8.2.1 Information, Demand, and Outcome Distributions

Banerjee (1992) and Bikhchandani et al. (1992) are the seminal herding papers that show how demand can be affected by previous customers’ choices. To make the process of dynamic demand more transparent we will follow the intuition set out in De Vany and Walls (1996). They assume for simplicity a single title that can be consumed sequentially (and not simultaneously) by a population of n potential customers. The particular sequence in which consumers make their choices is random and they have a common prior belief in the quality of the title. Denote the probability that a randomly chosen demander will choose to purchase and consume the title p. Letting X denote the number of people choosing to consume the title, then X is a binomial random variable; it follows that when consumers share a common prior over quality that the distribution of a title’s sales would be binomial. However, when quality is unknown and priors over quality differ among potential customers, p is a random variable. By conditioning on p and integrating over the binomial distribution, each of the n + 1 possible outcomes is equally likely; adding uncertainty to the priors transforms the distribution of revenue from the binomial to the uniform distribution.

Now consider information sharing where potential consumers can use information revealed over time to refine their prior on a title’s quality.² De Vany and Walls (1996) let the distribution of customers over screens be multinomial uniform, so the search problem – a search for quality with an unknown distribution – is similar to the search for price with an unknown distribution. Customers who do not know the distribution begin with a uniform prior and update the distribution sequentially. The result of this process is the Bose–Einstein distribution, which has the property that all of the possible outcome vectors are equally likely! This means the vector in which consumption of every title is equal to 0 is as likely as one in which all n trials go to only one title and every other vector is equally likely (Feller, 1957). The Bose–Einstein distribution has uniform mass over a space of possible outcome vectors, including the outcome where consumers choose to purchase nothing.

What is important about the evolution of choice probabilities under the Bose–Einstein choice logic is the way past successes are leveraged into future successes: as soon as individual differences emerge among the films, they are compounded by information feedback through the run into very large differences. Substantial promotional activity with large supply can produce large and rapidly growing sales, but it also can lead to early failure if the large number of consumers transmit negative information. As customers sequentially select titles, the probability that a given customer selects a particular title is proportional to the fraction of previous customers who selected it. This leveraging of success results when probabilities are not known and sampling reveals information that causes previous selections to attract subsequent selections.

There are several additional interesting issues that can be explored in the context of models of information sharing. For example, Chamley and Gale (1994) have an endogenous element in their model where consumers may wait to consume to see what others think first; this has a particular appeal to the extent that consuming a title is an investment rather than pure consumption, and in this context the paper is also related to the Ottaviani and Sorensen (2000) paper on herding and investment. Avery et al. (1999) also present a model in which the order of the choice sequence is endogenous. Sgroi (2002) investigates the strategy of focusing effort on early consumers. Smith and Sorensen (2000) examine confounded learning in a herding context as an alternative to the rapid success or failure in the older herding models. De Vany and Lee (2001) investigated the role of information in a simulation model, and showed that even with noisy information better movies usually win and that payoffs have the winner-take-all or heavy tails property. De Vany and Walls (1996) showed that box office revenues have a contagion-like property where the week-to-week change in demand is stochastically dependent on previous demand.

8.2.2 The Roles of Talent and Luck in Success

Rosen (1981) and Adler (1985) provide the now-classic analyses of superstardom. Adler (2006) in volume 1 of this Handbook provides a survey of superstar models in relation to talent. In these models there are a few ways that the talent of a performer can be related to whether or not the performer’s output becomes a bestseller or blockbuster.

In Rosen’s (1981) model consumer demand is characterized by substitution – so consumers will pay more for higher quality services or will substitute quantity for quality. Also, there are known differences of talent across performers and performers compete with one another on two margins: other performers may be equally talented or other performers maybe become substitutes at a given price or type of venue. Rosen simplifies things by assuming that quality does not deteriorate with quantity provided, but even in this simplified case we see that total revenue will increase with quality because both price and quantity provided increase with talent. Thus, in the Rosen model total revenue increases in talent more than proportionally and implies that revenues will be concentrated on the most talented people. Rosen further shows that if the services provided by more talented sellers are less affected by competition from other sellers, then even if there is more than one provider in the market it will still be the case that higher quality providers service larger crowds and the skewed distribution of success across performers will continue to hold.

Adler (2006) discusses competing models that relate talent to the distribution of success. In particular, he considers the model of MacDonald (1988) in which talent is defined in a different way. MacDonald, instead of focusing on the certain talent of a performer, focuses on the volatility of a performer’s talent. In his model, each performer can deliver good and bad performances, and the difference in talent is the probability that a particular performance will be a good one. Consumers value talent such that the probability of a good performance is higher for experienced performers than for inexperienced performers because only good performers will survive. Through this sort of selection, performers with a record of success will command higher ticket prices and supply larger audiences. In these circumstances, artists of equal talent do equally well.

