Abstract
The hot streak—loosely defined as ‘winning begets more winnings’—highlights a specific period during which an individual’s performance is substantially better than his or her typical performance. Although hot streaks have been widely debated in sports^{1,2}, gambling^{3,4,5} and financial markets^{6,7} over the past several decades, little is known about whether they apply to individual careers. Here, building on rich literature on the lifecycle of creativity^{8,9,10,11,12,13,14,15,16,17,18,19,20,21,22}, we collected largescale career histories of individual artists, film directors and scientists, tracing the artworks, films and scientific publications they produced. We find that, across all three domains, hit works within a career show a high degree of temporal regularity, with each career being characterized by bursts of highimpact works occurring in sequence. We demonstrate that these observations can be explained by a simple hotstreak model, allowing us to probe quantitatively the hot streak phenomenon governing individual careers. We find this phenomemon to be remarkably universal across diverse domains: hot streaks are ubiquitous yet usually unique across different careers. The hot streak emerges randomly within an individual’s sequence of works, is temporally localized, and is not associated with any detectable change in productivity. We show that, because works produced during hot streaks garner substantially more impact, the uncovered hot streaks fundamentally drive the collective impact of an individual, and ignoring this leads us to systematically overestimate or underestimate the future impact of a career. These results not only deepen our quantitative understanding of patterns that govern individual ingenuity and success, but also may have implications for identifying and nurturing individuals whose work will have lasting impact.
Main
According to the Matthew effect^{9,23,24}, victories bring reputation and recognition that can translate into tangible assets, which in turn help to bring future victories. This school of thought supports the existence of a hot streak in a career, which is also consistent with literature in the field of innovation showing that peak performance clusters in time, typically occurring around the middle of a career^{8,11,21}. On the other hand, the random impact rule uncovered in the arts^{10,21} and sciences^{10,18} predicts the opposite: the best works occur randomly within a career, and their occurrence is primarily driven by productivity. Following this school of thought, works after a major breakthrough are not affected by what preceded them, supporting the viewpoint of regression towards the mean. The two divergent schools of thought raise a fundamental question: do hot streaks exist in creative careers?
To answer this question, we collected data sets recording the career histories of individual artists, film directors and scientists (Supplementary Information S1) and traced the impact of the artworks, films and papers they produced, approximated by auction prices^{15}, IMDB ratings (https://www.imdb.com/)^{25} and citations garnered after 10 years of publication (C_{10})^{13,16,18,26}, respectively (see Methods). We started by investigating the timing of the three most impactful works produced in each career. In a sequence of N works by an individual, we denoted with N* the position of the highestimpact work within a career, N** the second highest and N*** the third. We found that each of the three highestimpact works was randomly distributed among all the works produced by an individual (Extended Data Fig. 1a–c), offering strong endorsement for the random impact rule^{10,18,21}.
However, as we show next, the randomness in individual creativity is only apparent, because the timing between creative works follows highly predictable patterns. We measured the correlation between the timing of the two biggest hits within a career, and compared it with a null hypothesis in which N* and N** each occured at random. The normalized joint probability, ϕ(N*, N**) = P(N*, N**)/(P(N*)P(N**)), is substantially overrepresented along the diagonal elements of matrices (Fig. 1a–c), demonstrating that N* and N** are much more likely to colocate with each other than would be expected from the random impact model across three domains. The diagonal pattern disappears if we shuffle the order of works within each career, thereby breaking the temporal correlations (Extended Data Fig. 1j–r).
To quantify the temporal colocation of hits observed in Fig. 1a–c, we calculated the distance between the two highestimpact works for every individual, measured by the number of works produced in between, ΔN = N*−N**. We compared \(P(\frac{{\rm{\Delta }}N}{N})\) of real careers with \({P}_{{\rm{S}}}(\frac{{\rm{\Delta }}N}{N})\) of shuffled careers by defining \(R(\frac{{\rm{\Delta }}N}{N})=P(\frac{{\rm{\Delta }}N}{N}){\rm{/}}{P}_{{\rm{S}}}(\frac{{\rm{\Delta }}N}{N})\). For artists, directors, and scientists, all \(R(\frac{{\rm{\Delta }}N}{N})\) exhibit a clear peak centring around zero and decay quickly as ΔN deviates from zero (Fig. 1d–f). Notably, \(R(\frac{{\rm{\Delta }}N}{N})\) is mostly symmetric around zero (Fig. 1d–f), indicating that the biggest hit is equally likely to arrive before or after the second biggest. The colocation patterns are not limited to the two highestimpact works within a career. We repeated our analyses for other pairs of hit works, such as N** versus N*** and N* versus N***, and uncovered the same colocation patterns (Extended Data Fig. 1d–i).
