ROGER DE PILES

Roger de Piles was born in Clamecy, France on October 7, 1635. He went to Paris at about the age of 15 and studied philosophy at the College du Plessis and then theology at the Sorbonne. He also learned to paint while in Paris. In 1662 Charles Amelot, president of the King's Great Council, hired him to tutor his seven year old son, Michel Amelot de Gournay. De Piles remained attached to Michel Amelot de Gournay for the rest of his life, subsequently serving him in the capacity of personal secretary.

With Michel Amelot de Gourney, De Piles traveled widely throughout Europe. In particular, in 1673 Amelot embarked on a 14-month grand tour of Italy and in 1682 he was appointed French Ambassador to Venice. De Piles followed his master and used these opportunities to view and become an expert on the important artists of his time and earlier periods. The Venetian style of painting, with its emphasis on warmth and color, had a profound effect on De Piles and most likely influenced the importance that he attached to colore. De Piles had other opportunities to mix travel and art: in 1685 the French government sent him on a spying mission to Germany and Austria under the pretext of studying and purchasing art for the Royal Collections. In 1692 he was sent to spy in Holland but his mail was intercepted, and he ended up spending four years in a Dutch prison.

While in prison, De Piles wrote his first work of art criticism, the Abrégé de la Vie des Peintres (1699), published two years after his release. During the same year, De Piles was admitted to the Royal Academy of Painting and Sculpture, and for the next ten years he was its leading critic and official theorist. In 1708, one year before his death, De Piles wrote the Cours de Peinture par Principes, in which his controversial table of ratings, the Balance des Peintres, was published. The table presents his rankings of 58 artists. He provided a numeric score for each artist in four categories: composition, drawing, color, and expression.

De Piles' ratings are controversial. As described in Ginsburgh and Weyers (2008), Julius von Schlosser, the leader of the second Vienna school of art history, hated the ratings. The art historian Ernst Gombrich, in his 1966 work, Norm and Form, described the exercise as a “notorious aberration” (p. 76). Thomas Puttfarken (1985), in his definitive work, Roger de Piles' Theory of Art, rejects both the quantitative approach and the rankings:

This article takes a second look at De Piles' rankings and shows that Puttfarken and other critics may have been hasty.

DE PILES' SAMPLE

Roger de Piles' chose 58 artists to rank, out of hundreds or thousands of possibilities. Before analyzing the ranking, it is worthwhile to look at his sample and just how restricted it is. Consider the 265 artists selected by Georgio Vasari (1511–1574) for inclusion in his 1568 edition of Lives of the Most Excellent Painters, Sculptors, and Architects. De Piles kept only 21 of Vasari's artists. De Piles extended his selection by including a limited number of other Italian painters in addition to Flemish, French and German artists in his rankings that were already known during Vasari's life, in addition to other painters that became famous after Vasari's death. The breadth of the sample no doubt resulted from De Piles' extensive travels.

How well have De Piles' selections done recently? In the 13th edition of Gardner's Art Through the Ages (2011), 90 different artists have paintings that are illustrated in chapters covering artists from Italy, Spain, and Northern Europe from 1400 to 1700; 23 of those artists, or 25 percent, were rated by De Piles. Of the 149 different artists currently on display with oil paintings at the Metropolitan Museum of Art and that are identified on their website as being painters from 1400–1800, but restricted to those who died before 1705, De Piles rated 30 of them or about 20 percent. Finally, in Wikipedia's Incomplete list of Most Important Old Masters, restricting the list to those who lived between 1427 and 1705 inclusive (the dates of the lives of De Piles' artists), we find that De Piles has rated 35 of these, or about 15 percent. Table 1 summarizes these results. Looking at it a different way, in a combination of Gardner's, the Met exhibition, and Wikipedia, only 14, or about 25 percent of De Piles' artists, did NOT show up in these illustrations, this exhibition, or this list.

TABLE 1

DE PILES' SELECTION

^a Illustrations of paintings included in chapters on Italy and Spain and Northern Europe from 1400 to 1700.

^b Oil paintings that are identified on their website as being painters from 1400–1800, but restricted to those who died before 1705.