Adler (1985) also describes another alternative model that can explain the emergence of a superstar performer. His model differs from the Rosen model because the star emerges from among several equally talented performers due to pure luck. The process is as follows. Consumers select a performer randomly in their purchase decisions, so it is only by luck that one performer supplies more consumers than another performer. Luck makes the superstar performer the most popular and because consumers have a preference for popular performers this initial advantage, which is purely the result of chance, is leveraged into further success. Adler notes that if the performers are knowledgeable of consumers’ selection mechanism, they may try to influence their perceived popularity rather than leaving it up to nature’s lottery.

8.2.3 Innovation Tournaments

De Vany and Walls (1997) propose using a tournament model as a way of understanding competition in the movie market. In their depiction of the movie tournament, each film must earn a critical level of box office revenue to survive to the next stage of competition. If films fail to earn this threshold level of revenue, exhibitors will choose to drop the film and replace it with another film that has a higher level of expected revenue. The revenue threshold will change in response to audience changes, over the course of a film’s theatrical lifetime, and as other competing films enter the tournament.

Books, movies, and music can all be modeled as if they are competing for the top positions in a tournament. In fact, weekly lists of sales rankings are reported in precisely this way, sometimes on the evening news. Suppose there are m(t) > 10 titles competing at time t for these 10 tournament positions. Rank sales revenues at time t from high to low in the m-dimensional vector r =  (r₁(t), r₂(t), …, r_m(t)), where low numerical ranks indicate high sales revenues. The threshold revenue that a title must earn to survive from week t to t + 1 is a random variable C(t) = r₁₀(t). A title i in release at time t survives if its revenue exceeds the critical level, r_i(t) ≥ C(t), and it is eliminated from competition otherwise. Every title in the top 10 sample meets this condition by definition. When a title does not meet the condition, it is dropped from the sample to be replaced by a competing title. A title may move up or down in rank during the course of its lifetime. The length of time that a title remains within each rank is a random variable.

If a title gains in popularity its rank will increase over time, but its revenues eventually must fall as it exhausts its potential audience. As it declines in rank, a title becomes vulnerable to newly released competitors and we might expect that its occupancy time in rank becomes more brief. This intuition of an innovation tournament where new titles enter the tournament as older titles are dropped corresponds well with statistical survival models as discussed toward the end of Section 8.3.

8.3.1 The Pareto Model (and Departures From It)

Now we demonstrate the observable implications of various types of demand dynamics discussed above on the distribution of sales across titles. This provides a correspondence between the conceptualization of demand and the testable implications to be confronted with data in the empirical analysis. The idea here is to examine the distribution of success across products and to test competing characterizations of the demand process. The approach taken here is discussed at some length by Brock (1999), who argues that empirical regularities in the form of scaling laws, such as the Pareto law, may be useful in providing information on the process that is generating the data.³ The Pareto law as used to characterize the distribution of success – across competing books, films, or songs – is similar to recent work in econophysics (Rosser, 2008).

Many empirical regularities in economics (as well as in the physical world) follows power laws (Gabaix, 2008). For example, Steindl (1965) found that the size of a firm S is systematically related to its rank R in an industry according to the Pareto law S × R^β = A, where β and A are constants. In the context of entertainment media consumption, titles are naturally ranked each week by their revenues so the Pareto law implies the following log-linear relationship between weekly revenue and rank: log Revenue = log A − β log Rank. Not only is the Pareto law shown above an empirical regularity for several economic quantities – income, firm size, financial returns – but Simon (1955) showed that it can be derived analytically assuming that the growth rate of revenues is independent of size and that there is a constant rate of entry for new titles. What this means is that in the absence of demand feedback the Pareto law would describe the distribution of outcomes.

The actual empirical distribution can differ from the Pareto law defined in the linear relationship between log Revenue and log Rank, and it is in this way that we can make inference on feedback in the demand process. In studying the distribution of firm sizes, Ijiri and Simon (1974) found that empirical distributions often deviated from the Pareto law: instead of Revenue and Rank following a log-linear relationship as implied by the Pareto law they found that the plot had curvature.⁴ In markets for competing entertainment titles we can interpret deviations from the Pareto law in the following way: an increase in revenues affects future growth through either (i) the information sharing between those individuals who consumed a title and potential purchasers, or (ii) herd- or contagion-like behavior on the part of potential purchasers. Regardless of the behavioral cause of the information feedback onto demand, the effect of the demand increase will ultimately diminish over time due to the saturation of the potential audience and the entry of new competing titles.