Do high impact works come in streaks within a career? We counted the number of consecutive works whose impact exceeded the median of all works within a career (Extended Data Fig. 2d–f). We calculated the length of the longest streak L for each career. We then shuffled the order of works within each career, and measured again their longest streaks L_{s}. P(L) was characterized by a much longer tail than P(L_{s}) (Fig. 1g–i), indicating that real careers are characterized by long streaks of relatively highimpact works clustered together in sequence. We tested the robustness of these results by controlling for individual career length, and by varying our threshold used to calculate L, and arrived at the same conclusions (Extended Data Figs. 2–4, Supplementary Information S2). Together, these results raise an important question: what mechanisms are responsible for the temporal regularities observed across diverse career histories?
Let us first consider a null model in which the goodness of works produced in a career (that is, log(price) for artists, ratings for directors, and log(C_{10}) for scientists) is drawn from a normal distribution \({\mathscr{N}}({{\rm{\Gamma }}}_{i},{\sigma }_{i}^{2})\) that is fixed for an individual. The average Γ_{i} characterizes the typical impact of works produced by the individual, and σ_{i} captures the variance. This null model can reproduce the fact that each hit occurs randomly within a career^{10,18}. However, it fails to capture any of the temporal correlations observed in Fig. 1. The main reason for this failure is illustrated in Fig. 2a–c, where we selected for illustration purposes one individual from each of the three data sets and measured the dynamics of Γ_{i} during his or her career. We find that Γ_{i} is not constant throughout a career. Rather, it deviates from a baseline performance (Γ_{0}) at a certain point in a career (t_{↑}), elevating to a higher value Γ_{H} (Γ_{H} > Γ_{0}), which is then sustained for some time before falling back to level similar to Γ_{0} (Fig. 2a–c):
This observation, combined with the shortcomings of the null model, raises an intriguing question: could a simple model based on equation (1) explain the temporal anomalies documented in Fig. 1?
To test this hypothesis, we applied equation (1) to real productivity patterns, allowing us to generatively simulate the impacts of the works produced by an individual (Supplementary Information S3.3). During the period in which Γ_{H} operates, the individual seemingly performs at a higher level than his or her typical performance (Γ_{0}), prompting us to call this model the hotstreak model (where the Γ_{H} period corresponds to the hot streak). We introduced to each career one hot streak that occured at random with a fixed duration and magnitude, and repeated our measurements in Fig. 1 on careers generated by the model. We find that, whereas equation (1) introduces only a simple temporal variation, the hotstreak model is sufficient to reproduce all empirical patterns observed in Fig. 1 (Fig. 1d–i and Extended Data Fig. 1s–u). Given the myriad factors that can affect career impacts^{9,10,11,12,18,22,27,28,29,30}, and the obvious diversity of careers we studied, the level of universality and accuracy demonstrated by the simple hotstreak model was unexpected.
The real value of the model arises, however, when we fit the model to real careers to obtain the individual specific Γ_{0}, Γ_{H}, t_{↑} and t_{↓} parameters (Supplementary Information 3.4), helping us to reveal several fundamental patterns that govern individual careers.
1. Hot streaks are ubiquitous across careers, yet at the same time usually unique within a career. The vast majority of artists (91%, Fig. 2d), film directors (82%, Fig. 2e) and scientists (90%, Fig. 2f) have at least one hot streak during their careers, documenting the practical relevance of the uncovered hot streak phenomenon. However, despite its ubiquity, the hot streak is likely to be unique within a career. Indeed, when we relaxed our fitting algorithm to allow for multiple hot streaks (up to three) with different values of Γ_{H}, we found that, among those who had a hot streak, 64% of artists, 80% of directors, and 68% of scientists were best captured by one hot streak only (Fig. 2d–f), documenting the precious nature of hot streaks. Second acts may occur but are less likely, particularly for film directors. Occurrences of more than two hot streaks are rare across all careers.
