• No results found

Analyzing listening behaviour from digital music services to measure the impact of live performances

N/A
N/A
Protected

Academic year: 2021

Share "Analyzing listening behaviour from digital music services to measure the impact of live performances"

Copied!
22
0
0

Bezig met laden.... (Bekijk nu de volledige tekst)

Hele tekst

(1)

Analyzing listening behaviour from digital music

services to measure the impact of live performances

By Bernd van der Wielen

Supervision by: Frank Nack

University of Amsterdam

26 March 2015

(2)

Abstract

Live performances are becoming increasingly important for artists in terms of revenue as revenue from other sources such as music sales and royalties are decreasing. This is caused due to the changing landscape of the music industry which is impacted by file sharing and legal alternatives that provide lower royalties. In this paper we explore the impact of live performance on users’ digital music listening behaviour. We have set up an empirical research which uses data provided by Last.FM and MusicBrainz to see if live performances cause more people to listen to music of the performing artist and look if there any differences between genre, popularity and the amount of active years of an artist. We provide evidence for a negative impact of live performances on users’ listening behaviour

(3)

Table of Contents

1. Introduction

page 1

2. Theoretical framework

page 2

3. Research

page 4

4. Results

page 6

5. Discussion

page 16

6. Future work

page 17

7. References

page 18

(4)

1. Introduction

The development of peer-to-peer (P2P) networks and popularized appliances such as Napster and 1

more recently The Pirate Bay have made it possible for people to easily share files with each other 2

over the internet, including copyrighted material such as music. This has caused a major decline in revenue of traditional music sales, such as physical CDs, since 2000 . Research has shown that 3

illegal sharing of copyrighted music is a major reason for people not to buy music anymore (Zentner, 2006). Because of this development artists earn less from their released records.

The music industry has been trying to reduce piracy through lawsuits against popular P2P platforms, but new alternative P2P platforms would emerge quickly when one was taken down. More success has been achieved with providing legal alternatives where users can buy individual songs for a low price instead of complete albums. One of the first efforts with support from major record labels such as Universal Music Group, Sony BMG and Warner Music was Apple’s iTunes Store , which 4

launched in 2003. The iTunes Store offers single songs for $0.99, popularzing the concept of buying individual songs instead of whole albums. Since then other similar platforms have been launched such as Amazon Prime Music and Google Play Music . This development has lead to a new stream 5 6

of revenue for the music industry that has become increasingly important. 78

While these new (digital) distribution channels have generated new revenue for record labels to compensate for the decline in traditional music sales, artists are seeing a lower royalty fee for digital sales when compared to traditional sales. Royalties are further decreasing due to the relative new phenomenon of streaming music offered by services such as Spotify . These services use a 9

subscription based model which allows users to listen to the service’s music catalogue unlimited or present advertisements to support a free version. An example of decreased royalties is that 131.000 song plays earn only $547.71 on Spotify . This is an average of 0.42 cent for every play compared to 10

a 7-10 cent royalty on a $0.99 downloaded song or around $1.00-2.50 for a traditional album of around $17.00.

Because of this development a large part of the recording and performing artists have started to rely more on live performances as a primary income. An illustration of this is that the live

performance industry has already surpassed record sales in terms of revenue in the United

Napster
 1

http://iml.jou.ufl.edu/projects/spring01/burkhalter/napster%20history.html

The Pirate Bay
 2

http://thepiratebay.cr/about

CNN: Music's lost decade: Sales cut in half


3

http://money.cnn.com/2010/02/02/news/companies/napster_music_industry/

Apple launches iTunes Music Store 
 4

http://www.apple.com/pr/library/2003/04/28Apple-Launches-the-iTunes-Music-Store.html

Amazon Prime Music
 5

http://www.amazon.com/gp/feature.html?docId=1002557791

Google Play Music 6

https://play.google.com/music/listen

CNN: How iTunes crushed music sales
 7

http://money.cnn.com/interactive/technology/itunes-music-decline/?iid=EL Mashable: Rise of Digital Music Over the Past 10 Years


8

http://mashable.com/2014/03/19/digital-music-chart/ Spotify

9

https://www.spotify.com/uk/about-us/contact/

As Music Streaming Grows, Royalties Slow to a Trickle 10

(5)

Kingdom . The business model of the live performance industry has largely remained unchanged in 11

the last few decades and could benefit from developments in the music industry related to digital music sales. Modern services such as iTunes and Spotify analyze users’ behaviour to attempt to boost sales or usage by using recommendation systems, which will be discussed in more detail later. Artists could use similar techniques to analyze the behaviour of their listeners to tailor the promotion for concerts or the concert itself in a specific region to increase demand for tickets or increased sales from digital music platforms. A first step in this process would be to discover the current impact of live performances on digital listening behaviour.