In practice one can test for the presence of positive or negative information feedback in the demand process by quantifying the curvature of the Pareto relationship by adding a quadratic term in rank to capture the non-linearity as suggested by Ijiri and Simon (1974): log Revenue = log A − β log Rank + γ(log Rank).² A non-zero value of the parameter γ indicates a deviation from the Pareto law and the sign of this coefficient indicates whether the information feedback is positively or negatively associated with growth in demand. Finding curvature away from the Pareto law such that γ > 0 is an indication of autocorrelated growth in revenues, and this is the observable implication of the increasing returns that can cause some titles to become ‘hits’ and others to become ‘bombs’ through information feedback. Finding that γ < 0 would indicate negative information feedback (or negatively autocorrelated growth) or that titles that did well last period will have falling demand this period. This test for information feedback in demand does not permit us to distinguish between the competing explanations of the source of the increasing returns, but it does permit us to test the hypothesis of information feedback using data generated in actual markets. It has been employed in several recent studies of books, music, the theatrical movie market as well as the market for DVDs. We will discuss these empirical studies in Section 8.4.

8.3.2 Stretched Exponentials

Concavity in log-log plots of size against rank, also known as a parabolic power law, are interpreted as evidence of increasing returns to information in the demand for movies (De Vany and Walls, 1996; Walls, 1997; Hand, 2001). Frisch and Sornette (1997) propose a multiplicative stochastic process that can explain the deviation of the data relative to a power law distribution. LaHerrere and Sornette (1998) show the stretched exponential distribution accounts for many deviations from power law distributions in both the physical and the social sciences. Sornette (1998) provides rigorous technical details on multiplicative processes leading to power laws and stretched exponentials. Let the probability density function of independent and identically distributed (IID) random variables m be p(m). Now consider the product X_n = m₁m₂…m_n, which has probability density function for X → ∞ and finite n. From the relation X_n+1 = X_nx_n+1 we can write the equation for the density function of X_n+1 in terms of the density functions of x_n+1 and X_n, and derive the probability density function by the formal application of the Laplace method (Frisch and Sornette, 1997). The tail of P_n(X) is controlled by the realizations where all terms in the product are of the same order, so that P_n(X) is to a first approximation the product of the probability density functions of the n terms in the product m₁…m_n. When the density function p(·) is exponential, then the density function of the product is a stretched exponential for large n and follows the Weibull cumulative distribution given by . When the exponent α in the Weibull distribution is unity, it is the exponential distribution; when the exponent is less than unity, the tails are heavier than exponential tails and the distribution is often referred to as the stretched exponential.

8.3.3 Black Swans and the Lévy–Stable Model

Taleb (2007), in his bestseller The Black Swan, emphasizes the importance of high-impact low-probability events. He is critical of the Gaussian distribution as a model to predict future events based on a sample of data. Taleb’s black swan idea is not purely statistical, but from our perspective we can view it as arguing that distributions of success outcomes have heavier-than-normal tails. We now discuss some particular models that can be useful in characterizing black swan outcomes.

8.3.4 Skew-Normal and Skew-t Models

Several statistical models have been proposed to account for the non-normal distribution of data from the physical and social sciences. The stable Paretian model was proposed by Mandelbrot (1963a,b) to account for the heavy tails of financial data; other authors have suggested the Student-t distribution with the appropriate degrees of freedom as a natural alternative to the normal distribution. The multivariate Student-t distribution permits regression analysis of film returns allowing for heavy tails, but it does not allow for asymmetry. Several recent empirical papers in finance have employed various statistical models to capture the skewness and heavy tails in the distribution of financial returns using non-standard distributions that make it difficult for applied researchers to condition the distribution on a vector of explanatory variables.⁹ The logarithmic versions of the skew-normal and skew-t distributions have been proposed as models of sales or revenue that explicitly account for skewness, and skewness and heavy tails, respectively, in a multivariate regression-like framework of analysis that is familiar to financial economists and other social scientists.