2. The hot streak occurs randomly within a career. We estimate the beginning of hot streaks, by measuring N_{↑}, the position of work produced when a hot streak starts (t_{↑}). We find that, across artistic, cultural, and scientific careers, N_{↑} is randomly distributed in the sequence of N works within a career (Fig. 2g–i, insets). This finding reconciles two seemingly divergent schools of thought^{9,10,18}, providing a further explanation for the random impact rule: if the hot streak occurs randomly within a career, and the highest impact works are statistically more likely to appear within a hot streak, then the timing of the highest impact works is also random.
3. Across different domains, hot streaks are considerably shorter than the typical career length recorded in our database. We measure the duration distribution of hot streaks (τ_{H} = t_{↑}−t_{↓}), finding P(τ_{H}) peaks around 5.7 years for artists, 5.2 years for directors, and 3.7 years for scientists, which is largely independent of when it occurs within a career (early, mid or late career; Fig. 2g–i).
4. Unexpectedly, individuals are not more productive during hot streaks. We measured the distribution of the total number of works produced during hot streaks P(N_{H}). We then constructed a null distribution, by randomly picking one work in a career and designating its production year to be the start of the hot streak. We found that the two distributions aligned well with each other (Fig. 2j–l). Therefore, individuals show no detectable change in productivity during hot streaks, despite the fact that their outputs in this period are significantly better than the median, suggesting that there is an endogenous shift in individual creativity when the hot streak occurs. For additional properties of hot streaks, see Methods and Extended Data Fig. 5.
To investigate the impact of hot streaks on individual careers, we focused on scientific careers and measured the collective impact of a scientist, g(t), defined as the total number of citations over time collected by all papers published by an individual (Fig. 3a). g(t) can be derived analytically by combining the hotstreak model (equation (1)) and an existing model^{16} for the citation patterns of papers (see Methods and Supplementary Information S5), consisting of two terms:
g_{0}(t) captures a career’s collective impact in the absence of a hot streak (that is, Γ(t) = Γ_{0}). Contributions from the hot streak are encoded in Δg(t), driven by both the timing and magnitude of hot streaks (see Methods). Varying hotstreak parameters leads to substantial changes in the collective impact of a career (Fig. 3b). Hence the hotstreak model captures a wide range of impact trajectories that are followed by real careers (Fig. 3c), and the accuracy of the model is documented by several metrics (see Methods). Given that individuals improve substantially during hot streaks, the uncovered phenomena can be particularly crucial for understanding the longterm impact of a career (Extended Data Fig. 6).
We further tested several alternative hypotheses, each associated with possible origins of the uncovered hot streaks (see Methods and Supplementary Information S6). Of all hypotheses considered, the hotstreak model is the simplest and least flexible. However, it is the only model whose predictions are consistent with real careers (Extended Data Figs 7, 8). The identification of the true origins of hot streaks is beyond the scope of this work. As such, the hot streaks uncovered here should be treated in a metaphorical sense, highlighting an intriguing period of outstanding performance during individual careers without implying any associated drivers. Crucially, though, the findings presented here hold the same, regardless of the underlying drivers.
Methods
Data description
We compiled three largescale data sets of individual careers across three major domains involving human creativity. The first data set (D_{1}) consists of auction records curated from online auction databases, allowing us to reconstruct the career histories of 3,480 artists through the sequence of works they each produced, together with the impacts of the artworks, approximated by hammer prices in auctions^{15}. D_{2} contains profiles of 6,233 film directors recorded in the IMDB database, each career being represented by the sequence of films directed by the individual. As metrics that quantify the impact of a film correlate closely with each other^{25}, here we use the IMDB rating to measure the goodness of a film. Finally, our third data set (D_{3}) includes publication records of 20,040 individual scientists through a largescale name disambiguation effort that combined the Web of Science (https://clarivate.com/products/webofscience/) and Google Scholar (https://scholar.google.co.uk/) data sets. The impact of each paper is measured by citations garnered after 10 years of its publication^{13,16,18,26} (C_{10}). Further details on data collection and curation are provided in Supplementary Information S1.
