An eminent artist will bring about a considerable change in the established modes of each of those arts, and introduce a new fashion of writing, music, or architecture.— Adam Smith (1759)

1 Introduction

Many scientists, entrepreneurs or artists are driven by the desire to be original, influential or successful. But little is known about how these variables interact. Adam Smith was the first economist to approach these topics, but since then few others have ventured into related inquiries, perhaps because originality is such an elusive concept. And yet, originality is fundamental to economic growth. Without the originality of an inventor, there would be no inventions, and hence no technological progress. And without the desire to become influential or successful, some inventors would never become original.

This paper sheds new light on how these concepts are related by formalizing a model of creative style, based on which we define and compute measures of originality and influence. We seek to understand in particular, whether originality, influence, and popular success are conflicting goals in the creative process—such that creatives trade them off against one another—or whether they go hand in hand, perhaps as by-products of the “quality” of creative work.

The concept of style is of general interest and has been studied in various contexts, including business, politics, economics, sociology, and finance (e.g., Bertrand & Schoar, 2003; Chan et al., 2018; Malmendier et al., 2021; Aggarwal & Woolley, 2019). Style has also a particular relevance in the arts, where it captures patterns of artistic production that are distinctive of an artist, place, or period. Our model of style will be applied to the arts, but its framework and insights are of general relevance. The model builds on a probability distribution over features (attributes) of creative works. Based on these distributions, we identify and measure the degree of originality and influence of a body of a creative’s work. The originality of a creator is the degree of uniqueness of her work relative to past work. This means that the reference point is the past: The more one deviates from the style of previous creators, the higher the originality. In contrast, influence looks to the future: The more predictive one’s work is of the style of future creators, the higher the influence.

We apply and test our framework in the context of Western classical music between the 15th and 20th centuries, which provides an ideal setting to analyze the interplay between originality, influence, and long-term success for several reasons: First, the data enables us to extract key features from a large number of musical themes—short but important excerpts of musical compositions. We are thus able to identify and measure the style of a composer or composition, going far beyond mere placement within a musical period. Second, the structured nature of classical music provides a setting where measures for such concepts as originality and influence can be constructed consistently and convincingly, and studied over the very long-term. Third, qualitative evidence from musicology and music history enables us to validate our methodology in a number of tests, and also to support our findings with various anecdotal accounts.

The core of our data has been collected from dictionaries of musical themes by Barlow and Morgenstern (1975, 1976) and includes 18,074 melodic themes from 6352 classical and operatic works by over 750 composers. The source provides sequences of musical notes, which are conveniently transposed to a common key, as well as a staff for each theme showing the original key and time signatures. We combine this data with measures of individual composers’ long-term success from modern consumption data based on Spotify streams as well as two measures of distinction: The Murray index (Murray, 2003) and the length of biographical entries in Grove Music Online, an encyclopedic dictionary of music and musicians. Finally, we collect information on the lives and work of composers from their biographies.

We find that more original composers tend to be more influential upon the work of their later peers and more successful with present-day audiences or experts. Influence and success increase monotonically with originality, though the relationship is more uncertain for very high levels of originality. The results hold also at the level of the individual theme, and are stable across different specifications as well as robust to the inclusion of various controls. The magnitude of the effects suggests an important role played by originality in determining both influence and success.

Finally, we develop a simple algorithm to randomly generate compositions. These tend to score highly on our measure of originality. We conclude that while it is not difficult to generate original works, it is a challenge to create original compositions that successfully pass through the selection implied by the composers’ creative process and of the (unknown) process by which compositions enter a compendium of famous works. In other words, maximizing originality is not a sufficient criterion to become influential or successful, but instead one’s output needs to conform to norms and standards of a given time.

Our work relates to scholarship on innovation (see Hall and Rosenberg, 2010), particularly to two studies that measure the influence of technological innovation (Kelly et al., 2021; Bussy and Geiecke, 2021), and the increasingly often studied concept of creativity. In a model of the creative process, Feinstein (2011) describes how creators explore and gather elements before finding ways to combine and reconfigure these elements into new original forms. Akcigit et al. (2018) model and analyze sources of the creator’s knowledge, which comes from interaction with other people or from external sources related to own explorations. Borowiecki (2022) explores whether teachers in creative fields leave an imprint on their students that shapes the style of their future work, using parts of the dataset presented in the underlying paper. In recent years, economists have also begun to study creativity empirically. For example, Graddy and Lieberman (2018) explore how artists’ creativity is affected during bereavement, while Borowiecki (2017) shows the overall influence of psychological well-being on creative production.

Related to creativity is the concept of style, which has traditionally been restricted to the arts (Munro, 1946). However, in recent years, style has become increasingly relevant in several other contexts. Business scholarship includes studies of leadership style (Rotemberg and Saloner, 1993; Bertrand and Schoar, 2003), style in product design (Chan et al., 2018), and cognitive style in teams (Aggarwal and Woolley, 2019). The communication style may denote voting outcomes of Federal Reserve committee members (Malmendier et al., 2021) and constitutes a particular feature of populism (De Vreese et al., 2018). Style characterizes the behavior of children (Caspi et al., 2003), parents (Kim et al., 2015), and financial investors (e.g., Keim & Madhavan, 1997).