8.3.5 Non-Parametric Kernel Regression

All of the techniques listed above – even the ones that use distributions accounting for skewness, heavy tails, and infinite variance – assume a functional form for the basic regression model. The validity of each model’s statistical tests and the resulting conclusions are predicated on the regression equation being properly specified. To avoid the model specification problem entirely, one could use non-parametric kernel regression. The distinguishing feature of non-parametric regression, in contrast to a parametric model, is that little or no a priori knowledge is assumed about the form of the true function which is being estimated. The function is still modeled using an equation containing free parameters, but in a way that allows an extremely broad class of functions to be represented. There are different types of non-parametric models, including neural networks. Non-parametric kernel regression is a powerful statistical technique suited to this purpose that can readily be added to the toolkit of applied researchers. The non-parametric kernel regression model used in several movie papers is based on the earlier work of Fan and Gijbels (1992, 1996) on local linear non-parametric regression that has been supplemented by cross-validated bandwidth selection and recent advances in generalized kernel estimation (Li and Racine, 2004; Racine and Li, 2004).

8.3.6 Other Tools and Models

8.4.1 Movies

Recent surveys of the economics of the movie industry are given by Chisholm (2003), in Volume 1 of this Handbook by De Vany (2006), and Walls (2008). A survey with a more management-related focus is provided by Hadida (2009) and Eliashberg et al. (2006) provide a marketing-related survey. A very well-written and thorough survey of empirical work on the economics of the movie industry is provided in McKenzie (2012). Here, we shall focus specifically on empirical findings that relate to the bestsellers and blockbusters.

Many papers on the movie industry have focused on the types of demand processes that are consistent with the observed distribution of product success. De Vany and Walls (1996), Walls (1997), Hand (2001), and Sornette (2002) look at the importance of blockbusters when tails follow a power law. De Vany and Walls (1996) provide evidence of maximal uncertainty in movie outcomes resulting from the way that demand feeds back on itself. The empirical evidence in their paper is based primarily on departures from the Pareto size distribution as discussed in Section 8.3.1: concavity in log-log plots of size against rank, which is known as a parabolic power law, are interpreted as evidence of increasing returns to information in the demand for movies. Numerous authors, including Walls (1997), Hand (2001), Maddison (2004), Akdede and Ogus (2006), and others, have applied the empirical framework from the De Vany and Walls (1996) paper to other samples of data and have confirmed their results of positive information feedback in the movie industry. All these papers base their results on deviations from the Pareto size distribution. The linkage to information feedback is indirect instead of being structural. One can interpret that as a primary weakness to the approach of identifying the process that generated the data by examination of the outcome distribution. De Vany and Lee (2001) investigated the role of information in a simulation model and showed that even with noisy information better movies usually win and that payoffs have the winner-take-all or heavy tails property.

The stretched exponential distribution has also been applied to movie success. Walls (2005a) finds strong evidence that the distribution of movie outcomes does not follow a power law or a even a parabolic power law as proposed by the papers discussed in the paragraph above. He presents empirical evidence from North American box office revenue showing that the stretched exponential distribution provides a statistically superior fit. The stretched exponential distribution does not truncate the upper tail in its estimates of the probability of a movie earning a larger amount than previous movies. The distribution also accounts for the deviation from the strict Pareto power law in a way that does not place artificial restrictions on the possibility that a movie can earn far more than our experience suggests. This is consistent with a model of consumer demand where shocks are multiplicative and there is decay in information transmission that leads to local audience saturation.

In related work, De Vany and Walls (1999, 2002) fit the Pareto distribution – not the rank-order size distribution – and find evidence of infinite variance in the outcome distribution. Sinha and Raghavendra (2004) find evidence of long tails using weeks in top 60, opening week gross, and cumulative lifetimes in theatrical exhibition. These empirical results do not correspond directly to a behavioral model in the way that deviations from a Pareto size distribution are consistent with increasing returns to information. Instead the Pareto model allows the possibility of heavy tails and the possibility of infinite variance. The De Vany and Walls (1999, 2002) papers find evidence of a heavy upper tail in the outcome distribution as well as evidence of infinite variance. This is interpreted as evidence of the ‘nobody knows’ principle made famous by screenwriter William Goldman (1983) and set out academically by Caves (2000). This line of research on the distribution of film outcomes is further developed by De Vany and Walls (2004) and McKenzie (2008). These researchers use the unconditional Lévy-stable model set out in Section 8.3.3. These results generalize the earlier studies that fit the Pareto distribution, which only models the upper tail of film outcomes. The unconditional Lévy-stable model captures the entire distribution, including skewness, kurtosis, and heavy tails that in the limit follow a Pareto distribution. These studies further confirm the uncertainty surrounding film success in that the variance is infinite.