Our hypothesis is that live performances positively impact users’ digital listening behaviour. We define impact as a (significant) change in the number of people that listen to songs of an artist or the number of songs they listen to of an artist in a specified timeframe. We believe this positive impact is true for the amount of people that listen to an artist but not that people will listen to more songs of an artist. We expect to find differences when looking at genre, the average popularity and the amount of active years of an artist and that more recently released songs are more popular than older songs.

To test our hypothesis we have defined the following research question: ‘Do live

performances impact digital listening behaviour?’ To be able to answer this question we will try to

answer the following sub questions:

■ Do live performances cause people to listen more to recently released songs of the performing artist than older songs?

■ Do live performances cause more people listen to songs of the performing artist? ■ Do live performances cause people to listen to more songs of the performing artist?

■ Is there a difference, regarding the previous three questions, in terms of popularity, genre and how long an artist has been active?

Before we discuss the used research method to answer these questions we will first look at existing theory about live performances, the music industry and digital music.

2. Theoretical framework

It is evident that technological developments have impacted the music industry in a negative way due to increased illegal file sharing, but these developments have also enabled the music industry to develop new distribution channels and models. File sharing has reduced the probability that people will buy music by around 30% according to the research of Zentner (2006) and because of this the model of relying on copyright to cover costs and risk is under pressure. The position of record labels has weakened due to reduced costs to produce and distribute music for artists without the aid of a record label via these new distribution channels and models (Schultz, 2009).

Artists and their source of income are also being affected by these changes. Artists vary in their distribution of revenue sources according the research of DiCola (2013). Some artists rely more on their income from royalties related to copyrighted material while for other artists this is only a small

The Guardian: Live music outperforms record sales

11

(6)

part of their income and for example rely more on their income from live performances. New

technologies such as streaming offer lower royalties than traditional music sales and the research of DiCola shows the increased economic importance of live performances for all music genres to compensate for the decreasing music sales. One exception of this behaviour is observed for classical music artists, who rely more on a salary income. The research of DiCola also shows that high-income, popular artists benefit more from income related to copyrighted materials.

Going back to the research of Schultz (2009), reliance on revenue from live performances can be considered an alternative business model for the music industry, with the production and release of albums as a tool to increase demand for live performances. This effect is already being observed with illegal file sharing boosting non-digital complementary productions (Mortimer et al., 2012). Mortimer et al. state that while file sharing has displaced music sales, it has facilitated a broader distribution which appears to raise awareness about smaller artists and has increased demand for their live

performances. This effect is not seen for more popular artists as their music is usually already widely available, but their music sales are hurt more by file sharing due to the availability of their music. This means the impact of file sharing and live performances is not the same for every artist and that it will be useful to look at different subsets instead of the music industry in general.

Until now we have been discussing the impact of file sharing and new distribution channels on the demand of live performances, this impact has led to an increase in revenue for smaller artists. This revenue consists mostly of ticket sales and revenue related to the production and booking, but also consists of merchandise and (digital) record sales. Digital music services, which aim to sell music or increase streaming, currently rely on recommendation systems to do so. Two important parts of these recommendation systems are user and item profiling (Song et al., 2012). These systems rely mostly on metadata, collaborative filtering and content or contextual data from looking at users´ listening behaviour. User profiling can be important for targeting users as certain groups might show different behaviour and require a different approach. Using a more tailored approach could lead to better results – such as increased sales or usage. Molteni & Ordanini (2003)’s research supports this notion and states that a small, dedicated group could be used as indication for a larger market and thus reduce the time needed for market research or allow for different marketing strategies such as co-branding or organizing virtual communities to increase loyalty.

An interesting question is if live performance data can also be integrated in these recommendation systems to boost sales or usage to introduce a reciprocity system where digital music boosts demands for live performances and live performances increases sales or usage. While little research has been done on this front and therefore is the underlying question of this research – researching the impact of live performances on digital music behaviour – artists have tried to keep ticket prices relatively low in the past to increase attendance and a consequence boost the sales of physical CDs (Koster, 2008). A new development in this area highlighted by Koster is that concert promoters are starting to become record labels. Curlen & Moreau (2009) suggest the opposite, namely that record labels should get into the concert industry to be able to get a share of the generated revenue and argue that this could also be valuable for artists as they can gain additional exposure and promotion. Merging these two industries within the music industry could be beneficial for the industry’s income and could lead to increased interest and development of such a renewed reciprocity system.


(7)

By looking at existing theory we have gained a clearer picture of the relation between the music industry, live performances and the impact of digital music and its possibilities. We will now discuss the method that was used to conduct the research.

3. Research

To be able to answer our research question ‘Do live performances impact digital listening behaviour?’ we need to collect data about live performances and data from digital music services about users’ listening behaviour. For this research we will retrieve most of the required data from a single source with some additional data from a secondary source.