This work also relates to research on innovation in history. The supportive intellectual environment in Europe since the 1500s as well as its political fragmentation explain the stark rise in innovations in Europe, which triggered the Industrial Revolution and economic progress (Mokyr, 2016). Freedom to pursue one’s own paths of inquiry results in increased originality of people (Feinstein, 2006; Aghion et al., 2008; Simonton, 2004). Relevant is also the important literature on the upper-tail human capital and economic growth (Mokyr, 2009; Meisenzahl and Mokyr, 2011; Squicciarini and Voigtländer, 2015)).Footnote 1 The modern economic growth in Britain and abroad was possible due to an increased desire of innovators to influence others as they spread their improving mentality (Howes, 2017). Creative activity has also been mapped and measured over time and place. For example, Murray (2003) selects leading innovators in the arts and sciences from 800 to 1950, whereas Gergaud et al. (2017) document the geographic spread of innovative individuals and cities. In cities where many similar individuals cluster, visual artists (Hellmanzik, 2010) and music composers (Borowiecki, 2013, 2015) are more productive.Footnote 2

This paper provides two main contributions. First, it introduces a new model that enables not only a formalization of style but also the measurement of the degree of originality as well as influence. The framework is versatile and can be applied in many contexts, including to individuals, products, or time periods, which makes it relevant for a large number of unanswered important research questions. Second, it provides an analysis based on this framework that shows originality to be positively associated with influence and several success measures among the composers in our sample. We conclude that originality, rather than being traded-off against influence and success, is a key driver of both.

The paper proceeds as follows. Section 2 introduces a model of style and proposes how to measure originality and influence. Section 3 presents the data sources and descriptives. Section 4 shows the results. Conclusions are in Sect. 5.

2 A model of style, originality, and influence

The aim of this section is to introduce credible measures of originality and influence. In order to do so, we first outline a model of style. Style is the way one typically does something. For example, style can be a particular manner or technique by which something is created, written, or performed. It is a distinctive characteristic of a person, group of people, place, or period.

Style plays a particular role in the cultural and creative sectors, because it enables the grouping of creative output or creatives into categories. The observation of style can thus be a useful tool in the study of the arts and music, advertising, architecture, design, film, publishing, video games, and many other creative domains. While any creative output is to some extent unique (otherwise it would not be creative), the manner in which it has been created can be categorized within a certain style.

In this section, we first outline our model of style, before turning to how we conceive of originality and influence on the basis of our model of style.

2.1 Modeling style

We propose to capture the style of creative output as a function of its features or key attributes. For example, in the visual arts such features may include the painting technique, materials used, or motifs painted. In the music it may be specific sequences of music notes, the key in which a piece is written, or more aggregated measures such as rhythm or mood.

In this framework, a style is linked to the frequency of these features. To make this empirically workable, the concept of style must be linked to creative output for which the frequency of features can be measured. We conceptualize style as a probability distribution over features and assume that the frequency of these features is the result of drawing from this probability distribution.

Formally, consider creative works by different people. Let F be the set of features that occur in at least one of the works under consideration, with \(f=|F|\) denoting the number of such features. The style of an individual creator i, \(s_i\), is a probability mass distribution over all features in F. In other words, style, \(s_i\), is a vector of length, f, the components of which sum to one.

Let \(W_i\) be the set of works by individual i. Then each work, \(w \in W_i\), is characterized by its associated vector of feature frequencies, \(x_w\). We assume that \(x_w\) is the result of the composer drawing from \(s_i\). The frequency distribution of features over all of individual i’s works, \(x_i= \sum _{w \in W_i} x_w\), is thus a noisy proxy for style, \(s_i\).

2.2 Modeling originality

We conceptualize originality in terms of deviation from past creative output. The style of an individual i, \(s_i\), is more original the more it deviates from past style, whereas a perfectly unoriginal individual would have an identical style to past style and the degree of originality would be zero.

In terms of the framework outlined at the start of Sect. 2, we first define a creator’s innovation in terms of her own style and past style.

Definition (Innovation): Let \(s_i\) be individual i’s style and \(s_p\) be the past style. Then i’s innovation is defined as:

$$\begin{aligned} v_i = \tilde{s}_i - \tilde{s}_p \tilde{s}_p' \tilde{s}_i \end{aligned}$$

where \(\tilde{s}_i=\frac{s_i}{||s_i||}\) and \(\tilde{s}_p=\frac{s_p}{||s_p||}\) are the normalized versions of \(s_i\) and \(s_p\); that is, they are rescaled so as to have length one.

Like an individual’s style, her innovation is an f-dimensional vector. But unlike style, it is not a probability distribution. Instead, it is the difference between an individual’s (normalized) style, \(\tilde{s}_i\), and that style’s projection onto (normalized) past style, \(\tilde{s}_p\). The innovation is orthogonal to past style by construction. It gives the direction of the individual’s deviation from the past.

In order to measure the degree of originality, which we refer to simply as originality, the innovation needs to be translated into a scalar measure. We do this by taking the length of the innovation vector, \(v_i\).

Definition (Originality): Individual i’s originality \(\omega _i\) is defined as the length of her innovation:

$$\begin{aligned} \omega _i =\Vert v_i \Vert \end{aligned}$$

Originality will therefore be greater, the more \(s_i\) deviates from \(s_p\). Since all entries of each style vector are non-negative, originality will equal zero if and only if \(s_i=s_p\). This makes originality a measure of dissimilarity between \(s_i\) and \(s_p\).Footnote 3 At most, originality may equal one if \(s_i \perp s_p\), a condition that may occur only if the sets of features with positive entries in \(s_i\) and \(s_p\) are disjoint.