Many empirical studies, including Wallace et al. (1993), Ravid (1999), Albert (1998), Sedgwick and Pokorny (1999), Albert (1999), Collins et al. (2002), among others, quantify the correlates of high-grossing films in the context of standard statistical models; a thorough discussion of these models and their results relating to the use of marquee talent or stars, genres, ratings, awards, and reviews is provided in McKenzie (2012). These sort of models are useful in that they permit straightforward examination of the impact of various explanatory variables on a film’s box office revenue or rate of return. However, these models fail to capture the important statistical properties, such as heavy tails, skewness, and possibly infinite variance, that characterize the film industry. In an effort to bring covariates to the Lévy-stable analysis, Walls (2005b) applied McCulloch’s (1998) symmetric-stable regression model to film returns and the empirical results confirmed that the variance of film returns was infinite even when conditioning on a vector of covariates; however, a major shortcoming of that model is that it imposes symmetry. Walls (2005c) applies the skew-Student-t regression model to film returns to accommodate skewness and heavy tails; although the model is ad hoc, it does accommodate skewness and heavy tails.

Walls (2010a) estimates the generalized skew-stable model set out in Section 8.3.3. That model quantifies the effects of budgets and stars on film outcomes while allowing for skewness, heavy tails, and infinite variance. The empirical results indicate that the effect of large budgets and marquee stars is to reduce the risk that a film earns very low rates of return while increasing the most probable rate of return. This is consistent with Ravid's (2004) and Basuroy and Ravid's (2004) hypothesis that the objective of studio executives is not to maximize expected returns, but instead to make projects that are unlikely to fail. The findings are also consistent with quantile regressions reported by De Vany and Walls (2005) showing that big budgets and stars increased the lower quantiles of film revenues, but that they did not increase the upper quantiles. The general stable regression model is revealing because it identifies how the shape of the distribution is changed. The skew-stable regression analysis also confirms that the distribution of film returns – even when conditioning on a film’s attributes – has infinite variance so point predictions are not very useful. Predictions of the conditional expectation of film returns have infinite variance even though one can estimate a least-squares regression and obtain estimates that ignore this (e.g. Litman, 1983; Litman, 1990, Litman and Ahn, 1998). A pragmatic approach suggested by De Vany and Walls (1999) and applied by Collins et al. (2002) is to estimate the probability that returns exceed a given threshold using discrete choice analysis. While this approach can be used, it is preferable to make use of all the sample information by modeling the entire distribution of returns using the skew-stable model rather than modeling a discrete probability. The general stable model does permit the conditional distribution of film returns to be quantified and it can be used to make probability statements that may be useful in practice (e.g. pricing contingent claims on a film’s stream of future returns).

An alternative to the skew-stable model is non-parametric regression in which the precise functional form need not be specified explicitly. Walls (2009a,b) applies non-parametric regression with data-driven bandwidth selection to model movie revenues and movie profitability, respectively. It is reported that the non-parametric regression model fits the data far better than the logarithmic regression model employed by most applied researchers and that one can reject the hypothesis that the log-linear model is correctly specified. The non-parametric regression estimates of revenue response to production budget differ from those obtained by other researchers. For example, Walls (2005c) using the skew-Student-t model reports an elasticity of box office revenue with respect to budget of 0.75; Walls (2010a) using the symmetric stable regression model reports a budget elasticity of 0.81; De Vany and Walls (2005) use a log-linear regression and quantile regression to report budget elasticities of 0.55 and 0.56, respectively; Basuroy and Ravid (2004) use a log-linear model to estimate budget elasticities of 1.44; Ravid (1999) estimates a budget elasticity of 1.35 and reports further that it is the most significant determinant of domestic box office revenue; and the Litman and Ahn (1998) results imply an elasticity of 0.23.¹² The estimates of revenue response to opening screens obtained from the non-parametric kernel regression also differ from those obtained by other researchers. Walls (2005c) using the skew-Student-t model reports an elasticity of box office revenue with respect to opening screens of 0.273; Walls (2010a) using the symmetric stable regression model reports a budget elasticity of 0.313; De Vany and Walls (2005) use a log-linear regression and quantile regression to report budget elasticities of 0.205 and 0.234, respectively; Litman and Ahn (1998) estimates imply an elasticity of box office gross with respect to opening screens of 0.65.¹³