The main source of data is Last.FM . This is an online service that tracks what users listen to 12

on platforms such as iTunes, Spotify and Google Music by using a plugin. It uses this information to provide users with insights and recommendations for music and events (concerts and festivals). The service has tracked over 43 billion song plays and is active in multiple countries. Because last.FM provides recommendations for concerts it also has a large dataset of previous concerts, which makes it an excellent source of data for this research as it includes all the data that is needed. Our secondary source of data is the MusicBrainz database to collect data about a songs’ release year when this is 13

missing from the Last.FM data.
 Analysis

Measuring the impact of live performances on digital listening behaviour requires comparing data from multiple timeframes. We have chosen to compare the data in the week before, during and after a live performance as this would measure any immediate impact. We will look at data of the top 500 artists from one city to measure the impact of a live performance on its direct environment. With respect to our sub question ‘Is there a difference, regarding the previous three questions, in terms of popularity,

genre and how long an artist has been active?’ we will categorize the data in three ways where

popularity is measured by the average number of listeners to an artist in the three analysed weeks. These three categories will be divided in several clusters to possibly discover differences in the impact of live performances. An example of this is that the genre category will be divided into different genres and the active years category in clusters such as ‘artists that started before 2000’. The popularity category will be divided in clusters such as ‘Artists with an average number of listeners of 200 till 600.

With the aid of SPSS software by IBM we will analyze each of these clusters for significant 14

differences in the average number of listeners, the average number of songs between the chosen timeframes and look if more recently released songs are more popular than others.

Last.FM 12 http://www.last.fm/about MusicBrainz
 13 http://musicbrainz.org/doc/About SPSS by IBM
 14 http://www-01.ibm.com/software/uk/analytics/spss/

(8)

Collecting data

Collecting the data that is needed for analysis from the chosen sources is done by utilizing the application programming interface (API) they provide for researchers and developers. Both the API of Last.FM and MusicBrainz provide data through their API in XML format. Data is collected by using multiple Python scripts which utilize external modules such as Pattern for parsing the XML and 15

isoweek for converting dates. Multiple Python scripts are used to increase performance, to allow 16

future changes and make it easier to identify the steps that were taken to collect the data. The method consists of three steps to collect all the data that is needed.

The first step is to collect basic data about artists and songs such as the amount of plays in a specified timeframe. This data is provided in timeframes of one week. This data is processed to only include data of the relevant top 500 artists. The second step is to determine which of these artists have performed in the chosen timeframe and location and then store this data in seperate text files. The last step consists of gathering additional information about artists that have performed in the chosen timeframe and location. This additional information includes genre, year of formation and the release year of individual songs. This information is combined with the basic data and stored in tabular (.csv format) files. These files need to be combined manually to be able to conduct analysis in SPSS. The reason for this is that the Python scripts couldn’t do this automatically and due to the nature of this research no resources were available to implement this feature without any errors. This manually curation is also needed to determine the genre of an artist to ensure valid genres and reasonable sized clusters.

Discussion

We have used data that was provided by Last.FM for our research with some additional data from MusicBrainz. While this data is clearly formatted, it might have been beneficial to store the data in a new structure. This would have allowed us to explore the dataset more without manually adjusting the data as the new structure would be fit for analysis purposes. Providing an own data structure also allows for further expanding of the research and integration in other services when a standardized format such as XML or JSON is used. Last.FM provides a solid stream of data but with a new data structure we could integrate other data sources more easily to gain more insights.

We will now analyse the results that are collected from conducting the described method before discussing the research and drawing our conclusions.

Pattern Python module
 15

http://www.clips.ua.ac.be/pattern

Isoweek Python module
 16

(9)

4. Results


We have chosen to collect data from users that are located in London, United Kingdom in 2014. The reason for this is that it can be considered a major city and is the capital of the United Kingdom. This means there is usually more data available through the Last.FM API when compared to smaller cities and can be compared to other capitals in the future. 2014 is chosen as it is the most recent year we could analyze and would give the most up to date results as the digital music landscape is constantly changing.

Collected data

To determine the 500 most popular artists in the chosen city we have collected and processed 301 weeks of data between 1 February 2009 and 2 November 2014. We have identified 1529 unique artists from this dataset of which the top 500 were used in this research, based on their cumulative number of listeners in this period. As Last.FM’s API only returned data until 2 November 2014 we only analyzed artists that performed in the chosen city before 2 November 2014 and of which enough song data was collected to conduct comparisons between the chosen timeframes. This has lead to a final number of 96 artists with a combined amount of 511 songs.

These group of artists can be considered diverse as they range greatly in popularity and the number of active years across multiple genres. Some artists have an average number of 200 listeners in a week while the most popular artists have more than 1500 listeners on average in a week. The dataset includes artists that have been active from around 1960 to artists that started after 2010. Genres that have been identified are Hip-Hop, Pop, Rock or Electronic for example.

Clusters

Table 1 (below) shows the clusters that were made for each of the three categories that will be used in the analysis. Clusters are not entirely equal in size due to the fact that for some artists more songs were found. The Popularity category consists of six clusters that include 511 songs. The clusters are based on the average number of listeners of an artist in the week before, during and after a live performance. An example of this that the first cluster (200-600) consists of artists that have between 200 and 600 listeners on average and holds 103 songs (N = 103).