Figure 1 illustrates the definitions of both innovation and originality. The normalized styles, \(\tilde{s_i}\) and \(\tilde{s_p}\), are vectors of length one. The innovation, \(v_i\), is the difference between \(\tilde{s_i}\) and its projection onto \(\tilde{s_p}\). Its direction is given by the feature space in which individual i’s style deviates from the past. That is, it shows which features are either more or less common in individual i’s style than they are in past style and to what extent they are more or less common. If these differences are large, then \(v_i\) will be longer, which is why its length defines our measure of originality, \(\omega _i\).

The figure also illustrates another view of our originality concept, namely in terms of a decomposition of style, \(s_i\), into two components: One in the direction of past style, \(s_p\), and another which orthogonally departs from the past, the innovation. The relative lengths of the two components indicate to what extent individual i follows the past or innovates.

2.3 Modeling influence

While originality measures the extent to which an individual’s style departs from the past, influence relates an individual’s style to the future. The individual is influential if she “bends” future work in the direction of her style.

The definition of influence relies on three concepts introduced above: Past style, \(s_p\), future style, \(s_f\), and individual i’s innovation, \(v_i\). Using these, we measure the extent to which future style follows individual i’s innovation.

To this end, we first project \(s_f\) onto the span of past style and innovation (which is equal to the span of past style and the individual’s style). The span of \(s_p\) and \(v_i\) is described by the matrix

$$\begin{aligned} A = \left( \tilde{s_p}, \frac{v_i}{||v_i||}\right) \text { ,} \end{aligned}$$

in which the two vectors are already normalized. The projection of \(\tilde{s_f}\), the normalized vector describing future style onto the plane defined by A is then given by

$$\begin{aligned} p_f = AA'\tilde{s_f} \text { .} \end{aligned}$$

This projection can be understood as the “shadow” of \(\tilde{s_f}\) on the plane which contains both \(s_p\) and \(v_i\). Having thus brought all three objects into the same plane, what remains to be done is to evaluate the deviation of \(p_f\) from \(\tilde{s_f}\) in terms of innovation \(v_i\).

Figure 2 illustrates this. The vectors, \(\tilde{s_p}\) and \(v_i\), are drawn as in Fig. 1. The projection, \(p_f\), of \(\tilde{s_f}\) is dashed. It is in general shorter than one, since \(s_f\) will not perfectly align with the plane defined by A except by coincidence. In general therefore, there is a component of \(\tilde{s_f}\) that points away from the plane, following neither the past nor individual i’s innovation.

In the left panel of Fig. 2, \(p_f\) deviates from \(\tilde{s_p}\) in the same direction as \(v_i\). In a sense, \(v_i\) points to the future. In this case, the creator’s influence is positive. In the right panel, by contrast, \(p_f\) deviates from \(\tilde{s_p}\) in the opposite direction. Here \(v_i\) points away from the future. In this case, the creator’s influence is negative.

To formalize this measure, we first project again, this time \(p_f\) on \(\tilde{s_p}\), and obtain individual i’s contribution to future style:

$$\begin{aligned} u_i = p_f - \tilde{s_f} \tilde{s_f}' p_f \end{aligned}$$

This contribution to future style is analogous to the innovation defined in Sect. 2.2. It too can be viewed as part of an orthogonal decomposition, here of \(\tilde{s_f}\), where the contribution to future style represents the component that runs parallel to innovation, \(v_i\).

But, as can be seen in the right panel of Fig. 2, \(u_i\) does not need to run in the same direction as \(v_i\) and the distinction is crucial. For this reason, we define influence as a directed length of \(u_i\), where the direction is given by \(v_i\).

Definition (Influence): Let \(u_i\) be individual i’s contribution to future style and \(v_i\) her innovation. Then individual i’s influence is defined as the directed length of \(u_i\):

$$\begin{aligned} \lambda _i = ||u_i|| \frac{{u_i}\prime v_i}{|{u_i}\prime v_i|} \end{aligned}$$

As with \(\omega _i\), individual i’s originality, influence cannot exceed one. This extreme case occurs when future style is equal in direction to the creator’s innovation, \(v_i\); that is, if the novel part of individual i’s style “becomes” the future.Footnote 4 At the other extreme, the influence of an individual i can theoretically approach minus one, though it cannot be exactly equal to minus one. Extreme cases occur when innovation is very small and future style is nearly orthogonal to past style.Footnote 5

Another instructive aspect of the behavior of influence as defined above concerns the length of \(p_f\). As can be seen intuitively in Fig. 2, if \(p_f\) were to shorten, while retaining its direction, this would move \(\lambda _i\) closer to zero. The length of \(p_f\) varies with the extent to which future style, \(s_f\), is aligned with the plane defined by A: If the future is very different both from the past and from individual i’s style, \(p_f\) will be short and influence will tend to be small in absolute value. In the extreme, if the set of features used in the future is disjoint from that used in the past and by individual i, then influence is necessarily zero.

2.4 Working with frequencies

So far, we have defined a creator’s innovation, originality, and influence in terms of style. Yet in reality style, defined as a probability distribution over features, is not known. Instead, any style will have to be substituted by its frequency equivalent.