McKenzie (2010) and Walls (2010b) have in independent studies using different datasets examined the distribution of DVD sales. McKenzie (2010) found that the DVD revenue distribution had thicker tails than the theatrical revenue distribution; this implies that the top-ranked DVDs earned a greater share of revenues than their theatrical counterparts. His comparison of revenues shows not only a high degree of correlation between the two markets, but a relationship that is non-linear and increasing at higher theatrical revenue levels. This is interpreted as being consistent with a word-of-mouth momentum effect and more institutional flexibility in the DVD market. Walls (2010b) finds that the DVD market differs markedly from the theatrical market in the dynamics of demand and in the distribution of payoffs.¹⁴ Where the theatrical market has strong information feedback as evidenced by autocorrelated growth in weekly revenues, Walls (2010b) finds no such evidence of information feedback in the analysis of weekly DVD revenue. While the theatrical movie market has a clear winner-take-all character where the few high-grossing films are improbably large to have been drawn from a (log) normal probability model, he finds no evidence of heavier-than-normal tails in the distribution of cumulative revenues over DVD titles. The DVD market appears to be much less uncertain than the theatrical market for movies and it is populated by a more diverse set of competitors, including not only theatrically released movies, but also films released directly to DVD in addition to a wide variety of repackaged programming originally broadcast on television and cable TV. Competition at the box office leads to a winner-take-all distribution of prizes while competition for DVD sales has a much less skewed, though not entirely normal, distribution of payoffs. These findings are complementary to those of Jozefowicz et al. (2008) who found that word-of-mouth, critical reviews, and other structural factors that would create positive information feedback did not have any systematic impact on DVD rentals.

De Vany and Walls (1997) model film survival on the Variety’s top 50 chart. Each film in their sample was tracked from its birth to its death, where birth is defined as the beginning of a film’s theatrical run and death was defined as falling off the top 50 chart. Modeling film lifetimes is important because lengthening the theatrical run of a film is the most common way to adjust supply to evolving demand and because long lifetimes are very important in creating high-grossing films. They estimated the survival model set out in Section 8.3.6, with the vector of explanatory variables including the number of first run bookings, the week’s revenue, the number of weeks the film previously had been in the top 50, the film’s rank, and the number of screens in which the film played in its initial theatrical release. The empirical results indicated that the number of first-run screens, revenue, and the number of weeks in which the film previously was in the top 50 increase significantly the survival time of a movie. For a re-released film, the number of weeks it ran in its previous release is a measure of its entertainment value that carries over to the subsequent theatrical runs. This can be interpreted as further evidence of information transmission and memory. Walls (1998) models movie lifetimes using a sample of data from Hong Kong. McKenzie (2009b) also uses a survival analysis to analyze life length on Australian box office charts. He finds, contrary to several earlier studies, that production budget, opening screens, and the use of marquee star talent are not important determinants of life length. Simonoff and Ma (2003) model the hazard rate of Broadway shows to quantify the correlates of survival, and find that the type of show, whether a show is a revival or not, and opening-week attendance are strongly related to show survival. They also find mixed evidence regarding critic reviews: winning major Tony Awards is associated with longer runs, but being nominated and then losing is associated with shorter post-award runs. Maddison (2004) and Akdede and Ogus (2006) also examine the lifetimes of live performances and find results consistent with those for movie lifetimes.

8.4.2 Music

Connolly and Krueger (2006) in Volume 1 of this Handbook provide a thorough overview of the institutions and empirics on the economics of the popular music business. Throsby (2002) provides an empirical overview of the global music industry.

Hamlen (1991, 1994) tests superstar models using voice quality and finds some evidence that greater talent received greater rewards, but that differences in talent far exceed differences in reward. Chung and Cox (1994) model sales as a dynamic demand process where customers’ demand is related to prior-period demand and this provides some support for the type of model discussed in Section 8.2.2 where previous success is leveraged into future success; this supports Adler’s (1985) model to the extent that initial advantages may determine future success more than differences in talent. Giles (2006) and Spierdijk and Voorneveld (2009) both provide updates to the Chung and Cox (1994) paper, finding that the statistical model – the Yule distribution – does not provide a very good fit for the upper tail of the distribution, which implies that the model does not provide a good description of superstardom.

Giles (2007b) models durations of #1 Billboard chart hits and Giles (2007a) uses data from the Billboard Hot 100 chart and finds evidence of increasing returns to information in the US market for popular music. These results are consistent with the large number of results reported for the movie industry in Section 8.4.1. Davies (2002) examines the lifetime total success of bands measured by the total number of weeks they were in the weekly top 75 list of best-selling recordings and finds evidence that the individual success of musicians follows a stretched exponential distribution. Strobl and Tucker (2000) find that chart success is substantially skewed to the right for UK pop music using data from the 1980–1993 British rankings charts. Success is substantially skewed to the right, whether measured by total weeks per artist, average weeks per album, or the total number of albums per artist. The determinants of chart survival are also examined, and these indicate that the type of album and initial popularity are important in album survival.

Pitt (2010) examines the distribution of royalty income, explicitly modeling skewness and heavy tails using the skew-normal and skew-t distributions in a regression framework. Pitt finds strong evidence of superstar effects, in that average royalty payments are dominated by the values found in the upper tail and a small number of performers earned a very large share of the total royalty payments from blockbuster hits.