The Genre category consists of seven clusters with each cluster representing a different genre and including 511 songs in total. An example of this is that the Electronic cluster consists of artists of which their music has been identified as ‘electronic music’ and holds 91 songs (N = 91). The Drum and Bass cluster will not be used in this research due to its small size.

The last category is ‘Active Years’ and consists of six clusters that hold 451 songs. This is a slightly lower number than the other two categories (511 songs) due to missing information data about when an artist became active. The clusters represent the number of active years of an artist. An example is that the cluster 1964-1994 represents artists that started between 1964 and 1994 and holds 98 songs (N = 98).

(10)

Table 1. Clusters and their size based on Popularity, Genre and Active Years

Analysis of clusters


Each of the clusters consists of three weeks that will be compared in pairs. These pairs are

‘week before – week during live performance’, ‘week during – week after live performance’ and ‘week before – week after live performance. To check for significant differences in these pairs we will use a paired or independent samples T-Test, depending on the data, and an alpha level of α = 0.05.

4.1. Measuring the impact of live performances on the average amount of

listeners

The first sub question that we will answer is ‘Do live performances cause more people listen to songs

of the performing artist?’. This will be done for each of the three categories’ clusters by looking at the

average number of listeners in the chosen weeks and comparing the pairs by using an independent t-test. An example of this is that the first cluster in the popularity category (artists with an average of 200 to 600 listeners) has an average amount of 177.51 listeners for each song in the week before their live performance and where most songs have 164 listeners, also called the median value. Every artist in this cluster has an average of 2.2 songs that are being listened to. This can be found in table 2 (below).

Category Clusters (N = amount of songs) Popularity (Average listener amount) 200-600 (N = 103)

600-821 (N = 110) 821-1019 (N = 99) 1019-1285 (N = 104) 1285-1950 (N = 95)

Genre drum and bass (N = 2)

electronic (N = 91) hip-hop (N = 8) indie (N = 219) pop (N = 47) rock (N = 86) singer-songwriter (N = 58) Active Years 1964-1994 (N = 98) 1995-1999 (N = 85) 2001-2005 (N = 93) 2006-2008 (N = 100) 2009-2014 (N = 75)

(11)

4.1.1 Popularity

Table 2. Average amount of listeners for each song where m stands for median and a for the average of the amount of songs for an artist. (Popularity)

In table 2 we can see two phenomena in the popularity category. Each of the clusters show an initial decline in the week during the performance, compared to the week before. Two different effects are seen when comparing the week after the performance to the week during the performance. Cluster 3 (821-1019) and cluster 5 (1285-1950) show a further decrease but the other clusters show an increase over the last week, which in the case of cluster 2 (600-821) and cluster 4 (1019-1285) result in a higher average amount of listeners for each song than any other week, resulting in a ‘V-Shape’ when visualized.

Table 3. Significance levels (Popularity) where significant differences are bold.

The observed behaviour is only significant in two of the clusters, cluster 3 (821-1019) and 5 (1019-1285), as can be seen in table 3. Significant differences are only found when comparing the week before the performance to the week after it for both of these clusters.

week before week during week after

200-600 177.51 (m = 164, a = 2.2) 162.71 (m = 163, a = 1.41) 172.53 (m = 150, a = 1.15) 600-821 180.06 (m = 167, a = 4.00) 177.00 (m = 156, a = 3.05) 197.86 (m = 171, a = 3.14) 821-1019 244.17 (m = 203, a = 8.64) 220.93 (m = 176, a = 6.64) 206.51 (m = 173, a = 5.00) 1019-1285 295.35 (m = 230, a = 10.33) 274.57 (m = 252.5, a = 8.44) 308.59 (m = 237, a = 8.89) 1285-1950 404.08 (m = 279, a = 11.38) 342.05 (m = 278, a = 11.00) 300.49 (m = 249, a = 10.63)

week before - week during week during - week after week before - week after

200-600 .808 .186 .623

600-821 .758 .143 .181

821-1019 .201 .420 .035

1019-1285 .393 .191 .635

(12)

4.1.2. Genre

Table 4. Average amount of listeners for each song where m stands for median and a for the average of the amount of songs for an artist. (Genre)

Clusters in the genre category show similar behaviour as the clusters in the popularity category, with the exception of the electronic and hip-hop clusters. These two clusters see an increase in the average amount of listeners for each song in the weeks during and after the performance. Indie, pop and singer-songwriter music see a decrease over two weeks and rock music is the only cluster to show the previous observed ‘V-Shape’ behaviour.

Table 5. Significance levels (Genre) where significant differences are bold.