To study the impact of works across the three domains, we measured the distributions of hammer prices, IMDB ratings and paper citation counts in our data sets. Both hammer price (D_{1}) and C_{10} (D_{3}) follow fattailed distributions, well approximated by a lognormal function (Extended Data Fig. 2a, c), and the IMDB rating follows a normal distribution ranging between 1 and 10 (Extended Data Fig. 2b). To make sure C_{10} is not affected by citation inflation^{14,18,31}, we also measured a rescaled C_{10} (see Supplementary Information S1.3) and found that it also followed a fattailed distribution (Extended Data Fig. 2c, inset). Therefore, we take the logarithmic of hammer price and C_{10} (log(price) and log(C_{10})) to approximate the goodness of an artwork and scientific publication. Note that the choice of logarithmic for hammer prices and C_{10} is meant to be consistent with prior studies^{14,18}, and does not affect any of the conclusions of the paper. Indeed, the logarithmic function is a monotonically increasing function, hence it does not change the rank ordering of top hits in a career. Note that while the data sets we used in this paper cover a large collection of career histories across a wide range of domains, the datadriven nature of our study indicates that the scope of our data is limited to individuals who have had sufficiently long careers to provide enough data points for statistical analyses (Supplementary Information S1).
Additional properties of hot streaks
How much does an individual deviate from his or her typical performance during a hot streak? Do people with higher Γ_{0} also experience more performance gain from hot streaks? We explored correlations between Γ_{0} and Γ_{H}, finding them to be well approximated by a linear relationship (Extended Data Fig. 5a–c). Hence, individuals with better typical performance also perform better during their hot streaks. It is interesting to note that the coefficients are slightly less than 1 (Extended Data Fig. 5a–c). Hence ΔΓ ≡ Γ_{H} − Γ_{0} decreases with Γ_{0} (Extended Data Fig. 5a–c, insets), suggesting that individuals with smaller Γ_{0} benefit more from hot streaks. These results are again independent of when the hot streak occurs along a career (Extended Data Fig. 5a–c).
The temporally localized nature of the hot streak is also captured by its proportion over career length τ_{H}/T (Extended Data Fig. 5d–f). We compared the duration of hot streaks with typical career length, finding that the median hovers around 20% (0.17 for artists, 0.23 for directors, and 0.20 for scientists).
Analytical solutions for the collective impact of a scientific career, g(t)
Brought into the spotlight by popular websites such as Google Scholar, g(t) is playing an increasingly important role in driving many critical decisions, from hiring, promotion and tenure to awarding of grants and rewards. Many factors are known to influence it, ranging from productivity^{17,28} to citation disparity and dynamics^{13,14,16,22,23,29} and temporal inhomogeneities along a career^{11,17,18,21,22,30}. As our goal is to understand impact, here we bypass the need to evaluate the inhomogeneous nature of productivity^{17,18} by rearranging the publication time of each paper, such that an individual produces a constant number of papers each year, denoted by n (Fig. 3a). To calculate g(t), we need to incorporate the citations patterns of papers into our hotstreak model (equation (1)). A recent study^{16} suggested that the citation dynamics of a paper published at time t_{0} can be approximated by
where m is a global parameter describing the typical number of references a paper contains, and Φ(·) is the cumulative normal function, characterized by μ and σ, which capture the typical citation life cycle of a paper. The paper’s ultimate impact is determined by its fitness^{16}, λ. To adapt equation (3) into our framework, we replace λ with Γ(t_{0}), and for simplicity assume that μ and σ are fixed for different papers published by an individual. The resulting model, combining equations (1) and (3), can be solved analytically (Supplementary Information S5), allowing us to express g(t) in terms of hotstreak parameters:
Equation (4) consists of two terms. g_{0}(t) captures a career’s collective impact in the absence of hot streaks (that is, Γ(t) = Γ_{0}). Contributions from the hot streak are encoded in Δg(t), driven by both the timing and magnitude of hot streaks (t_{↑}, t_{↓}, Γ_{H}, and Γ_{H} − Γ_{0}).