8.4.3 Books

General background on the economics of books, including institutions and empirics for the book publishing industry, are provided in Canoy et al. (2006) and also in van der Ploeg (2004). The empirical literature on the economics of bestsellers and blockbusters in relation to books is much less extensive than the literature on music or movies. There are studies, such as Clerides (2002), that focus on industrial organization issues related to the pricing of books, but few studies examine the distribution of outcomes across competing titles.

One paper that does examine the market for books in a way similar to the studies for movies and music is Gaffeo et al. (2008). This paper uses a form of the rank–revenue model set out in Section 8.3.1 to study the dynamics of the demand for books in Italy. Gaffeo et al. (2008) disaggregated the Italian book publishing industry into three broad submarkets: Italian novels, foreign novels, and non-fiction. Within each submarket the authors find evidence that the distribution of sales can be well-fitted by a power law distribution. However, in contrast to the studies of music and movie markets, a significant departure from the Pareto law was not found, so the authors conclude that the information transmission in the Italian market for books is characterized by constant returns. However, their results are consistent with the nobody-knows principle, the winner-take-all property, and the success-breeds-success principle.

Chevalier and Mayzlin (2006) investigate word-of-mouth impact on book sales using book reviews at two popular online booksellers: Amazon.com and Barnesandnoble.com. Book reviews are found to be positively associated with book sales, but the marginal impact of bad reviews is greater than the impact of good reviews. They also find evidence to indicate that customers actually read the text of online reviews rather than relying solely on the numerical review. This final result is probably the most interesting of their study in that it provides some evidence for models that are predicated on the sharing of quality information rather than models that are based solely on quantity signals; in this sense the Chevalier and Mayzlin (2006) paper is consistent with the De Vany and Lee (2001) paper, which in a simulation compares models of quantity signals to models of quality signals. Beck (2007) also estimates the impact of word-of-mouth on sales of novels. In an interesting empirical study, Sorensen (2007) finds that appearing on the bestseller list leads to a modest increase in sales for the average book and that the effect is bigger for first-time authors. Sorensen finds that the market expansion effect of bestseller lists appears to dominate any business stealing from non-best-selling titles.

8.5.1 Heavy Tails or the Long Tail?

Changes in the way goods are sold may affect the distribution of success. Anderson (2006) made popular the hypothesis that firms in the Internet age may be able to profitably sell small volumes of less popular titles to many customers instead of only selling large volumes of very popular titles. The total sales of a large number of non-hit titles is called the long tail. Does the evidence allow us to distinguish between these competing hypotheses? Are the results different across industries? Leskovec et al. (2007) analyze a person-to-person recommendation network, consisting of 4 million people who made 16 million recommendations on half a million products. They study the propagation of recommendations and the cascade sizes, and analyze how user behavior varies within user communities defined by a recommendation network. Product purchases are found to follow a ‘long tail’ where a significant share of purchases belongs to rarely sold items.

Tucker and Zhang (2011) find evidence of a steep tail effect where customers tend to herd to the most popular items, though there is a business creation effect that may stimulate demand for less popular products. This is not unlike the business creation effect of best-selling titles found by Sorensen (2007) in the book market. Elberse and Oberholzer-Gee (2007) examine data from music and movie Internet sales, and find no evidence to support the long tail theory. A few studies have examined the market for DVDs in comparison to the theatrical market for movies. McKenzie (2010) using Australian data found that the DVD revenue distribution had thicker tails than the theatrical revenue distribution so that top-ranked DVDs earned a greater share of revenues than their theatrical counterparts; thus success was even more highly concentrated on a small number of DVDs than it was in the theatrical market. Walls (2010b) using North American data finds that the DVD market differs markedly from the theatrical market in the opposite way: the theatrical movie market was found to have a clear winner-take-all character where the few high-grossing films are improbably large to have been drawn from a (log) normal probability model, but he finds no evidence of heavier-than-normal tails in the distribution of cumulative revenues over DVD titles. The North American sample of DVDs is populated by a more diverse set of competitors, including not only theatrically released movies, but also films released directly to DVD in addition to a wide variety of repackaged programming originally broadcast on television and cable TV.