As we can see in table 5 only electronic, indie and singer-songwriter music show a significant difference. The increase in the average amount of listeners for each song for electronic music is significant when comparing the week after the performance to the week during and before the performance. Hip-Hop music doesn’t show a significant difference even though it shows similar behaviour as electronic music. This could be explained due to the small size of the cluster as shown in table 2.

week before week during week after

electronic 213.53 (m = 188, a = 3.35) 227.03 (m = 187, a = 3.41) 291.49 (m = 223, a = 4.88) hip-hop 157.13 (m = 142.5, a = 1.60) 166.75 (m = 130, a = 0.8) 182.25 (m = 138, a = 0.8) indie 314.96 (m = 207, a = 6.94) 301.70 (m = 257, a = 5.13) 260.66 (m = 222, a = 4.55) pop 256.97 (m = 201, a = 2.60) 252.10 (m = 191, a = 2.60) 250.34 (m = 177, a = 2.53) rock 197.45 (m = 165, a = 5.20) 169.75 (m = 146.5, a = 3.47) 174.11 (m = 151, a = 3.13) singer-songwriter 212.76 (m = 188.5, a = 8.29) 187.69 (m = 172, a = 5.57) 172.10 (m = 157, a = 3.00)

week before - week during week during - week after week before - week after

electronic .438 .008 .001 hip-hop .795 .821 .557 indie .546 .030 .008 pop .892 .961 .857 rock .066 .778 .131 singer-songwriter .091 .243 .011

(13)

4.1.3 Active Years

Table 6. Average amount of listeners for each song where m stands for median and a for the average of the amount of songs for an artist. (Active Years)

Clusters in the active years category, where artists are clustered based on their amount of active years, show the same behaviour as those in the popularity category with the exception of the last cluster (2009-2014). Cluster 1 (1964-1994), cluster 3 (2001-2005) and cluster 4 (2006-2008) show a decrease in the average amount of listeners for each song. Cluster 2 (1995-1999) shows the ‘V-Shape’ behaviour and Cluster 5 (2009-2014) shows an increase in the amount of listeners.


Table 7. Significance levels (Active Years) where significant differences are bold.

Table 7 shows that that only cluster 1 (1964-1994) and cluster 4 (2006-2008) show significant differences. These differences signify a decrease in the amount of listeners when compared to another week. Two significant differences have been observed for cluster 4 and just one for cluster 1. 4.1.4 Conclusion

We observed several patterns in the impact of live performances on the average amount of listeners for songs of an artist. Most of the clusters show a decline in the amount of listeners. Only a few clusters show an increase in the amount of listeners, such as electronic music in the genre category or artists that started between 2009 and 2014 in the active years category. This can been seen in table 4 and 6. A different pattern is the ‘V-shape’, which means there is an initial decrease of the amount of listeners before increasing again. In some cases this even leads to a higher amount of listeners than the week before the performance. As the patterns are not the same for each cluster we can say that the impact of a live performance depends on genre, popularity and the amount of active years. The observed impact however is not significant for every cluster, which mean this impact might be neglected.

week before week during week after

1964-1994 194.83 (m = 173, a = 5.94) 178.60 (m = 160.5, a = 3.63) 164.57 (m = 144, a = 2.31) 1995-1999 223.67 (m = 167, a = 5.15) 207.15 (m = 168, a = 4.54) 268.36 (m = 173, a = 4.54) 2001-2005 284.68 (m = 208.5, a = 6.00) 268.66 (m = 204.5, a = 4.67) 259.12 (m = 217, a = 3.93) 2006-2008 382.50 (m = 246, a = 5.88) 335.78 (m = 268, a = 4.94) 277.31 (m = 238, a = 5.13) 2009-2014 234.95 (m = 201, a= 5.00) 237.33 (m =208, a = 3.91) 247.68 (m =219, a = 5.18)

week before - week during week during - week after week before - week after

1964-1994 .175 .230 .032

1995-1999 .426 .052 .164

2001-2005 .550 .723 .348

2006-2008 .222 .035 .003

(14)

4.2 Measuring the impact of live performances on the amount of songs

The next sub question that we will answer is ‘Do live performances cause people to listen to more

songs of the performing artist?’. This will be done for each of the three categories’ clusters by looking

at the average amount of songs for an artist in the chosen weeks and comparing the pairs by using a paired samples t-test. An example is that people listen to an average of 2.2 songs of an artist in the first cluster (200-600 average listeners) in the week before their live performance as can been seen in table 8.

4.2.1. Popularity

Table 8. Average amount of songs for each artist (Popularity)

Table 8 shows the same behaviour as was observed in section 4.1.1 when looking at the amount of songs that are being played for each artist. We see the ‘V-shape’ behaviour in cluster 2 (600-821) and cluster 4 (1019-1285) while cluster 1, 3 and 5 are decreasing. A different effect is seen for the first cluster, which does not see the ‘V-shape’ behaviour like it does in section 4.1.1, but instead shows a further decrease.

