The Wine Gourd: August 2019

Monday, August 26, 2019

The relationship of price to wine-quality scores

This is a topic that I have periodically addressed before. For example:

Quantifying value-for-money wines - part 1
   — issues with quantifying value for money
Quantifying value-for-money wines - part 2
   — empirically comparing wines within a specified wine group
Quantifying value-for-money wines - part 3
   — formulae for assessing individual wines against a baseline wine
Quantifying value-for-money wines - part 4
   — empirically comparing wines across wine groups
How many wine prices are there?
   — Everyday wines, Better, Premium, Affordable Luxury, Luxury, Icon, Dream
Luxury wines and the relationship of quality to price
   — Bordeaux pricing
Calculating value-for-money wines
   — identifying wines that are good and poor value for their assessed quality

In particular, I have shown that the expected relationship between wine price and wine quality is mathematically known as exponential, as shown in the graph above. This means that the price increase exceeds the quality increase at an ever-increasing rate — value for money is never to be found among high-price wines. Was it ever otherwise?

The Wine Enthusiast data

In my last post, I discussed a dataset from the Wine Enthusiast, covering the years 1999–2017. For our purposes here, there are c. 121,000 wines with both a quality score (80–100) and a price (in $US). We can thus look at the relationship between these two variables using a simple graph, as shown below.

In this graph, the quality points are listed on the horizontal axis, with the prices shown vertically, so that each point in the graph represents one wine. Note that the price axis is logarithmic, so that the exponential relationship will form a linear pattern in the graph. The exponential relationship can thus be shown by the straight lines super-imposed on the points. [Note: if you want to see the same data plotted the other way around, then you can look at: Visualizing wine reviews.]

Quality-price relationship from Wine Enthusiast 1999-2017

The interesting thing about these data is that they cover a much wider range of assessed quality than do other data sets that I have looked at, all the way down to 80 points — most other datasets do not bother with wines <85 points (even though many of us are very happy to drink these as everyday wines).

What the graph shows is that below c. 86 points, there is very little increase in price associated with an increase in quality score. I have not seen this point made before, and yet it seems to be an important one, at least for wine buyers. Why buy an 80-point wine when you can get an 86-point wine for pretty much the same price?

Beyond 86 points, the price starts to increase rapidly, as expected. Here, the value-for-money wines will be those that are the furthest below the straight line, and the rip-offs will be those furthest above the line. The graph seems to indicate that there are actually some good-value wines in the vicinity of 91 points. This is slightly higher quality than I have reported from previous analyses, where the best values were around 89 points, instead.

There are a few caveats to the analysis that I have presented here. The main thing is that the dataset combines all of the numerical variation due to different times (eg. price inflation over 18 years), different tasters (eg. different interpretations of points), and different wine types (eg. countries, grapes). In one sense, this is a “good thing”, because any general data pattern that can be detected among all of this variation must be an important one. In this sense, the graph above is very revealing.

On the other hand, we might learn a lot more from the dataset if we start to subdivide it. In this case, we might discover different patterns through time, or differences among the critics or the wine types. I hope to look at this in future posts.

Biases in the quality scores

Using the same dataset, we can also look at another topic that I have addressed before. This is the distinct bias almost everywhere in the wine world towards awarding 90 points instead of 89. For example, see:

This bias is quite blatant in the Wine Enthusiast data, based on the c. 130,000 quality scores in the dataset, as shown in the next graph. The scores are arranged horizontally, with the vertical bars indicating the number of wines with each score.

Quality scores from Wine Enthusiast 1999-2017

In previous posts I have used the Weibull frequency model as a comparison to evaluate this bias. In this case, it indicates that simply swapping the frequencies for 89 points and 90 points would produce an unbiased dataset. I conclude from this that it is actually arbitrary whether a wine gets awarded 89 or 90 points!

This seems sad, in one sense. Presumably it matters for the sales of any wine whether it gets 89 or the magical 90 points, and yet there is apparently no objective basis for distinguishing between these two choices. What, then, is the role of the wine critics? They are just tossing a coin, for this pair of points.

Perhaps it is worth noting that this type of bias is not detected among those people who use a 20-point quality scale — there is no magical number!

Monday, August 19, 2019

Wine descriptions are related to perceived wine quality

One does not have to spend more than a few seconds reading wine writing to realize that this is not normal human discourse. The writing style used for descriptions of wine seems to cover the full gamut of human expression with the sole exception of normalcy. That is, it covers hyperbole, flowery, literary, allusive, obscure, flamboyant, and sometimes (sadly) even pretentious.