8.5.2 Piracy and the Distribution of Success

Piracy is one of the most challenging problems faced by the movie, book, and music industries. One of the biggest problems that has hampered studies of piracy is the counterfactual methods by which effects are usually estimated. De Vany and Walls (2007) develop and estimate a statistical model of the effects of piracy on the box office performance of a widely released movie. The model discredits the argument that piracy increases sales, showing unambiguously that Internet piracy diminished the box office revenues of a widely released movie. The model overcomes a major weakness of counterfactual or ‘but for piracy’ methods widely used to estimate damages. These counterfactual methods violate the ‘nobody-knows’ principle because they forecast what the product would have earned in the absence of piracy. The model does not violate this basic principle of movie uncertainty. To summarize the economics, piracy adversely affects the dynamics of demand and supply in the theatrical market. The immediate effect on the dynamics is to increase the rate at which revenue and the number of screens showing the movie decline. The longer-term effect is to alter the information dynamics in such a way that demand in later weeks is also reduced. Thus, a pirated movie will play off more rapidly and lose revenue at an accelerated rate during its run. They test this economic hypothesis by empirically modeling the rate of decline in weekly revenues in relation to the number of pirated prints available for Internet download.

Are illegal downloads responsible for declining music/book/film sales? One theory is that instead of making a purchase a customer obtains the title online for free. However, sharing may lead to learning about performers and this extra information may increase legitimate sales. Oberholzer-Gee and Strumpf (2007) conclude that file sharing has a nearly insignificant impact on CD sales. Hong (2013) examines a panel of household expenditures at the time of the introduction of Napster. He concludes that file sharing accounts for 20% of the decline in music sales. Moreover, he concludes that this is primarily driven by households with children ranging between 6 and 17 years of age. Rob and Waldfogel (2006) asked a panel of undergraduate students at the University of Pennsylvania how much music they obtained from file sharing and from purchases. They estimate that for albums that sold more than 2 million copies from 1999 to 2003, there is no relationship between downloads and CD sales. However, for lower-sales albums, five album downloads correspond to one less album purchased. Rob and Waldfogel (2007) conduct a similar analysis for films. Liebowitz (2006, 2008) finds that file sharing accounts for at least all of the decline in US music sales. McKenzie (2009a) explores illegal music file-sharing activity and its effect on Australian sales of singles in the physical and digital retail markets. Using 15 weeks of Australian weekly chart rankings of physical and digital sales, combined with a proxy for download activity, his evidence suggests no significant impact of download activity on legitimate sales. Liebowitz (2011) standardizes the metric of sales decline across most of the studies of music sales mentioned above and concludes that in most cases the entire decline in music sales can be attributed to file sharing.

Mortimer and Sorensen (2007) study the relation between file sharing and concert attendance. They conclude that while sales of recorded music declined after the introduction of file sharing, concert revenues and the number of artists performing concerts increased dramatically. Overall, the patterns in the data suggest that while file sharing may have eroded profits from CD sales, it also increased the profitability of live performances. Hendricks and Sorensen (2009) investigate the role of information in the distribution of success across products with an empirical application to recorded music. Specifically, their paper models the role of information discovery in demand. Their results show that the release of a new album is associated with an increase in the sales of a performer’s previously released albums. Hendricks and Sorensen call these effects ‘backward spillovers’, and they support a structure of information flows where demanders discover the artist from the newly released album and subsequently make purchases of the performer’s previously released albums. They conclude that the distribution of sales across performers is more skewed than it would be if consumers were more fully informed. Examining the role of piracy on the dynamics of demand and the distribution of product success appears to be a fruitful area for future research.

8.5.3 Customer Reviews, Word of Mouth, and Social Media

How do customer reviews affect customer choice? How do rankings lists affect customer choice? Studies such as Chevalier and Mayzlin (2006) and Liu (2006) used online product reviews to predict product sales. Most studies such as Chevalier and Mayzlin (2006) show that online product reviews have a significant effect on product sales. However, other studies such as Liu (2006) show that it is the volume of product reviews that significantly influences sales while the valence does not show any significant effects. Duan et al. (2008) provide a preliminary analysis of online word-of-mouth impact on film sales. The way in which online reviews impact the entire distribution of sales is not clear. This appears to be a fertile area for research.

A dynamic component of online communication that has been rapidly increasing in volume are blogs and other forms of social media. Social media are a fast-growing form of online communication and they represent a widely used method of personal expression. Asur and Huberman (2010) use the chatter from Twitter.com to forecast box office revenues for movies. They show that a simple model built from the rate at which tweets are created about particular topics can outperform market-based predictors. They also show that sentiments extracted from Twitter can be further utilized to improve the forecasting power of social media. A number of interesting questions arise. Do the conceptual or theoretical models yield refutable implications that allow us to distinguish between competing models? Does the empirical evidence permit us to test one model against another? Do changes in electronic social interaction have implications for herding or learning in the process that leads to bestsellers and blockbusters?