Table 9. Significance levels where significant differences are bold (Popularity)

The significant differences that are found when looking at the amount of songs being played for each artist both mean a significant decrease in the amount of songs. These are found in cluster 1 (200-600) and cluster 3 (821-1019). Cluster 1 shows a significant decrease when comparing the week after the performance with the week before it. Cluster 3 shows a significant decrease when comparing the week during the performance with the week before.

week before week during week after

200-600 2.2 1.41 1.15

600-821 4.00 3.05 3.14

821-1019 8.64 6.64 5.00

1019-1285 10.33 8.44 8.89

1285-1950 11.38 11.00 10.63

week before - week during week during - week after week before - week after

200-600 .130 .161 .029

600-821 .403 .931 .397

821-1019 .049 .170 .093

1019-1285 .265 .669 .543

(15)

4.2.2 Genre

Table 10. Average amount of songs for each artist (Genre)

When looking at changes in the amount of songs for each artist in the genre category we see similar behaviour as in section 4.1.2. Electronic music sees an increase in the amount of songs whereas singer-songwriter, rock and indie music see a continuing decrease. Hip-Hop and Pop music also see a decrease in the amount of songs but also show signs of stability in different weeks. The ‘V-shape’ behaviour is not observed for any genre.

Table 11. Significance levels where significant differences are bold (Genre)

No significant differences have been founding for any genre when looking for changes in the average amount of songs played for each artist, meaning that the amount of songs remains stable and live performances do not have an impact on the amount of songs being played.

week before week during week after

electronic 3.35 3.41 4.88 hip-hop 1.60 0.80 0.80 indie 6.94 5.13 4.55 pop 2.60 2.60 2.53 rock 5.20 3.47 3.13 singer-songwriter 8.29 5.57 3.00

week before - week during week during - week after week before - week after

electronic .750 .177 .170 hip-hop .294 1 .242 indie .140 .409 .066 pop 1 .719 .943 rock .279 .207 .229 singer-songwriter .502 .200 .157

(16)

4.2.3 Active Years

Table 12. Average amount of songs for each artist (Active Years)

More differences in the amount of songs for each artists are found when looking at the amount of active years of an artist. Cluster 1 (1964-1994) and Cluster 3 (2001-2005) show a continuing

decrease while cluster 2 (1995-1999) shows an initial decline but remains stable in the week after the performance. Cluster 4 and 5 show the ‘V-Shape’ behaviour of which cluster 5 sees a higher amount of songs when compared to the other weeks.

Table 13. Significance levels where significant differences are bold (Active Years)

While the clusters in the active years category show more diverse behaviour, most of these are not of significant nature. Only cluster 3 (2001-2005) shows a significant decrease in the amount of songs for an artist when comparing the week during the performance with the week after it.

4.2.4 Conclusion

When looking for an impact of live performances on the amount of songs that are being played we have found different patterns between popularity, genre and the amount of active years of an artist. Findings are similar with those of section 4.1, where the majority showed a decreasing effect. A new observation is that in some cases the amount of songs remain stable, meaning that a live

performance doesn’t have any impact on the amount of songs. This is found when looking at genre and the amount of active years. With regard for the significant levels, less significant differences are found in the amount of songs for an artist. This means the amount of songs that are being played of an artist remain relatively stable compared to the amount of listeners as shown in section 4.1.

week before week during week after

1964-1994 5.94 3.63 2.31

1995-1999 5.15 4.54 4.54

2001-2005 6.00 4.67 3.93

2006-2008 5.88 4.94 5.13

2009-2014 5.00 3.91 5.18

week before - week during week during - week after week before - week after

1964-1994 .270 .135 .053

1995-1999 .731 1 .740

2001-2005 .474 .044 .274

2006-2008 .326 .861 .270

(17)

4.3 Measuring the impact of live performances on recently released songs

The last sub question to answer is ‘Do live performances cause people to listen more to recently released songs of the performing artist than older songs?’. This will be done for each of the three categories’ clusters by looking at the amount of songs in a certain timeframe. An example of a timeframe is ‘songs released before 2000’. We will look in which timeframe the most songs occur and if this changes in the weeks during and after the live performance. An illustration of how this is

represented in the tables that are used is that in the the first cluster (200-600 average listeners) of the popularity category seven songs are being played from before 2000 in the week before the live performance and just two in the week after the performance. This can been seen in table 14. 4.3.1. Popularity

Most of the songs that are being played can be considered recent as these songs are released from 2010 or later. A large part of these recent songs is even released in the chosen year 2014. We see a similar effect as in section 4.2.1 where the amount of songs is decreasing for most of the clusters. An interesting pattern is that more popular artists, in the last two clusters, rely more on their newer songs than less popular artists. This is illustrated in table 14 when looking at the timeframe ‘before 2000’ and ‘between 2000 and 2005’. These clusters also show stable behaviour for songs released in 2014 where the other clusters show a decrease.