^c Painters who lived between 1427 and 1705, inclusive.

Sources:

See column titles and the text.

To what extent was De Piles just following previous established reputations? As indicated in Table 1, 45 of the artists rated by De Piles died earlier than 1650, and therefore must have an established reputation already; thus he only chose 13 artists in the half century before he produced the table. But, with these latter artists only one artist, Francesco Albani (1578–1660), in this group did not show up in Gardner's, The Met's exhibitions, or Wikipedia.

We also counted the number of words on each artist with a biography in the online Benezit Dictionary of Artists for both De Piles' ranked artists and for all of the artists listed in Wikipedia's list. The average number of words for De Piles' rated artists is 1,091 and the average number of words for artists that are included in this list but not included in his list is 649 (a statistically significant difference). Clearly, he had some ability to choose artists. But, if we then compare the average word count of De Piles' 58 rated artists (1,091) with the word count of the 58 artists with the highest word counts that were not rated by De Piles, the average word count for these artists is 1,288, which is statistically significantly higher than the word count for the De Piles' sample at the 10 percent level. Clearly, De Piles had some ability to choose artists, but, in hindsight, he did not find those artists from that era that are considered most important by today's standards.

Online Appendix Table 1 presents a list of all of De Piles' rated artists, ordered by their total score.Footnote ² Note that, importantly, De Piles did not include the total score in his table because doing so would have implied each score had equal importance.

The price data were compiled from two sources. First, we collected the prices and sizes of paintings by artists that were rated by De Piles from the online version of Art Sales Index.Footnote ³ Only paintings with clear attributions were included (those with qualifier such as “attributed to” or “school of” were left out). Information on paintings from this source only goes back to 1922, with relatively sparse data prior 1950. Auction prices were found for all of De Piles' 58 rated artists.Footnote ⁴ We exclude three artists. Guido Reni and Polidoro de Caravaggio were left out because De Piles had not given them complete rankings. We also set aside David Teniers the Younger. He was not included because Art Sales Index listed over 1,600 sales of his work from 1920 to the present, which is about four times the amount of the next most prolific artist. The number of sales raised concerns about attribution in addition to concerns about an unbalanced data set. Roger de Piles lumped Agostino, Annibale, and Lodovico Carracci together as one artist, “The Carracci.” Separate price and critical data were gathered on each of the three Carracci artists. The effort yields 4,136 observations on 55 different artists.

The second data set draws from Reitlinger's The Economics of Taste (1961, 1963, 1971) the prices of paintings by artists that were rated by De Piles. The time span of this data set is from 1740 to 1960 and includes 761 observations of both private sales and sales at auction. Reitlinger includes only a selection (24) of De Piles' artists and their paintings, and it does not give characteristics of the works such as height and width.

TABLE 2

CHARACTERISTICS OF ART SALES INDEX DATA

Notes:

Prices are adjusted using the composite price index from the 2003 report, Consumer Price Inflation Since 1750, published by the Office of National Statistics (ONS), combined with recent CPI data from the website of the ONS.

Tables 2 and 3 present summary statistics of prices and various characteristics. The nature of the data set does change over the decades. Namely, the average “quality” of the artists sold at auction—as measured by De Piles' ratings—declines over time in the Art Sales Index data set.

TABLE 3

CHARACTERISTICS OF REITLINGER DATA

*For the period 1731–1750, there was one sale reported at 6,500, which was over four times the amount of the next highest sale. This one data point is excluded.

Sources:

Prices are adjusted using the composite price index from the 2003 report, Consumer Price Inflation Since 1750, published by the Office of National Statistics (ONS), combined with recent CPI data from the website of the ONS. The index before 1750 was gathered from the Phelps Brown and Hopkins (1956).

It may be that better paintings during this time period are not being resold but are being held in museums or in private collections. The same decline in De Piles' ratings is not present in the Reitlinger data set, in which all sales occur before 1970, and in which the bulk of the sales occur before 1960. The fact that a decline appears to take place only in the Art Sales data set may either reflect the late twentieth-century boom in the art market or that Reitlinger only collects prices on a subsample of De Piles' artists.