The frequency equivalent of a style is the empirical style, a frequency distribution over features drawn from an observed corpus of creative work. For instance, while individual i’s own style, \(s_i\), cannot be observed, her empirical style, \(x_i\), can. It is the frequency distribution in individual i’s collected works over all features in F. Like \(s_i\), \(x_i\) is thus an f-dimensional vector whose entries sum to one. Empirical past style, \(x_p\), and empirical future style, \(x_f\), are defined analogously.

To avoid cumbersome references to empirical style, we will refer to it as “style” in the empirical section of the document, but denote it \(x_i\), \(x_p\), and \(x_f\) for clarity. Our baseline measures of composer i’s originality, \(\omega _i\), and composer i’s influence, \(\lambda _i\), are calculated on the basis of empirical style, using all features used by composer i and the full corpora of past and future themes.

In our empirical application, we observe many composers for whom only very few themes are available. The resulting sparsity of their empirical style \(x_i\) induces a correlation between the number of themes available for a composer and their originality estimate. To address this issue, we correct our baseline originality measure for sparsity. Appendix C describes this correction.

In addition to the baseline measures, we compute alternatives in which different themes are taken into consideration for \(x_i\), \(x_p\), and \(x_f\). Among these alternatives are measures computed at the level of the individual theme, such that each theme is considered to have its own style. In addition, we also compute an influence measure at the composer level, but restricting the corpus considered for future style.

2.4.1 Theme-level originality and influence

The theme-level approach yields estimates of originality, \(\omega _{it}\), and influence, \(\lambda _{it}\), for each theme, t. To aggregate these values at the composer level, we compute simple averages over all themes of a given composer. While this may seem redundant, the resulting composer-level average originality, \(\bar{\omega _{i}}\), has a key advantage. It does not suffer from the sparsity problem described above. This is because the themes tend to be of similar length. No correction is therefore applied for theme-level originality.

A similar theme-level measure is computed for those themes for which a composition year (or interval) is available. For each of these 3,038 themes, the corpus of past themes is assembled from all themes which were either composed before the start of a given theme’s time interval, or which were composed by a composer who died before the start of the interval. The future corpus is assembled analogously, including those themes composed after the end of a given theme’s time interval or by composers who were less than ten years of age at the end of the interval. This measure is only available for a reduced sample but has the advantage that past and future corpora contain works closer to the date of composition of a given theme.

2.4.2 Near-term influence

To account for the possibility that a composer’s influence may change over time, we create a near-term influence measure to be compared with overall influence. A composer might be highly influential on the next generation of composers but then be mostly forgotten, or ignored by the next generation and rediscovered much later. The near-term influence of composer i measures her influence on the first generation of composers who were born less than 30 years after composer i’s death and were not yet ten years of age at composer i’s death.

2.5 Modeling relationships

Originality, influence, and success are related in various ways. Before investigating these relationships empirically, it is useful to discuss them conceptually.

Figure 3 represents our thinking diagrammatically. By its construction, originality precedes the others in time. It is constructed on the basis of a creator’s own style and of past style, meaning that all information on which originality is based was available at the end of the creator’s productive life. The other two measures are only realized later, since influence depends on future style and success is measured in the present day. For this reason, we are mainly interested in the consequences of being original and treat it as exogenous to influence and long-term success.

More original creators may be more influential by being more inspirational to their future peers. At the same time, they may not be to the taste of their audience or critics, making them less well known to both future creators and future audiences. The relationship between originality and influence or success is thus ex-ante unclear and remains to be explored empirically in Sect. 4.

The relationship between influence and measures of success, however, is hard to interpret. By baseline measures the two are positively correlated, though not strongly (\(20\%\)). Various causal explanations are plausible. Both could be affected by common factors, such as originality as just discussed; success may bring a creator to the attention of others, driving influence; or a creator’s influence on others may become public knowledge, increasing her fame and thus her success. Of course, negative causal effects are also possible. Creators may avoid drawing on successful past creators, so as not to be seen as pedestrian.

3 Data and descriptives

3.1 Features of musical themes

We collect data on features of a large number of musical themes from two volumes of musical dictionaries by Barlow and Morgenstern (1975, 1976). The authors declare that they cover “all the themes the average and even the more erudite listener might want to look up” and include 18,074 musical themes from 6352 classical and operatic works written by 769 composers.

These dictionaries have been compiled for readers to check details on a musical theme (e.g., the title, who composed it, etc.) based solely on the melodic tune. Therefore, for each theme the dictionaries provide a sequence of musical notes intended to enable the reader to uniquely identify each theme. We use this information to obtain n-grams—sub-sequences of n notes (duplets, triplets, quadruplets). Conveniently, the sequences are transcribed into a notation index, which is transposed to the key of C (either major or minor) and reports each note as a letter with a trailing symbol to indicate a sharp (half-step raise) or a flat (a half-step lowering).Footnote 6

For example, the theme of Symphony No. 40 in G Minor by Wolfgang Amadeus Mozart is included in Barlow and Morgenstern (1975) as notes on a staff (Fig. 4). In addition, the first six notes of the theme are included in the notation index after being transposed to a common key: C minor as the original theme is in a minor key. The first six notes of the original theme in G minor Eb – D – D – Eb – D – D become, after being transposed, Ab – G – G – Ab – G – G.