If you want a good laugh, then plenty of people have ridiculed the style of wine descriptions, in everything from books to cartoons. You could try The Complete Guide to Wine Snobbery, if you want a well-written and educational introduction to the topic (by a New York children's-wear manufacturer co-incidentally named Leonard Bernstein); or if you want a rough-around-the-edges look at the topic then try The Illustrated Winespeak (by cartoonist Ronald Searle).

Way back in 2007, Mike Steinberger wrote about Why wine writers talk that way. He started by quoting Fran Lebowitz (well-known for “her sardonic social commentary on American life”), to the effect that: “Great people talk about ideas, average people talk about things, and small people talk about wine.” Steinberger was not impressed with where this must leave people who talk about wine writing.

Well, I have been interested to write a data-analysis blog post about wine words, and wondered how I might do it, while simultaneously avoiding Ms. Lebowitz's sharp tongue. I stopped worrying about it when I realized that someone had already done it for me, and so I could hide behind him, instead.

The basic idea is that there are often commentaries to the effect that the modern idea of giving wines a quality score may not be the best way to discuss wine, and that the descriptions are better. This leads to the obvious query about whether the points and words are actually related.

This has been looked at by Olivier Goutay (2018): Wine ratings prediction using machine learning. He studied the online wine reviews from Wine Enthusiast for the years 1999–2017, totaling c. 92,400 unique wine reviews. The data of interest, for each wine, concern the quality score reported plus the word description.

The first thing he notes about the relationship of points and descriptions is that the higher the points the longer the description. This is shown in the graph here, reproduced from the original.

The quality points are shown horizontally, and each score has a separate box-and-whisker plot above it, showing how long are the descriptions (ie. the number of words). The boxed area shows the range for the middle 50% of the description lengths, with the horizontal center-line indicating the median (50% of the lengths are above the median and 50% below). The vertical line (whisker) on each side of the box indicates the range of most of the rest of the lengths. However, unusual (outlying) values are shown by individual symbols.

The lengths of the descriptions increase continuously from 80 to 95 points, more than doubling from 15 words to 35. This is somewhat reassuring — higher points should indicate a better quality wine, which is then worth a few more words, showing appreciation. The lengths plateau after 95 points, which may also be reassuring — editors must realize that there should be a maximum length to wine descriptions! The sudden increase in length at 99 points (40 words) and 100 points (45 words) is not unexpected, I guess, as the writers wax lyrical about perfection.

Olivier Goutay's purpose in his data analysis was somewhat more ambitious than this: Is it possible, through the computer technique called machine learning, to predict a wine rating (in points) based on its word description? Technically, this is a type of text analysis, sometimes called sentiment analysis.

He concludes that it is possible to produce a very good prediction, at about 97% precision. I will not bore you with the details, which you can read for yourself by checking out the original. However, it should be obvious from the above graph that simply counting the number of words in a Wine Enthusiast description will allow you to work out the number of points pretty accurately, at least up to a score of 95.

As a last point, we might leave the final word to Jon Cohen, who, when describing wine descriptions (Jabberwiney, 2000), noted: “Why do these people write this way? Is this what happens when your job requires you to drink before noon?”

Monday, August 12, 2019

How similar is the wine quality of the Bordeaux First Growth chateaux?

We are always being told about the Premier Cru (First Growth) wines of Bordeaux, as though they are the pinnacle of red wine-making. Now, I have had only a few of these wines, but I have realized that they are not really to my taste; so their outrageous prices do not concern me, personally.

Nevertheless, I have sometimes wondered about the quality relationships between these wines. Which chateaux get high scores in the same years, and which ones do not? This seems like precisely the sort of question that I try to answer in this blog.

This is actually straightforward, because I have already used suitable data in some previous posts (How large is between-critic variation in quality scores? ; Wine-quality scores for premium wines are not consistent through time ; The sources of wine quality-score variation).

The data consist of wine-quality scores for the five premier cru Bordeaux chateaux (Latour, Lafite-Rothschild, Margaux, Mouton-Rothschild, and Haut-Brion) for each of the vintages 1988–2014, inclusive. This makes a total of 195 wines. The scores come from: Jeff Leve, Robert Parker, Jean-Marc Quarin, Jancis Robinson, La Revue du Vin de France, and the Wine Spectator.

There are thus three sources of possible variation in the set of quality scores:

between the 5 chateaux
between the 6 commentators
between the 27 vintages.