One feature of the data that is apparent by comparing these tables is the difference in works included in the two data sets. For example, the 1921–1930 period in the Art Sales Index Data lists an average price of £851 (£34,814 in 2010 GBP), and the average price for the 1911–1930 period in the Reitlinger data set is £20,200 (£1,146,170 in 2010 GBP).

This large price difference is not solely explained by different artists in the two data sets, but is likely explained by Reitlinger only including selected reported sales in his data set, mainly private sales between individuals. Because of the differences in the two data sets, the two data sets are not combined but are analyzed separately. As shown below, results are broadly consistent across the two samples and when they differ they are consistent with the favorable sample bias in Reitlinger.

DE PILES' RANKINGS

Because Roger de Piles' assigned numerical values to his painters, it is possible to go beyond just studying his sample as compared to other samples. A common criticism of De Piles' is that in his scores, he failed to recognize Michelangelo's genius. While this may be true, a very cursory analysis of the number of words given in Benezit's Dictionary of Art to the top half of De Piles' artists in relation to the bottom half of De Piles' artists shows average word scores of 1,330 for the top half, and only 851 for the bottom half. These numbers are highly statistically significantly different from one another. While Michelangelo might have been a mistake, and perhaps Poussin wasn't given his due, on average, De Piles' rankings have held up according to the space allocated to the artists in Benezit's.

As an initial analysis to determine the relevance of De Piles' rankings over a long period of time, we regress the natural log of the price of painting i at time t, p_it, on De Piles' overall ranking for each artist, z_ī, (ī denotes that that rankings vary by artist, not by painting—and we will cluster the standard errors by artists), a set of dummy time variables, δ_t, and an error term, ∊_it.Footnote ⁵ Because of the difference in sample sizes and time spans, we use year fixed effects in the Art Sales data set and decade fixed effects in the Reitlinger data set. The regression is

(1)

TABLE 4

MAIN DATA INPUTS FOR 1774 INCOME ESTIMATES

* = Significant at the 5 percent level.

Notes:

These are robust errors clustered by artist.

We find that the total ranking in the Reitlinger data set has a statistically significant effect (at the 1 percent confidence level) on price (the coefficient on total ranking is 0.048 with a robust standard error of 0.007; the R-squared is 0.45 with 644 observations). For the Art Sales Index data set, the effect is not significant (the coefficient on total ranking is 0.024 with a robust standard error of 0.016.) However, if we control for height and width, the effect becomes significant at the 10 percent level with a coefficient on total ranking of 0.029 and a robust standard error of 0.017; the R-squared is 0.56 with 4,136 observations. These results are presented in Table 4. Over the longer time period of the Reitlinger data set, namely 1730 to 1970, De Piles' original rankings have had a significant effect on price, more recently (1920 through 2010), the results are ambiguous.

A price index can also be used as a measure of De Piles' long-standing success. We focus solely on prices fetched for artists within De Piles' sample, rather than also focusing on prices of artists not rated by De Piles because of the ambiguity regarding which artists should be included in a not-rated sample. Ideally, one would like to include all of the artists from that era that appeared at auction, but because of the difficulties of price data collection and the number of artists this would entail, this becomes a nearly impossible task. Any results from a selected sample will be colored by the way in which the sample was chosen. This became evident when counting number of words in Benezit biographies: as noted above, if all Benezit biographies that were included in Wikipedia's list are taken into account, De Piles' selection looks very good. But, if De Piles' selection is compared with other long biographies of artists that were not ranked by De Piles, De Piles' selection looks less prescient. We therefore stick to developing price indices and comparing them to De Piles' rankings for the artists that were ranked by De Piles.

Because the data sets are so different, we construct separate indices for the Reitlinger data set and for the Art Sales Index data set. The econometric model used to construct the index is a variant of equation 1 above, except that for the painting characteristics, X_i, artist dummy variables as well as ln height and ln width are used for the Art Sales Index data set. As we do not have information on size in the Reitlinger data set, artist dummy variables alone are used for painting characteristics in this data set.Footnote ⁶ The index is then calculated as exp^yt and the annual return over a period t₁ to t_n is calculated as where t_n is year n and t₁ is year 1. The regression coefficients on the time dummies for the regressions are presented in Appendix Table 1 for the Art Sales database and in Appendix Table 2 for the Reitlinger database. In addition, two indices are constructed by splitting the data set using the median overall artist rating (44) in the entire data set as a break point. Artists whose overall ratings are greater than the median are included in one data set and those whose overall ratings are less than or equal to the median are included in the other data set.