Along with the sequence of notes, we collect time signature and key signature for each theme. A time signature, which can be indicative of rhythm, specifies how many beats are contained in each measure (segment of time) and which note value is equivalent to one beat. In Symphony No. 40, the C symbol at the beginning of the music staff means “common time” and is another way of writing the 4/4 time. A key signature indicates notes that are to be played higher or lower than the corresponding natural notes. In our example, the very title of the work tells us the key, and other hints come from the key signature, two flats reported at the beginning of the staff, and the harmonic structure of the opening.

Yet the key could not be successfully recorded for all themes. While the number of flat and sharp notes reported at the beginning of the staff can normally be observed for all themes, two possible keys—one major and one minor—are associated with the same set of sharp or flat symbols. For example, the key signature for G minor has two flats and its relative major is B-flat major.

To identify the key we perform a series of tests that look at the presence of certain notes in the musical theme (see Appendix A for details on how we identify the key signature). Themes for which the key identification process led to ambiguous results were dropped from the main dataset. Our baseline focus on 2-grams (sets of two consecutive notes), time signature, and key signature, enables us to cover 10,890 musical themes and 684 composers. The total number of unique features extracted in this way is 193.Footnote 7 Including the key and time signature then, the final list of musical features for Mozart’s Symphony No. 40 is [AbG; GG; GAb; AbG; GG; 4/4; G minor].

When applying the model presented in Sect. 2 to this data, we can observe that the two vectors of style of the past, \(x_p\), and future, \(x_f\), are generally similar (i.e., the angle between them, at around \(30^{\circ }\), is small). This is due to considerable continuity in this art form over the centuries. The two vectors also typically have few zeros, since any feature that occurs at all is likely to be seen at least once in both the past and the future.

Information on the years of compositions is obtained from Petrucci Music Library (2020). Unfortunately, the years of composition are only available for a small subset of 3038 themes, typically as an interval of several years, during which the associated work was composed.Footnote 8 For this reason, we do not rely on this data in our main analysis, but we provide results for a theme-level approach based on composition years. In the main analysis we rely on the long time span of the data, defining past and future corpora using the birth and death years of our composers. If a theme was composed by someone who died before a given composer turned ten, we consider it to be in the composer’s past.Footnote 9 On the other hand, if a theme was composed by someone who was not yet ten when a given composer died, we consider it to be in the composer’s future.

3.2 Success

To capture a composer’s success (or eminence as one of our sources insists on calling it) we collect data from three different sources. We are thus able to cover popular success as well as success designated by experts.

First, we rely on modern consumption data from the music streaming service Spotify. We collect the number of followers of each composer and a popularity index created directly by Spotify. The latter ranges from zero to 100. Both Spotify measures are retrieved for a period of 12 consecutive months and then an average of the scores is retained. This should limit concerns about seasonality of music or short-term shocks affecting the measures. As the two measures are highly correlated, we will, in the following sections, employ the number of followers as our baseline measure for success, but the results would remain qualitatively the same if any of the other success measure was used instead.

Second, we obtain the Murray quality index from “Human Accomplishment” (Murray 2003), which scores individuals from many disciplines who lived between 800 BCE and 1950. For Western music, the author combines 16 different international sources including encyclopedias, dictionaries, and surveys and aggregates them providing a list of the 522 most significant composers along with a weighted index that ranks these composers on a scale from one to 100. We assign a Murray index of zero to composers not listed by Murray but covered by Barlow and Morgenstern (1975, 1976). In this historiometrical effort the author argues that it is excellence in accomplishments that underlies the measures of eminence compiled in the overview, rather than mere fame.

Finally, we count the number of words contained in the main text of each composer biography published in Grove Music Online (2018)—a leading encyclopedic resource on music and musicians. The biographies are written accounts of the lives of composers including life events and their creative outputs.

3.3 Biographical information

We obtain detailed biographical information for all composers for whom we have information on musical themes and on measures of success. We collect individual-level data from the Grove Dictionary of Music and Musicians (now Grove Music Online 2018), including information on the years of birth and death as well as place of birth and nationality.Footnote 10

For composers for whom information on the birth or death years or nationality are missing we used data from Pfitzinger (2017) and from Barlow and Morgenstern (1975, 1976), who occasionally report years of birth and death of the composers, and, if needed, other reference and online resources. For very few remaining composers—only seven—for whom either birth or death years were still not available, we estimated these dates using the information available (e.g. birth, death, or “floruit”, the period during which the artist was reported active) as well as information from the lives of their contemporaries.

We have also obtained for each composer data from the biography in Grove Music Online (2018) on her teachers and students, including information on which music conservatories she attended and when, as well as traveling data, which includes the list and number of cities visited. These data are used in additional analyses and robustness tests in the Appendix E.1.

3.4 Descriptive statistics

Table 1 reports the century and country of birth for the 684 composers included in our main dataset. Though some of these composers were born as early as the 15th century, \(62\%\) were born in the 19th century. Most of our sample covers European composers, while a considerable share of composers in the 19th century is North American.Footnote 11

Our baseline measures for influence, originality, and success are summarized in Table 2. Panel (A) summarizes originality measures (see Sect. 2.4 for descriptions). The first measure is computed for each composer, the second for each theme. Both measures are standardized and range from a minimum of zero to a maximum of 100. While the two variables cannot be directly compared, once theme-level originality is collapsed to a composer-level mean, the correlation between the two measures is positive and higher than \(50\%\).

