Curiosity about the limits of machine learning led former trader, UCL academic and startup founder, Dr Tristan Fletcher, to apply complex AI techniques to the — on the surface — rather chaotic arena of fine wine pricing, comparing them with trading techniques used for more typical asset classes.
“Prices of wines are all over the place. You can get one wine sold at an auction say in London, and the same wine might get sold a few days later in Hong Kong and the price differences would be humongous,” says Fletcher. “So I realized it was a really inefficient market and I might be able to profit from those inefficiencies.”
Trading wine is expensive because of the price per unit traders have to pay. It’s not a market set up for high frequency trading, as with shares. Rather wine lovers typically buy a couple of cases, drink one (slowly), and then sell the second one after enough time has elapsed as a way to fund drinking the first. It’s trading to drink, not to profit.
So while Fletcher’s machine learning approach to predicting fine wine pricing turned out to be able to more accurately forecast prices than other more traditional trading methods, the research that led to it is perhaps more of an academic exercise in testing the boundaries of machine learning than a technique with widespread commercial application, owing to the complexities involved.
That said, Fletcher does also run a quantitative wine asset management startup, called Invinio, which collaborated on the UCL research and intends to continue working with the university to help refine the algorithms — and potentially use some of the research to improve the tools it provides to wine investors. So there are some immediate commercial applications, too.
There is also potential for applying similar machine learning techniques to other alternative asset classes that people trade, reckons Fletcher, such as classic cars, spirits, rare books and even perhaps fine art. Although, spirits excepted, each of those other markets has very different characteristics to fine wine — so there’s unlikely to be uniform results.
A study detailing the fine wine research has been published today in the Journal of Wine Economics. It’s co-authored by Fletcher, along with professor John Shawe-Taylor, co-director of the UCL Centre for Computational Statistics & Machine Learning and Head of UCL Computer Science. The primary author is UCL MSc graduate, Michelle Yeo.
“I first started collecting wine data a few years ago,” says Fletcher, who has a background in AI research, explaining how the project came about. “I used to work as an algorithmic trader at a hedge fund and I wasn’t allowed to trade normal assets, and wanted to trade something that you could still buy and sell on exchanges, because that’s what I knew about, and wine was something that the Financial Conduct Authority are happy for people to trade without any registration.”
He then wondered what he could do with the fine wine pricing data he had amassed, and put a call out at UCL for research students who were interested in working with him on applying AI-techniques to the data — which is how Yeo got involved. “Can we try and predict which way the wine prices would move?” was the core aim of the study.
Specifically the team looked at data for 100 of the most sought-after fine wines from the Liv-ex 100 wine index. They tested two forms of machine learning on the data-set — including ‘Gaussian process regression’; and the more complex ‘multi-task feature learning’, the latter in a bid to extract the most relevant information from a variety of sources.
Looking at the data the team found the wines fell broadly into two groups — with about half showing evidence of strong negative autocorrelations for one day; so if they went up one day, the price would go down the next. And the other group not showing that characteristic.
For the wines with the negative autocorrelations Fletcher says the team was able to get “quite significance outperformance” using Gaussian techniques (vs more standard trading metrics) to predict whether the price would go up or down the next day. For the other group of wines the technique was less successful but, according to Fletcher, it was still better than more traditional techniques.
Yeo also applied the even more complex multi-task learning technique to the data in an attempt to predict next day wine prices — not just up or down moves.
“This multi-task learning was able to find which prices in a time series — in a long list of prices being paid — were useful; extract that information and then predict what the price was going to be on the next day. It was a technique that turns out to be really useful for this group of wines where the autocorrelation was negative,” says Fletcher.
“We’ve taken these two machine learning techniques — Gaussian processes has been around for 5-10 years, but is only now getting really used, whereas multi-task learning has been around for less time and is not used that much anywhere — and both of them showed a lot of outperformance for some of the wines. Both in predicting whether the wine was going to go up or down in price, but also how much so.”
The big question is, is the complexity of these techniques justified? Would you use these things in the real world?
“I think the big question is, is the complexity of these techniques justified? Would you use these things in the real world?” he adds.
The complexity of the techniques means the necessary calculations would take too much time to be applicable in a high frequency trading scenario. But that’s not something which applies to the fine wine investment world, so Fletcher believes there could be worth in applying the research to extend the services Invinio offers, in the first instance.
“We’ve got this company Invinio where we’re trying to create these portfolios of fine wines and we might start incorporating this research to give people a vague idea about if they were going to buy a wine — recommended for diversification purposes — is today the best day to buy it or should you wait a few days” he says.
That said, it’s still only a small advantage in a small, low-frequency trading market gained via some very complex math. And Fletcher argues there are plenty of trading scenarios where complex AI techniques aren’t suitable, or where accuracy gains are small vs computation efforts required.
“My view is that a lot of the time the complexity of using these techniques isn’t justified in the performance they bring you,” he says. “They sound good and a lot of places like investment banks and hedge funds like the PR connected with saying they use machine learning methods but they might be doing something a lot more pedestrian in reality.”
“[With the fine wine research] we’ve got a big edge on these old traditional techniques, but the reality is we’re not going to have huge accuracy on predicting when stuff goes up or down — if we did we’d be way ahead of any other market. No one in any market really knows whether things are going to go up or down over any sensible time-period. It’s a very hard thing to do, and wine is no different.”
He argues generally that AI works best as an “augmentative intelligence” — supplementing and helping human operators, rather than replacing human activity entirely.
“AI can help you do a lot of things, automate a lot of things, it can make you a lot more effective but you will always need a human being for those unforeseen circumstances, and that’s true in the world of trading, just as much as it is in many other aspects of our lives where people are worried that the robots are taking over. I think there’s always going to be room for people,” he adds.
“You can’t completely replace [human] traders. As much as many people would like to because they’re expensive and seen as not having much value… but you do need people who can think laterally. A program… can’t conceive of things that it hasn’t witnessed before. It can’t deal with the unknown unknowns. But people can.”