Table 14. Amount of songs in a specified timeframe for each cluster (Popularity)

4.3.2 Genre

Electronic music was the only genre to display an increase in the amount of songs in section 4.2.2 and table 15 (below) shows that this increase is caused by songs released after 2005 in the week after a live performance. Other genres see a continuing decrease but also rely more on recently released songs and for a big part on songs that were released in 2014. An interesting observation is all of the songs in the Pop genre are from 2010 or later. This can be explained by the nature of Pop music where people only tend to listen to new released music that is being played on TV, radio or is featured in popular charts.

< 2000 2000 - 2005 2005 - 2010 > 2010 2014 200-600 7 3 2 2 0 0 5 3 3 48 41 29 35 31 20 600-821 2 1 1 3 2 0 7 7 4 65 50 44 40 26 25 821-1019 9 4 2 5 1 2 7 5 2 63 59 44 36 35 31 1019-1285 0 0 0 0 0 0 9 2 1 61 55 53 30 30 30 1285-1950 0 0 0 1 2 2 5 4 3 71 68 67 33 33 33

(18)

Table 15. Amount of songs in a specified timeframe for each cluster (Genre)

4.3.3 Active Years

If we look at the amount of active years of an artist we see similar behaviour as when looking at popularity or genre. Most artists in these clusters depend on recently released songs and show a declining pattern, as can been seen in table 16. A large part of the songs is released in 2014. Artists that started between 1964 and 1994 however also rely for a big part on older songs that were released before 2000. These songs still show the decline in listeners, but this is seen for songs in every timeframe and thus remain relatively important. One of the clusters (2006-2008) show the ‘V-shape’ behaviour in each of the timeframe of which songs were analysed.

Table 16. Amount of songs in a specified timeframe for each cluster (Active years)

4.3.4 Conclusion

Artists generally rely more on recently released songs but show an overall decrease in the amount of songs, meaning that people listen to less songs of an artist after a live performance when compared with the week before the performance. Artists that have been active longer (starting from 1964 to 1994) also rely on their older songs for a substantial part. The overall decrease is seen in every cluster of each category and we can thus say that live performances do not cause people to listen more to recently released songs compared to older songs. Now that we have answered all sub questions we will discuss our hypotheses and the outcomes of the research before drawing our conclusions. < 2000 2000 - 2005 2005 - 2010 > 2010 2014 electronic 1 1 1 1 1 1 2 2 4 42 43 43 23 24 25 hip-hop 0 0 0 0 0 0 1 1 1 5 2 1 2 1 1 indie 7 2 1 3 1 0 15 11 4 150 130 119 89 74 73 pop 0 0 0 0 0 0 0 0 0 30 31 29 14 15 13 rock 3 2 2 5 3 3 10 3 2 38 30 28 23 20 17 singer-songwriter 7 3 1 1 0 0 6 4 2 38 30 16 20 16 10 < 2000 2000 - 2005 2005 - 2010 > 2010 2014 1964-1994 15 8 5 3 2 1 6 3 1 45 36 22 22 17 11 1995-1999 1 0 0 6 2 3 7 6 3 47 48 41 28 27 25 2001-2005 0 0 0 0 0 0 12 5 2 51 45 39 21 22 20 2006-2008 0 0 0 0 0 0 4 2 3 79 68 70 51 43 46 2009-2014 0 0 0 0 0 0 1 1 1 48 38 38 33 23 22

(19)

5. Discussion

Our research has provided insights on the impact of live performances on users’ listening behaviour via digital services by using data from Last.FM’s users in London, United Kingdom in 2014 that was collected with the aid of several Python scripts. We looked at differences between popularity, genre and the amount of active years of an artist in the week before, during and after a live performance. This way we have identified several patterns.

Our first hypothesis was that more people would listen to songs of an artist after a live performance. Analysis has shown that the opposite is true and that less people listen to songs of an artist after their live performance for most cases, but differences have been found between genre, popularity and the amount of active years. Artists that started between 2009 and 2014, which can be considered new artists, see an increase in the amount of listeners in the week after the performance. This is the same for artists in the electronic music genre. An interesting pattern is the ‘V-shape’ behaviour, this means there is an initial decline in the amount of listeners in the week of the

performance but increases again in the week after the performance. Most observed changes however are not significant which means we can not say for certain that the observed impact of live

performance is valid and might be normal fluctuations in listening behaviour.

Our second hypothesis was that we do not expect any impact on the amount of songs that are being listened to of an artist. We have found an decrease in the amount of songs in the weeks during and after the performance compared with the week before when looking at differences

between popularity, genre and amount of active years in most cases. These changes are however not significant for most of the analyzed clusters which means our hypothesis that live performances do not cause a change in the amount of songs that are being played of an artist is true.

Finally, our last hypothesis was that we expect that people will listen more to recently

released songs after a live performance. Results has shown however that this is not the case. People do listen more to recently released songs and this is independent of genre, popularity or how long an artist is active but is not being impacted by live performances.

An interesting point of discussion is that the results that we have found might be influenced by the scope that we used in this research. We have chosen to analyze data from one city in three specific weeks. Expanding this scope to multiple cities or a nation could already see different results as live performances also attract people from other places. Another interesting aspect is that we do not know the behaviour of listeners in other weeks. Seeing the bigger picture might show other insights than this research does as we have focussed on the immediate impact of live performances.