Influence measures are shown in Panel (B) of Table 2 (see Sect. 2.4 for descriptions). The measures are standardized while preserving the meaning of the zeros. The process results in influence variables having different negative boundaries and a maximum of 100. Influence measures are available for only 432 composers. This is because influence can only be evaluated when a corpus of future compositions is available, which is not true of composers who live at the end of the sample period. Influence highly correlates with both near-term influence and theme-level influence (\(>85\%\)).

Success is shown in Panel (C) of Table 2. All success measures are taken in their logarithmic form and, despite measuring different aspects of the popularity of a composer, are all highly correlated (around \(70\%\)). Before taking the logarithm, we set the index to zero for all composers in our dataset who are not explicitly mentioned by Murray and add one to all values.

Histograms of originality, corrected originality, and theme-level originality as well as the three influence measures are available in Appendix C (Figs. 14 and 16).

Table 3 reports summary statistics for originality, influence, and success by historical-musical period into which Western music is traditionally divided. Average values of influence decrease over time, while average originality remains fairly stable, with the Medieval, Baroque, and Romantic periods being characterized by the highest degree of originality. Success is on average much higher for composers of the Romantic period, which is also the one for which we have the highest number of composers in our dataset.

The changes in originality and influence over time are shown in Fig. 5. Instead of simply aggregating the originality and influence measures by year we create a “rolling window” indicator, which measures the originality and influence of the work of all composers who were at least ten years of age in a specific year—the contemporaneous composers. Their combined corpus is compared against past work by composers who died before the youngest of the contemporaneous composers turned ten and against the combined corpus of future work by composers who were at most ten years old when the oldest of the contemporaneous composers died. The beginning and end of the sample period have to be interpreted with caution: The number of past themes is limited for composers near the beginning of our sample. Similarly, the number of future themes is lower for composers born near the end of the sample. This could limit the accuracy of our measures at the beginning and end of our observation window.

Focusing on the more central parts of the sample period, we observe two peaks in originality—around 1500 and at the turn of the 1700s, the Baroque period. The first peak is followed by almost two centuries of relatively low originality. In this period music by others was often appropriated for re-use, and such borrowing would not raise any concerns of plagiarism. The second peak—around the mid-18th century—comes at a time when originality had become increasingly praised and eventually became regarded as superior to imitation (Grove Music Online 2018).Footnote 12 As a result, the quality of a composer was assessed by the inventiveness of new music as opposed to a skillful manipulation of existing material.

Figure 5 also shows that the peaks in originality are not closely followed by peaks in the influence measure. Moreover, both measures clearly decrease during the last roughly two centuries, which is possibly a reflection of the “low-hanging fruit” phenomenon. A more formal explanation is provided by the humanities’ theory of anxiety of influence. The body of musical works increases with time, which instills in young composers a form of anxiety as they struggle against past musical predecessors to create something original and to achieve success. The correlation between originality and influence is equal to 0.31.

4 Results

In this section we first present the results of our baseline approach, which involves estimating originality and influence at the composer-level and considers all work before a given composer turned ten years of age to be the past and all work after her death to be the future. We then present the results based on the variants of the approach discussed in Sect. 2.4 and investigate the relationship between a composer’s age and her originality and influence. Finally, we assess the relationship between near-term influence and overall influence.

4.1 Originality, influence, and success

This section explores how the baseline measure of originality is related to, on the one hand, the success measures and, on the other, the baseline measure of influence.Footnote 13 We first explore these relationships graphically. The graphs in Fig. 6 plot baseline originality against our different measures of success: The number of Spotify followers, the Murray index, and word count measures.Footnote 14 In all cases, the graphs suggest a positive relationship between originality and each of the success measures.

Panel (A) of Fig. 7 shows non-parametric regressions of the success measure—the (log) number of Spotify followers—on baseline originality. Panel (B) shows an analogous regression of baseline influence measure on baseline originality. The graphs suggest that both success and influence are approximately linearly related to originality. Interestingly, rising originality increases influence on average, but the effect is more volatile.

Next, we explore these relationships using regression analysis. In Table 4, the success measures are regressed on baseline originality. Estimates report results for a Seemingly Unrelated Regression (SUR), where the error terms are allowed to correlate across the three columns of each set. The results show that originality is positively associated with each of the three success measures.Footnote 15 These relationships are barely affected by the inclusion of controls for the number of past and future themes or nationality fixed effects (columns 4–6).Footnote 16 To quantify the effect, looking at column (4) in Table 4, we can read that an increase in originality by 1% is associated with an increase in the number of Spotify followers by close to 10%.Footnote 17 To express this differently, for a composer with 145 thousand followers (the log-mean of the distribution), this would imply an additional about 14,000 followers.Footnote 18

Table 5 reports results for an OLS regression in which baseline influence is regressed on baseline originality. Different specifications progressively introduce fixed effects for the half-century of the composer’s birth, fixed effects for the nationality of the composer, and controls for the number of each composer’s themes as well as the number of themes in each composer’s past and future. Reading the results from column (4), Table 5, we note that an increase in originality by 1% is associated with an increase in influence of about 0.3\(\%\). An attempt to visualize the relationship between originality, success, and influence, is reported in a three-dimensional heat map in Appendix F.