Table 5 presents the annual return estimates. The returns to De Piles' rated artists are similar to the returns estimated in other larger data sets. For example, the average annual return for 1740–1960 is 1.83 percent and is between William Baumol's (1986) and William Goetzman's (1993) average returns for similar time periods. The more recent returns, from 1950 to 1999, are less than the Jianping Mei and Michael Moses (2002) estimated returns, but the Mei and Moses returns are high when compared with other studies.Footnote ⁷

TABLE 5

NOMINAL RETURNS TO DE PILES’ ARTISTS

Notes:

In their 2002 paper, Mei and Moses only report real returns. We use Shiller's calculations of CPI to inflate these returns to the nominal returns shown above.

Comparing the two samples of De Piles' artists that were split by De Piles' ranking is extremely interesting. In both data sets, the average returns to the top sample exceed the returns to artists with lower ratings, and the coefficients on the final time period are statistically significantly different from one another at the 1 percent level for the Art Sales data set but not for the Reitlinger data set.Footnote ⁸ The fact that several centuries later De Piles' top rated artists significantly outperformed De Piles' lower-rated artists over a 90-year period is astounding.

TABLE 6

PRICE LEVELS AND OBSERVATIONS

Sources:

Prices are adjusted using the composite price index from the 2003 report, Consumer Price Inflation Since 1750, published by the Office of National Statistics (ONS), combined with recent CPI data from the website of the ONS. The index before 1750 was gathered from the Phelps Brown and Hopkins (1956).

The economic history of a painting, namely, the price achieved at auction and in private sales, has in many ways validated De Piles' judgment as an art critic and contradicts much of the previous literature on expert opinion, which has largely found expert opinions do not hold up well in the long run (see Ashenfelter and Jones 2000; Ginsburgh and Van Ours 2003; Ginsburgh 2003; Landes 2004).

Why did De Piles' top-rated artists return more than his lower-rated artists over the two time periods? It could be that in the early years, De Piles' top-rated artists were already superstars—i.e., were priced more highly—than his lower-rated artists. Table 6 compares the average prices (in 2010 GBP) received for these two groups of artists. Dividing both data sets into thirds, the averages for the initial prices are higher for the higher rated artists in the first third of the time period, though they are not statistically significantly higher. The prices for the higher-rated group are statistically significantly higher in the last third of the time period. This data is consistent with a superstar model, where the top-rated artists become even more valuable and rare, which is reflected in their prices. These results argue against the theory that tastes have changed to become more aligned with De Piles' ratings: De Piles' lower-rated artists had lower prices in the period 1920–1950. Furthermore, as De Piles rated each characteristic separately, even if he favored one characteristic over another (his opinions on color and drawing are discussed below), forcing himself to rate each of these characteristics separately likely decreased the degree to which his personal preferences on the individual characteristics influenced the total ratings and thus strengthened the success of his combined ratings.

One must also worry about sample selection of particular paintings, in particular because as Table 6 shows, the paintings by the lower-rated artists have proliferated since the first part of the twentieth century. The ratio of number of paintings sold by highly rated artists to low-rated artists from 1920–1950 is about 4 to 1. By 1980–2010 the ratio is less than 1; far more paintings by lower-rated artists than highly rated artists were sold during this period. There are two possibilities that we could not distinguish between. It could be that the supply of highly rated artists has almost dried up and thus they benefit from a “scarcity” premium. But it could also be that many more inferior paintings by lower-rated artists have come on the market in recent years because even for them important paintings have been snapped up by museums or taken off the market—either explanation would contribute to the lower returns for the lower-rated artists. An interesting project for future research would be to gather data on the dates of acquisitions by museums of paintings by these artists.

