What do wine snobs, and Chinese anti-corruption laws have to do with failed statistical arbitrage? Just by reading the previous line aloud I am sure most of the readers already have either fallen sleep or clicked out of this blog.

I have written in the past about the perils of finding spurious correlations in Finance, and particularly about the inability of machine learning algorithms (and some very smart people) to identify the causal model that drives the prices of financial assets. Below we can see a typical model-breaking example, but with tastier assets than subprime mortgage backed derivates.

Look at the graph below: between Jun 2006 and April 2009 the prices of Clerc Milon and Duhart Milon 2005 were ‘co-integrated’ (they pretty much moved in tandem) — however from April 2009 to September 2011 the prices diverged enormously!

Those wines are the typical ones that only wine snobs would know about — they are neither the record breaking ultra expensive premier crus nor the cheap 2-buck Chuck from Trader Joe’s. Hence, people buy them for their ‘taste’.

Look at Google Maps: Clerc Milon is just accross the street from Duhart Milon and Lafite Rothschild (you can bike around it in about 10 minutes). The physical closeness pretty much removes any argument of ‘terroir’ differences between the wines. Wine bores who use Parker tasting’s as reference can point out that in a 2006 tasting the Clerc Milon was scored at a 90-92 range (released at £270) just the same as the Duhart Milon (released at £240). Real wine geeks go wild for 100 points, so a ‘mere’ 92 is not good enough to drive the price up so wildly.

Full ratings around April 2006 for the 2005 vintage en-primeur sale session:

  • Duhart Milon JR: 17.5. WS: 92–94. RP: 90–92. MJ: 17.5. GD: 90–92.
  • Clerc Milon — JR: 17.5. WS: 92–94. RP: 90–92. MJ: 17. GD: 89–92.

By 2006 there was no discernible difference between the wines — if anything, Clerc Milon was able to release their wine at a higher price.

The clue lies in the full name of Duhart-Milon Rothschild:

Notice the (Lafite)

In the label, you can see the full name of the Chateau: ‘Domains Barons de Rothschild (Lafite)”. For many reasons (including some urban legends that claim that Lafite is easy to pronounce — read this article) the Chinese Market loved buying Lafite premier cru. As Lafite prices went up, the bubble propagated to the rest of Rothschild-Lafite wines (see the graph below: Lafite started rising in the middle of 2007, but Duhart Milon lagged almost one year). In mathematical geek speak, we have a price ‘confounder’:

A confounder is a

is a variable that influences both the dependent variable and independent variable causing a spurious association.

And it is quite possible this confounder killed 1855.com!

1855.com was a French company that deals with the wine ‘en-primeur’ market. When your annoying wine bore buys wine ‘en-primeur’ he/she is paying for wine that is not bottled yet. The Bordeaux Chateaux do not want to handle the retail paperwork, so many intermediaries exist to market the wine, handle the payments and future delivery (even storage) — collectively these apparatus is known as ‘Place de Bordeaux’.

If you are a financial professional, you will immediately notice that such an intermediary can ‘short-sell’ wine: sell the wine to the retail customer, keep the money and buy the wine only when delivery is required.

Statistical arbitrage requires you to sell the ‘expensive’ asset and buy the ‘cheap’ one: the clear trade above is to sell Duhart Milon and buy Clerc Milon. If your model captures the ‘fundamental value’ of wine (terroir, vinification, taste) you can think you are ahead of the game.

And 1855.com did seem to think in this terms. Looking at the history pre-2006, I am pretty sure their statistical analysis showed them that ‘en primeur’ was regularly overpriced and that they could buy back cheaper: 1855 Le pyramide du ponzi du vin (in French):

“[the 1855.com founder] believed he could buy cheaper wine than in the ‘Place de Bordeaux’’

In fact, if they had sold the Duhart Milon 2010 en primeur and bought it back in early 2015 they would have gained about 40% (the percentage terms that the price fell). As usual, it was all about timing; 1855.com had run out of money by 2013. Had they started their ‘strategy’ in 2008 they would have richly profited:

Since 2008 punters have lost money on every single vintage except 2012. [Life Spectator]

the 2010 vintage: back to fundamentals

This looks like an amazing schadenfreude story: wine bores and dot com investors get their comeuppance after China enforces anti-corruption laws. Whats not to like ?

But let’s look closely at the chronology:

  1. December 2006: 1855.com IPO
  2. June 2012: formal complaints from clients against 1855.com at “ Tribunal de grande instance de Bordeaux”
  3. September 2011: Duhart Milon prices start to collapse.
  4. November 2012: Anti-corruption campaign under Xi Jinping. Effects felt by whiskey: sales down year on year by March 2013
  5. January 2015: 1855.com entered liquidation

Duhart Milon collapsed almost one year before the official start of the anti-corruption campaign. There should be some additional factors we have not uncovered (read “China and Bordeaux story”):

  • the oversupply of ‘exceptional’ wine after the 2008 (‘lucky’ Chinese number), 2009 and 2010 vintages. It looks like the 2005 was the ‘vintage of the century’ at least until the 2009 arrived, followed by the 2010 (see ‘The last five great Bordeaux vintages’)

Using Judea Pearl’s causal diagram we can model (not very seriously) the factors impacting the price:

Moral of the story:

If paid professionals (like 1855.com) can go bankrupt because they are unable to fully grasp the drivers of price even by using external knowledge (alternative data) about:

  • geography (Google maps, GPS chateaux location),
  • Parker scores (3rd party ratings)
  • cultural understanding (sentiment analysis)

How could we expect a machine learning algorithm to do any better ? As Judea Pearl noted, creating a causal model from pure data is beyond the ability of current machine learning systems (Theoretical Impediments to Machine Learning)

And if that can happen in a small market with ‘simple’ fundamentals, how can we understand what could happen in much more complex financial world? How can a Data gathering technique identify regime changes without having a causal model of the world?

I leave you with a link to one of my favourite wine story (self serving — I do not care if it is really true), which argues that humans have evolved to enjoy wine:

Dr Carrigan investigated the evolutionary history of this enzyme [ADH-4] by comparing the form we carry with the one found in apes, monkeys and other primates.

What he found was that about 10 million years ago a mutation appeared in the common ancestor we share with gorillas and chimpanzees, which allowed that primitive ape to tolerate alcohol.

“… given the choice between starving and eating rotten fruit, it appears eating rotten fruit is advantageous, as long as you have the enzyme to process the alcohol and not get intoxicated.”


Co-Founder of Lamat, a company specialized in solving high-value problems in finance by applying cutting edge numerical methods. www.lamat-uk.com

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store