In summary, originality is shown to be positively associated with success or influence, and remains stable across various specifications. Even when we include half-century and nationality fixed effects, as well as the number of past, future, and composer themes, the coefficients remain large and positive. Next, we turn to theme-level data.

4.2 Theme-level data

Exploiting theme-level data has two advantages. First, it allows us to reassess the results previously obtained, that originality is positively associated with success and influence. Second, by including composer fixed effects, it becomes possible to measure the effect of originality on influence over the lifetime of a single creator.

Table 6 reports OLS regressions as before, but where influence and originality are calculated at the theme level and every observation is a unique theme. As the table shows, the effects of originality on both influence and success remain positive and highly significant. For the within-composer effects, shown in columns (2) and (3), this means that the more original themes of a given composer tend to be more influential on future music than less original themes of the same composer.

Comparable results emerge in Table 7, which provides a more restricted estimation of the previously shown model. In particular, we consider now only themes for which a date of composition is available.Footnote 19 As past and future are defined using the composition period of each theme, instead of the life of the composer, compositions of contemporaries are included in the calculation of influence and originality. This may explain results in columns (4) and (5), where originality loses its significant relationship with composer’s success.

It should be noted that themes for which a year of composition was retrieved where more likely to be composed by relatively successful composers. Figure 8 clearly shows that the two success distributions are only partially overlapping, with the distribution of success for composers with known composition years extending further to the right.

4.3 Composer’s age, originality, and influence

We exploit the available sample of themes for which the composition year was recorded and calculate composer’s age at the time of composition.Footnote 20 Figure 9 shows the relationship between composer’s age when the theme was composed and theme-level originality (Panel A) and theme-level influence (Panel B). In both panels, the fitted line is estimated using non-parametric regressions of the composer’s age on originality (Panel A) and influence (Panel B). The scatter plots show weak but suggestive patterns whereby originality appears to be smaller for themes composed very early or very late in a composer’s life and influence appears to decrease with the composer’s age.

To test for these patterns, we perform quadratic regressions for each relationship. The results are shown in Table 8. While the concave relationship between originality and composer’s age is not significant with composer fixed effects and controls for the number of past and future themes, the negative relationship between influence and composer’s age remains consistent throughout both specifications.

4.4 Near-term influence

The relationship between near-term influence (described in Sect. 2.4) and baseline influence (or \(\lambda _i\) in Eq. 6) deserves some attention. While near-term influence reflects the most immediate impact of a given composer (i.e., on the next generation of composers), overall influence should capture the influence over future generations.

For a fraction of composers (\(30\%\)), the two measures coincide as they were born at the end of the sample. For those for whom the two measures differ, we can observe several cases in which the overall influence is much larger than the near-term one. These composers are overall quite influential, but they seem to be much more appreciated in the long-term than by the next generation of composers. Figure 10 illustrates this relationship. Composers such as Luca Marenzio, Claude Le Jeune, and John Dunstable have an overall influence measure above 50, much higher than their near-term influence. They all appear to be composers whose style was transitional in between two movements. While recounting the style change in Italy from the late Renaissance to the early Baroque, Giustiniani (1961) describes Marenzio’s work as follows: “In a short space of time the style of music changed and the compositions of Luca Marenzio appeared with delightful new inventions, either that of singing with several voices or with one voice alone accompanied by some instrument, the excellence of which consisted in a melody new and grateful to the ear, with some easy fugues without extraordinary artifices.” Note that Marenzio scores high in our measure of originality, \(\omega _i\), placing himself among the 25% most original composers.

In the same period, Claude Le Jeune, criticized by his contemporaries for his use of the counterpoint, profoundly influenced some composers of the 20th century. For example, in 1948 Olivier Messiaen composed “Cinq rechants” an hommage to Le Jeune’s “Printemps” (Dobbins and His, 2019). Dunstable, too, was a transitional composer who “influenced the transition between late medieval and early Renaissance music” (Encyclopaedia Britannica, 2020).

4.5 Simulating themes

As discussed in Sect. 1, we view the compositions in our data as having originated through, first, the creative process of each composer and, second, a selection process by which they are included in the canon of sufficiently well known works as proxied by Barlow and Morgenstern (1975, 1976). Our results above show that originality is positively associated with influence and several success measures among the composers in our sample. Therefore, originality appears to determine influence and success, rather than being traded-off against either. Does this mean then that a composer should at any rate maximize originality or are there any barriers to achieving high originality? To shed some light into the extent of such barriers, we simulate musical themes similar to those we find in the data in order to evaluate their originality against the corpus of themes in the data.

Proceeding conservatively, we draw notes one by one to assemble random themes, excluding relatively rare notes and evaluate each simulated theme’s originality.Footnote 21 Fig. 11 shows the resulting distribution of originality scores over simulated themes and real themes of equal length.Footnote 22 Despite being composed of only the most common notes, the simulated themes tend to score higher in originality than those in the data.

This suggests that it is not the case that highly original themes are rare per se. Rather, it appears that the creative and selection processes impose constraints that disfavour themes which are as if randomly generated. These constraints may include a need for a theme to conform to norms that are not built into our model or to be pleasing to hear in order to achieve some initial success with audiences, peers, and critics. The creative artist should thus not maximize originality at any cost, but rather optimize it in consideration of—unobservable to us—norms and standards of a given time.