The first one is the one I am studying, which means that I need to decide how to address the other two, so that the chateaux comparisons are correctly made. One simple way to address the third source is to average the scores across the years, for each commentator. This leads to two data analyses: one based on the original data, and one based on the averaged data — their difference will tell us how much vintage variation affects the relationships.

As I have done in previous posts, I have used a network to visualize these data, with the network being used as a form of exploratory data analysis. I first calculated the similarity of the different wines, based on the quality scores, using the correlation coefficient. I converted this to a distance measurement; and then used a neighbor-net analysis to display the between-chateaux similarities as two phylogenetic networks.

For the interpretation of such networks, see the post on Summarizing multi-dimensional wine data as graphs, Part 2: networks.

The first network, based on all of the scores for each vintage and each commentator, is rather uninformative. This means that the between-vintage variation is fairly large, and is thereby obscuring the topic of interest (the relationships among the chateaux). However, there is some indication of similarity in quality scores between Latour and Lafite, and between Mouton and Haut Brion.

Network of wine quality for premier cru Bordeaux wines

So, we need to investigate these patterns further, by reducing the effect of between-vintage variation. As noted, this can be done by averaging the scores across the vintages, before producing the network. The result is shown in the next graph.

Network of wine quality for premier cru Bordeaux wines - averaged across years

Things are now much clearer. Indeed, the quality scores for Latour and Lafite are very similar to each other, as are the scores for Mouton and Haut Brion. Margaux is placed in an intermediate position between the latter two.

So what do we conclude from this? First, we can note that there actually are patterns in the data, rather the whole thing being random — certain chateaux do follow each other in terms of variation in wine quality (at least as assessed by the 6 score sources included here).

Second, we might not be surprised that the Margaux wine is different from the others, since it is geographically separate from the others, in the Margaux commune of the Haut Médoc. Similarly, we might not be surprised that the the Lafite Rothschild and Latour wines are similar, since they are both located in the Pauillac commune of the Haut Médoc (although one is in the north of the commune and the other is in the south).

However, our third point might be that Mouton Rothschild is also located in this same commune of Pauillac, and yet its wine quality is most similar to that of the Haut Brion wine, which comes from the Pessac area of Graves. Indeed, these two chateaux are almost the furthest apart of the 5, geographically, while the closest chateau to Mouton is actually Lafite. You might like to think of an explanation of this pattern for yourself.

This obviously leads on to an assessment of the 1855 classification of Bordeaux wines, which I will tackle in a future post.

Monday, August 5, 2019

Swedish wineries — who'd have thought it?

The recent heatwaves throughout the northern hemisphere have set record temperatures in a number of locations. Not the least of these have been traditional grape-growing regions (eg. Europe's heatwave cooks French vines). This has focused attention in the wine world on the current suitability of these regions for making premium wines, at least with their current grape varieties (eg. Bordeaux winemakers allow new grapes to fight climate change). The Millennial “warm climate” used to be the Baby Boomer “cool climate”.

The European Union (EU) has a map, reproduced here, indicating several traditional climatic regions for growing grapes (EU Wine Market Data Portal). The current premium regions are almost all in regions B and C, with only Germany and Austria producing much wine from region A. However, the regions whose wine-making has been expanding this century are all from region A, including Belgium, the Netherlands, Poland, and southern Britain (noting that the regions whose wine-making is being resurrected are from regions B and C, including Croatia, Hungary, Romania, and Slovenia).

The potential of the other northern regions is currently unknown, and we have very little data about their current levels of production. This is because the EU collates production data only from countries that have “a minimum planted area of 500 hectares (ha) of vineyards”, so that only 17 member countries are included in the recent statistics (Vineyards in the EU — statistics).

The most northern of the new vineyard areas are in Sweden (as shown on the map above), although there is at least one vineyard in Finland and two in Norway. There seems to be little information about the Swedish wine-making scene, so I though that I might collect some of it here.

As far as I can tell, there are currently about 250 grape-growers¹ in Sweden, although most of these are hobby growers. Only about 50 growers work commercially; and only about 40 produce wine that is commercially available. The farms vary from a few hundred square meters up to 15 hectares (35 acres). The total national area covers only c. 100 hectares (250 acres).

This is an increase from 2005, when it was estimated that 4,335 liters of wine (1,160 liters of white and 3,175 liters of red) was produced from an area of about 8 hectares. The latest official figures I can find (from the agriculture ministry, Jordbruksverket) are for 2009, when Sweden apparently produced 17,859 litres of wine for commercial use (6,695 litres of red, 7,737 litres of white, 3,427 litres of rosé) from a vineyard area of 20 hectares.