DE PILES' RATINGS AND CHANGING TASTES

Before further regression analysis, let us review De Piles' four characteristics (composition, expression, drawing, and color) and their meaning to art critics. Composition involves how a work is ordered and laid out; the way that various objects are placed on a canvas. Expression is the emotional effect that a work has on its viewer, which can be achieved by representation of expressiveness or human emotions (Grove Art Online). De Piles' drawing (disegno) characteristic represents more than just the physical ability to draw or depict an object. According to David Rosand (1982), “drawing is viewed as the key to the entire imaginative process, the medium of the painter's very thought as well as of its concrete expression. From the initial conception of the idea through its formal statement in sketches to its final execution in a finished cartoon, the entire creative procedure is defined by Vasari [1511–1574] essentially in terms of disegno.” (p. 16). According to Rosand (1982) and Puttfarken (1985), color was often philosophically compared with elocution or ornamentation; De Piles broke with tradition and did not consider color as simply accidental ornamentation, but the main condition of an object's visibility. Thus color, to De Piles, was part of the natural order of painting.

Most of De Piles' life was spent outside the Royal Academy of Painting and Sculpture and his views, especially on color, did not coincide with the leading critics of the time (Charles LeBrun (1619–1690) and André Félibian (1619–1695)). De Piles, in particular, disagreed that a painting should be viewed as a story with episodes. For him, a picture should be more in tune with nature, in which colors and shapes are of great importance (Holt 1994). His views were undoubtedly influenced by the importance of color in Venetian painting, but they may have reflected the popular viewpoint, rather than the expert theoretical views of the time, though by the time the Balance de Peintres was published, De Piles was a member of the Royal Academy.

Figure 1 presents histograms of the various characteristics. It is interesting to note that the average drawing rating is actually slightly higher than the average color rating. Furthermore, the histogram becomes approximately normal when the characteristics are combined.

The econometric model used to test whether De Piles' individual characteristics have an effect on price is as follows,

(2)

In the Art Sales Index data set, p_it is the price of work i in year t, X_i is a vector of characteristics of painting i that include De Piles' ratings on composition, drawing, color, and expression, which vary by artist, in addition to height and width of the paintings, which vary by painting. τ_T are a set of decade dummy variables that are interacted with the painting characteristics, and β_T are the coefficients on the characteristics that vary by decade.Footnote ⁹ Year dummy variables are also included (equal to 1 if painting i is sold in year t and zero otherwise) with corresponding yearly coefficients, γ_t. The error term, ∊_it, varies by item i at auction date t.

In the Reitlinger data set, p_{i t} is the price of work i in year t, X _i is a vector that contains only De Piles' ratings for each artist, τ_T are a set of 20-year dummy variables that are interacted with De Piles' characteristics, and β_T are the coefficients on the characteristics that vary by 20-year periods. Decade dummy variables are also included with corresponding decade coefficients, γ_t. The error term, ∊_it, varies by item i at auction date t. Table 7 presents the results of this analysis for the Art Sales Index database, and Table 8 shows the results for the Reitlinger database. Some coefficients on the characteristics are statistically significant in some decades, but others are not. The majority of individual coefficients are insignificant. However, with the exception of composition in the Art Sales data set, by using an F-test on a subset of variables we can reject that each characteristic in all time periods is zero. In other words, all other characteristics are jointly statistically significantly different from zero over time in both data sets. Though there is joint significance, the individual effects are not constant or even consistent, and in many periods the characteristics do not appear to be correlated with price.

FIGURE 1

PLOTS OF CHARACTERISTIC COEFFICIENTS OVER TIME

Sources: See the text.

TABLE 7

ART SALES INDEX REGRESSIONS

* = Significant at the 5 percent level.

Notes:

Number of observations = 4,136, R-squareds = 0.585 (interacted) and 0.572 (restricted). Regressions include 74 year fixed effects and a constant. Robust standard errors, clustered by artist, are in parentheses.

For the Art Sales Index data set, if we include decade fixed effects and a constant in the regressions, 53.5 percent of the variance is explained. Including length and width interacted with decade dummy variables increases the R-squared to 55.2 percent, and then including De Piles' characteristics interacted with the dummy variables increases the R-squared to 58.5 percent. Thus, De Piles' characteristics only explain about 3 percent of the variance in price in the shorter data set; nonetheless, expression, color, and drawing are all jointly significant. The restricted total score only explained about 2 percent of the variance in the Art Sales Index data set.