5 Conclusion

Identifying the style of a creative person or her work has long been of interest across several domains and is widely regarded as difficult. It is even more challenging to identify the extent of a person’s influence upon others, or her originality. Measurement of these concepts is practically non-existent due to the elusive nature of style.

In this paper we overcome some of the difficulties by developing a novel approach to define and measure style. This enables us to rigorously identify and estimate the originality and influence of a creator or her creative output. Based on this framework, we study the interdependencies between originality, influence, and success in the context of music composition.

Our results show that more original composers tend to be more influential upon the work of their later peers and are more successful with present-day audiences or experts. Originality is also predictive of influence and, to a lesser extent, success at the theme-level. We conclude that originality, rather than being traded-off against influence and success, is a key driver of both.

The empirical context and background of this study are distinctive, and so is the focus on a small, albeit important group of composers. However, the model proposed and mechanisms examined here are likely applicable to most settings where creative output is produced, especially since creative processes are likely comparable across domains and disciplines (van Broekhoven et al., 2020). This means that, for example, musicians and scientists may create two different kinds of work with different intentions and outcomes, but the process they use to get there appears to be similar (see Azoulay et al. 2017). Moreover, the method presented and tested in this paper, by enabling the measurement of the abstract concepts of style, originality and influence, opens new areas for future research, including on the role of design, leadership, and investment styles, or styles of political speeches and corporate communication. Finally, it is not only the measurement of originality or influence that is important, but also the very specific question of how these concepts are related to success. In addition to being of great importance to a composer, her success also brings benefits to her audience and to cultural production in her artistic field. On aggregate, success in the creative industries is relevant for economic growth.

Many creators, including artists, scientists, and entrepreneurs, are driven by the desire to become successful and influential. How to achieve that success or influence is fairly unknown. The ex ante uncertainty of the outcome (or the nobody knows theorem) is a notorious feature of cultural production and applies also more broadly to many areas of the creative industries (e.g., Schulze, 2005). However, by considering the mechanics of originality, some creators may be better capable to navigate their mainstream success or have a greater chance to become influential by bringing a “change in the established modes” (Smith, 1976) in their fields.

Fig. 1
figure 1

A diagram of innovation and originality. Notes: Originality, \(\omega _i\), is the length of individual i’s innovation, \(v_i\)

Fig. 2
figure 2

A diagram of positive and negative influence. Notes: Influence, \(\lambda _i\), is the (directed) distance between the projection of future style, \(p_f\), and the vertical line, the direction of which is given by past style

Fig. 3
figure 3

Relationship between originality, influence, and success

Fig. 4
figure 4

Theme of the Symphony No. 40 in G Minor by Wolfgang Amadeus Mozart

Fig. 5
figure 5

Originality and influence in Western music over time. Notes: The plot presents originality and influence of the work of all composers who were at least ten years of age in a specific year. The data was collected by the authors (see Sect. 3 for details)

Fig. 6
figure 6

Scatterplots of originality and different success measures. Notes: Scatter of baseline originality and different success measures: The log number of Spotify followers (correlation of 0.56), the log of the Murray index (correlation of 0.54), and the log of the word count (correlation of 0.48). Dots are labelled for top-5 composers based on the success measures

Fig. 7
figure 7

Originality, influence, and success. Notes: Panel A shows non-parametric regressions of the the success measure—the (log) number of Spotify followers on baseline originality. Panel B shows the same for baseline influence on baseline originality. Both graphs are the result of kernel regressions with a bandwidth of ten and using a Gaussian kernel. Adjusted predictions with \(95\%\) confidence intervals

Fig. 8
figure 8

Distribution of success by availability of composition periods. Notes: The histograms show the distribution of success for two groups of composers: Composers for whom at least one composition period could be retrieved and the remaining composers. For details on the data see Sect. 3.1

Fig. 9
figure 9

Originality and influence over age. Notes: Panel A shows for each theme the age of composer when the theme was written and theme-level originality. Panel B shows the same for the theme-level influence measure. Each observation in the scatter represents a theme for which a composition period was retrieved and theme-level originality and legacy were computed. Both graphs are the result of kernel regressions with a bandwidth of ten using a Gaussian kernel. For details on the data see Sect. 3.1

Fig. 10
figure 10

Near-term influence and overall influence. Notes: The scatter plot shows the relationship between near-term influence and overall influence. Near-term influence of composer i restricts the observation window of influence to the first generation of composers after the death of composer i. Overall influence considers influence to any composer after the death of composer i. See Sect. 2.3 for details on calculation and Sect. 3 for data sources. Dots are labelled for influential composers whose overall influence is more than five times larger than their near-term influence. The dashed line shows a linear fit

Fig. 11
figure 11

Distribution of originality for real and simulated themes. Notes: The histograms show the distribution of originality for actual themes of six notes (the most common length) and simulated themes of the same length. The simulation (see Sect. 4.5) draws uniformly from the set of notes that appear in our data, excluding rare notes that are less common in our data than 0.2%. For details on the data see Sect. 3.1

Table 1 Descriptive statistics for composers by century and region of birth
Table 2 Descriptive statistics for originality, influence, and success
Table 3 Descriptive statistics for composers by musical period
Table 4 Originality and success
Table 5 Originality and influence
Table 6 Theme-level originality and influence
Table 7 Theme-level originality and influence using theme composition period
Table 8 Originality, influence, and composer’s age