Cultivation in Sweden is largely along the coastal regions, such as the southern Skåne and Blekinge coast, the western coast of Halland, and the eastern islands of Gotland and Öland ² — see this Google Map (reproduced below). The sea provides a protective environment for a longer growing season, but only a subset of the areas have a climate that makes wine growing effective — the main issue seems to be early spring frosts, as it also is elsewhere. There are also a few vineyards around Vättern, a large lake half-way between the east and west coasts.

The southern third of Sweden, with vineyards marked

The grape types commonly grown in Sweden are listed in Wikipedia. You will probably not have heard of most of them. Many are particularly hardy hybrids developed in Germany (eg. Solaris is the most common white-wine grape) or Czechoslovakia (eg. Rondo, a common red-wine grape). However, Chardonnay, Riesling, Cabernet franc, Merlot and Pinot noir do also feature. The wines include whites, reds, rosés and sparkling.

Commercial viticulture was officially sanctioned by the EU from the year 2000 (the same year as Denmark), and most of the commercial wine producers were founded from 2000 to 2015. However, the earliest winery in Sweden actually dates from 1981, which long predates any discussion of global climate change. This is Domaine Bellevue, which was established in the suburbs of Malmö (Sweden's third largest city) as a “little piece of Alsace”. To this day it still has only 200 vinestocks, and a wine production of 80 liters. Nordanvik was then founded in 1996, followed by Domän Sånana in 1998, both along the southern Skåne coast.

Originally, the three biggest wine companies were Gute Vingård, on Gotland (founded 2000), Blaxta Vingård, in Flen (founded 2000), and Håks Gård, on Öland (founded 2000). More recently, places such as Arilds Vingård (founded 2006), Vingården i Klagshamn (founded 2001) and Flädie Vingård (founded 2007) have greatly expanded. Blaxta Vingård is the most northerly vineyard and also the furthest from a large body of water, being just south-west of Stockholm — its claim to fame is its Icewine, made from Vidal blanc.

For reference, I have included below a list of those wine-producers who, as far as I know, have had wine commercially available. Some of the wines have won international awards (eg. Upplev Vinvägen i Skåne lists 11 for the Skåne wineries in 2018).

Föreningen Svenskt Vin (Swedish Wine Association) (formerly called Svenska Vinodlare) was formed in 2001, as an association working for cooperation between wineries and grape-growers, as well as for promoting Swedish wines (nationally and internationally). They run training courses, publish instructional books, run conferences, and are also trying to institute a formal evaluation of wine quality (including diplomas for individual wines). They are responsible for the map linked above (which is not completely comprehensive).

Another organization is VitiNord, intended to “promote the advancement of viticulture and oenology in northern environments that are characterized by cool or short summers and/or cold winters.” Their main activity is a triennial conference (2006, 2009, 2012, 2015, 2018, so far).

There is only one government report about grape-growing in Sweden, which is dated 2006 (Marknadsöversikt — Vin). This has very little information, obviously, as there were only four commercial producers at the time. There is also one book, Tjugo Skånska Vingårdar, from 2013, which (as the title says) describes 20 wineries from Skåne.¹ It is discussed at Wine-Searcher.

You can read a bit more about the Swedish wine scene in a number of online articles:

Finally, the main limitation of Swedish wines at the moment is that they are not good value for money. The economics of the industry is discussed by Mårtensson et al. (2013. Swedish vineyards: a utopia? International Journal of Wine Research 5: 39–45), who note the high production costs associated with the small volumes.

As but one typical example, Jack Jakobsson recently described the Inger-Lena Solaris Barrique 2018, from Vingård Hall, as “the best Swedish Solaris wine I have tried so far”. He gave it 14 / 20 points, which, based on his scores for international wines (see the graph in Calculating value for money wines), would cost an average of c.150 sek (15 usd) — it actually retails for 240 sek (25 usd).³

So, I am very happy to serve Swedish wines to visitors, but I cannot afford to do it too often.

¹ Swedish is confusing, because vin can be translated as either grape or wine, depending on the context. So, a vingård could be either a vineyard or a winery (or both). Also, try not to be put off by the fact that Swedish has three more letters than English: å, ä, ö.

² These are all the names of counties.

³ Sadly, Jakobsson also noted that “one of the best Swedish wines I have tried so far” is Skepparps Mousserande Tirage 2013, made by Skepparps Vingård, but from grapes sourced in Italy. Sigh. Indeed, there is also a company called Högberga Vinfabrik that makes wine from Tuscan grapes, which it then re-exports to Italy!