(20)

Conclusion

This research can be considered a first step in understanding the impact of live performances on digital music behaviour. Before physical CD sales begun to fall live performances were used to increase music sales and this behaviour was hypothesized to also take place in the age of digital music services. We called this effect a reciprocity system. We analyzed user behaviour by using data from Last.FM. For most of the analysed cases that showed a significant difference this was not the case and in fact showed the opposite effect. Analyzing user behaviour could help artists to drive music sales or to tailor their live performance with the goal to improve their income to make up for the decreasing royalties. This research has used a relatively small dataset and could benefit from using a larger dataset to further explore the effects of live performance on digital music services, which is possible with adjusting the used method for data gathering and analysis slightly.

6. Future work

This research can be seen as a foundation for further research as we have identified initial patterns and problems when analyzing users their listening behaviour via digital music services. A logical to step to validate the found results is to expand the scope of the research and compare several groups with each other. While this research is limited and statistical analysis had to be partially done manual, a new data structure could be developed to automate this process and would allow more sources for additional data. Using more sources than just Last.FM can lead to better insights and this is already illustrated by the fact that we used MusicBrainz for limited data. An example of expanding the scope is to not just look at one week before or after an performance but several weeks. Data could be

collected from a country or at least multiple cities in a country instead of one city and possible be compared with other countries to discover possible differences between location or culture.

(21)

7. References

Curlen, N. & Moreau, F. The Music Industry in the Digital Era: Toward New Contracts. Journal of Media Economics (22):2. 2009. Last accessed on 3 march 2015 via:

http://www.tandfonline.com/doi/full/10.1080/08997760902900254


DiCola, C. P. Money from Music: Survey Evidence on Musicians’ Revenue and Lessons

About Copyright Incentives. Arizona Law Review. 2013. Last accessed on 18 February 2015

via:


http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2199058##

Koster, A. The Emerging Music Business Model: Back to the Future? Journal of Business Case Studies (4):1. 2008. Last accessed on 20 February 2015 via:


http://www.cluteinstitute.com/ojs/index.php/JBCS/article/view/4812/4904

Molteni, L. & Ordanini, A. Consumption Patterns, Digital Technology and Music

Downloading. Long Range Planning (36):4. 2003. Last accessed on 3 March 2015 via:

http://www.sciencedirect.com/science/article/pii/S0024630103000736


Mortimer, J.H., Nosko, C. & Sorensen, A. Supply responses to digital distribution: Recorded music and live performances. Information Economics and Policy (24):1. 2012.

Last accessed on 20 February 2015 via:


http://www.sciencedirect.com/science/article/pii/S016762451200008X#

Song, Y., Dixon, S. & Pearce, M. A Survey of Music Recommendation Systems and Future

Perspectives. 9th CMMR. 2012. Last accessed on 20 February 2015 via:


http://cmmr2012.eecs.qmul.ac.uk/sites/cmmr2012.eecs.qmul.ac.uk/files/pdf/papers/ cmmr2012_submission_42.pdf

Schultz, M.F. Live Performance, Copyright and the Future of the Music Business. 2009.

Last accessed on 21 February 2015 via:


http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1337914

Zentner, A. Measuring the Effect of File Sharing on Music Purchases. Journal of Law

Economics (49):1. 2006. Last accessed on 21 February 2015 via:


(22)

8. Appendix

Due to the size and length of the data and the Python scripts that were used in the research, We have chosen to not include these directly into the thesis itself. The following data can be downloaded as a .zip file at bsc.berndvdw.com/Research.zip:

■ Python scripts used in this research

■ raw data, collected through the API with the Python scripts ■ processed data

Referenties

GERELATEERDE DOCUMENTEN

proces waarbij een chemische stof via een bioche­ misch of ander mechanisme wordt getransformeerd tot een metaboliet die giftiger is dan de stof zelf combinatiewerking van twee

Indien investeerders niet in staat zijn de bias in de verwachtingen te herkennen, weten zij niet wat de verwachtingen inhouden en of een positieve earnings surprise

Cognitive underload can lead to (passive) fatigue, which has been shown to result in disengagement from the task and higher distractibility and can subsequently degrade

Researchers often assume that metacognitive activities mediated the effect of scaffolding on learning, but there is little empirical evidence for this assumption (Veenman et

reg GapinTotalReturnsHK4Q recessionHKLag2 CrAvGrHKLag2 RiskAversionHK... reg GapinTotalReturns4Q Recession4Q

Data Ethnography Interviews Participant observation Documents/ artifacts Reflexivity Audio-visual registrations Sensitizing concepts... ‘Doing being shanty choir’; hoe een

Different kinds of measurements have been performed: steady and unsteady fuselage pressure distributions, steady and unsteady loads and moments and rotor trim angles, and