TABLE 8

REITLINGER REGRESSIONS

* = Significant at the 5 percent level.

Notes:

Regressions include 20 decade fixed effects and a constant. Robust standard errors, clustered by artist, are in parentheses.

In the Reitlinger data set, if only the 20-year fixed effects and a constant are included in the regressions, 37 percent of the variance is explained. When the interacted characteristics are also included, the R-squared increases to 52.5 percent, suggesting that the interacted characteristics explain about 15 percent of the variance in price. The restricted total score explained about 8 percent of the variance in the Reitlinger data set. Reitlinger did not include information on painting dimensions in his published price data. If the length and width characteristics are excluded from the main regression in the Art Sales Index data set, the coefficients on the De Piles' characteristics change very little, indicating that the lack of these characteristics is having little effect on the Reitlinger regressions. The composition coefficients remain jointly insignificant, and the expression, drawing, and color coefficients remain jointly statistically significant.

When the coefficients are restricted across years as in Table 9, in the shorter time period data set collected from Art Sales Index, only composition, color, and width are statistically significant; expression and drawing have little bearing on price. Not surprisingly, when the characteristics coefficients are restricted across the 220-year period in the Reitlinger database, none of the restricted coefficients are significantly different from zero. So, we find that while the coefficients are almost all jointly significantly different from zero, when the coefficients are restricted to be identical across time, the significance disappears for all of the characteristics in the Reitlinger data set and for some of the characteristics in the Art Sales Index database. This change in significance strongly indicates that while De Piles' ratings influence price over some time periods of the study, the effects differ over time and De Piles' individual ratings do not consistently affect price. This variability appears when we plot the characteristics coefficients by year, especially over the historical period in the Reitlinger data set, as in Figure 2 below. The plots that stand out amongst the others are the plots of the point estimates of the coefficients on drawing and color in the Reitlinger data set that spans from 1736–1960. The plot clearly shows a decline in the importance of drawing over time and less clearly but still visible, an increase in the importance of color as reflected by prices, though many of these coefficients are statistically insignificant.Footnote ¹⁰

TABLE 9

HEDONIC REGRESSIONS

* = Significant at the 5 percent level.

Notes:

These are robust errors clustered by artist.

The debate between the importance of drawing and color (disegno and colore) is longstanding. It dates back to Giorgio Vasari, often taken as the founder of art history and who is well-known for his biographies of Renaissance artists. Vasari, along with the art establishment at that time, firmly believed that the importance of a painting lay in its design rather than its execution. For Vasari and other later critics such as Félibian, color was considered mere ornamentation. Some art theorists believe this view began with Plato's rejection of the image; for others the dichotomy between color and drawing was drawn by Aristotle. According to Jacqueline Lichtenstein (1993), “Plato condemned painting because of its colors and Aristotle reprieves it for its drawing” (p. 62). Breaking with this tradition by asserting the importance of color, in 1673 De Piles published the Dialogue sur le Coloris in support of the Venetian style and their use of color. At the time, the followers of Nicolas Poussin (1594–1665) were on the side of design (disegno) and the followers of Peter Paul Rubens (1577–1640) were on the side of color (colore). The supporters of Rubens ultimately triumphed when Rubens became known as the greatest European Master (Grove Art Online). De Piles emphasis on color over drawing was prescient, and our statistical results demonstrate that the importance of drawing appears to have declined over the decades, while that of color has not and it may have increased somewhat. In his emphasis on colore vs. disegno, De Piles foresaw this change in taste.

FIGURE 2

PLOTS OF CHARACTERISTIC COEFFICIENTS OVER TIME

Sources: See the text.

The extent to which De Pile's emphasis on color has helped his total rating withstand the test of time is debatable. While it is possible that tastes are simply more aligned now with De Piles' ratings because of the increasing importance of color on price, the fact that he rated each of the characteristics separately makes this change of taste less likely to drive his total rankings. The act of rating each of the characteristics separately has very likely increased the durability of his total ranking.