorithms they applied predicted the past price fluctuations of 100 fine wines with greater accuracy than the simpler, more traditional methods.

And no, apparently it's not as simple as suggesting that wines are worth more the older they get.

One technique they used is called Gaussian processes, and Fletcher explained it as expanding on the simple prediction methods you might envisage—like drawing a straight line on a graph to simply continue a consistent trend into the future—by "allowing that trend to exist in strange spaces." Imagine you drew that line on a sheet of paper, but then that paper got folded up in weird ways.


In the paper, the authors conclude that "using machine learning techniques in the field of fine wine price prediction is completely unprecedented and our results point to a huge potential of applying more of such machine learning research to the field."

Fletcher said that another advantage of using machine learning algorithms is that you can get information beyond a prediction, such as "confidence bounds" on that prediction (basically, estimating how likely it is). That could be useful if you were actually looking to trade.

However, Fletcher pointed out that actually trading wine isn't very efficient because of the high transaction costs; it's more of a long-term investment. "At the moment, realistically, trading wines with any frequency is insane," he said.

But despite limits on immediate practical applications, one of his main motivations was to show how people can use these AI techniques in unusual ways.

Continuing the collaboration with UCL and his company Invinio, Fletcher's hoping to apply similar techniques to other items of investment, such as spirits and classic cars.

"What I'm trying to do is take these esoteric things that come out of the academic community in the world of artificial intelligence and try and put them in the real domain," he said. "I know predicting wine prices is pretty pedestrian and won't win any Nobel Prizes, but it's a step in the right direction."