Breaking the Code: Multi-level Learning in the Eurovision Song Contest

Abstract

Organizations learn from the market, political, and societal responses to their actions. While in some cases both the actions and responses take place in an open manner, in many others, some aspects may be hidden from external observers. The Eurovision Song Contest offers an interesting example to study organizational level learning at two levels: organizers and participants. We find evidence for changes in the rules of the Contest in response to undesired outcomes such as runaway winners. We also find strong evidence of participant learning in the characteristics of competing songs over the 70-years of the Contest. English has been adopted as the lingua franca of the competing songs and pop has become the standard genre. Number of words of lyrics has also grown in response to this collective learning. Remarkably, we find evidence that four participating countries have chosen to ignore the "lesson" that English lyrics increase winning probability. This choice is consistent with utility functions that award greater value to featuring national language than to winning the Contest. Indeed, we find evidence that some countries -- but not Germany -- appear to be less susceptible to "peer" pressure. These observations appear to be valid beyond Eurovision.

Figures (28)

• Figure 1: Over its history, Eurovision has experienced dramatic changes in the range of nations participating and voting, in the rules underlying the contest, and in its outcomes. These changes can be interpreted response of the organizers to an increased homogeneity of songs competing.a, The number of countries participating in Eurovision has increased steadily over time. A major change, however, occurred in 2004 when there was an increase in the number of participants and nations allowed to vote by nearly 50%. This increase required the use of semifinal contests. We split this 68-year history into 3 periods that we denote "Formation", "Consolidation", and "Expansion". b, Changes in the voting process resulted in changes in the distribution of the votes received by participating nations (see Fig. S\ref{['fig:vote_distribution']}). Importantly, it also affected the fraction of total votes received by the top 3 performing songs. It is visually apparent, nonetheless, that the system reached a stationary state after 1974. c, While the voting share of the winners remained stationary after 1974, the voting patterns did not. The top panel here shows the number of top three finishes for the countries participating in the 1957 to 1975 Contests. It is visually apparent that some countries systematically do well while many others do not. Indeed, the top performing 4 countries took the majority of top 3 finishes. This concentration of winning decreases over time to almost parity in the Expansion period. This evolution suggests that the advantage held by some countries initially was lost over time .
• Figure 2: Despite the diversity of European languages, English becomes the de facto language of Eurovision.a. Fraction of songs using the languages of most of the initial competing countries. During the "formation" period, there is a great diversity in the language of the competing songs. This diversity increases through most of the "Consolidation" period. Suddenly, in 1998, there is a transition to a new state, where English becomes the dominant language. In this new state of affairs, over 70% of songs are entirely in English or a mix of English and another language. b. When considering the languages of the top performing songs, the situation is quite different. Initially, French is the dominant language but, already by the 1970s, nearly 40--60% of the winning songs are in English (see full black line, which shows 5-year running average). By the early 1990s, nearly 80% of winning songs are in English. While the rate of increase is less pronounced for the top 3 and top 5, the data is consistent with the learning by participants that songs with English lyrics have an advantage. The increased dominance of English as the language of the top 3 songs is accompanied by a decrease of French, which nearly disappears as the language of choice among the top 5 songs after 1990.
• Figure 3: Song lyrics also show sings of change over time consistent with participant learning.a. Time evolution of mean lyrics size (as measured by number of words) for all songs, winning song, and Top 3 songs. For all songs, the shaded region sows the 95% confidence interval for the estimate of the mean. As for song language, we find a step-like increase in lyrics size in 1999. Demonstrating the presence of a signal to be learned, we find that the lyrics size of Top 3 songs was already increasing during the period 1975--1998. Interestingly, the winning songs, are characterized by extreme lyrics sizes, either very small or very large. As shown in Fig. \ref{['fig:lyric_sizes']}, there is no significant dependence of lyrics size on country. b. Time evolution of the percentage representation of the four most informative song themes for all songs and Top 3 songs. The black circles show the average over all competing songs for a given year, the line and shaded regions show 99% confidence intervals for the linear fit, and the colored lines show running 10-year running average of theme representation among Top 3 songs. All four themes display strong and statistically significant temporal trends. While top performing songs follow the common trend for Nostalgia, they focused less on Pain between 1990 and 2000, but since 2010 are focusing more on Pain than the average of all competing songs.
• Figure 4: Despite the diversity of music styles and traditions in Europe, pop has become the dominant music genres for participating songs, but not for winning songs.a. Annual mean audio feature scores over time of competing songs. The solid lines show the mean value of the attribute for the songs competing in a given year, while the red shaded areas display the $95\%$ confidence interval for the mean. The gray lines in the rightmost plot show fits to the functional forms discussed in the text. b. Mean annual genre score of all competing songs and winning songs. The mean pop score of competing songs has been increasing steadily over the existence of the competition. Intriguingly, the genre of the winning songs shows much different trends (see black line, which shows the running average for 5-year time windows). Since about 1970, half of the winning songs have been pop songs. This pattern is also present when considering the top 3 or top 5 songs (Fig. \ref{['fig:music_genre_top3-5']}) but occurs at a slower rate. These results suggest that, unlike language, where English is the standard, there is more opportunity for "bucking the trend" in song genre.
• Figure 5: Participant learning of winning strategies.a. Average Acousticness and Danceability scores of songs submitted to Eurovision by different countries across two time periods. It is visually apparent that, prior to 1980, competing songs had high Acousticness, but after 1980, competing songs --- across all countries --- had high Danceability. We find similar results for lyric sizes (Fig. \ref{['fig:lyric_sizes']}). b. Linear regression of danceability of a country's song in year $y$ against the average danceability of all other countries in the previous 3 years. b. Regression coefficients for the 11 countries for which the linear model is statistically significant and $R^2$ of the linear fit. d. Fraction of songs with lyrics in a given language submitted to Eurovision nations across two times periods. Prior to 1999, most nations submitted songs in one of their national languages (see Switzerland pre-1999). Post-1999, most countries submitted songs with English lyrics. The exceptions were France, Italy, Portugal and Spain, which were mostly submitting songs in their national languages. e. Logistic regression of a country's song having English lyrics in year $y$ to the average fraction of songs placing in the top 5 with English lyrics in the previous 3 years. f. Regression coefficients for the 10 countries for which the logistic model is statistically significant and pseudo-$R^2$ of the fit.