Evaluating the accuracy of the hotstreak model
We quantify the accuracy of our model in equation (4) using three metrics.
To account for the inherently noisy career trajectories, we first assign an impact envelope to each career, explicitly quantifying the uncertainty of model predictions (Extended Data Fig. 6g). We simulated g(t) for each individual by assigning a Gaussian noise \({\mathscr{N}}(0,{\sigma }_{{\rm{s}}}^{2})\) to the fitted Γ_{0} and Γ_{H}. For each paper i we randomly draw its Γ_{i} from a normal distribution, depending on whether the paper was published within the hot streak (Γ_{H} during hot streak, Γ_{0} otherwise). The standard deviation σ_{s} represents the inherent noise of the goodness parameter defined in Supplementary Information S3.5. For each individual, we simulated g(t) for 1,000 realizations, allowing us to obtain a distribution of g(t), with one standard deviation offering an uncertainty envelope. We repeated the same procedures for the null model. We measure the fraction of g(t) that falls within the envelope, finding that the distribution peaks close to 1 (Extended Data Fig. 6h), which indicates that most career trajectories are well encapsulated within the predicted envelopes.
The superior accuracy of our model is also captured by the mean absolute percentage error (MAPE). We compared the distribution of MAPE between the data and the predictions of the model (Extended Data Fig. 6f), finding again that the hotstreak model outperformed the null model. The improvement was most pronounced for an early onset of hot streaks (Extended Data Fig. 6i), which is also consistent with our model’s predictions.
To account for model complexity, we also calculated the Bayesian information criterion (BIC) measure, which penalizes the number of parameters in the model. Compared with the null model, the hotstreak model has systematically smaller BIC (Extended Data Figs 6e), documenting that the hotstreak model better captures the collective impact of a career than the null model.
Implications of hot streaks for longterm career impact
The analytical framework presented here not only offers a new theoretical basis for our quantitative understanding of dynamical patterns governing individual career impact, but also may have implications for comparing and evaluating scientists (Extended Data Fig. 6). Indeed, for individuals whose hot streaks are yet to come, ignoring the hot streak may lead to underestimation of their impacts (Extended Data Fig. 6a, b), especially given the ubiquitous nature of hot streaks (Fig. 2f). On the other hand, an early onset of a hot streak leads to a high impact that peaks early but may not be sustained unless a second hot streak occurs (Extended Data Fig. 6c).
Testing alternative hypotheses
To explore the possibility that alternative hypotheses might explain the observed patterns, we tested several models that capture different dynamics of hot streaks (Supplementary Information S6.3), each associated with possible origins of the uncovered hot streaks. (A) A right trapezoid (Extended Data Fig. 7b) captures a sudden onset of a hot streak with a more gradual decline, as innovators may stumble upon a groundbreaking idea, which manifests itself in the forms of multiple artworks, films, or publications. Hence from an evolutionary perspective, the duration of a hot streak may characterize the time it takes for the temporary competitive advantage to dissipate. (B) An isosceles trapezoid model (Extended Data Fig. 7c) captures hot streaks that evolve and dissolve gradually over time, which may approximate social tie dynamics, as one individual’s hot streak could be the result of a fruitful, repeated collaboration^{27,32}. (C) Furthermore, individual performance may peak at a certain point in a career, prompting us to test invertedU shape (Extended Data Fig. 7d) and tent functions (Extended Data Fig. 7e). (D) Last, a left trapezoid function (Extended Data Fig. 7f) captures a gradual startup period with a sharp cutoff, corresponding to career opportunities that can augment impact but last for a fixed duration.
We tested the validity of the four alternative hypotheses (A–D) by comparing each model’s prediction with empirical observations on the relative order of the top six hits within a career. The symmetric patterns of ϕ and \(R(\frac{{\rm{\Delta }}N}{N})\) observed in real careers suggest that the biggest hit is equally likely to appear before or after the secondbiggest hit. The randomness of the relative ordering among hits is not limited to the two biggest hits. Indeed, we measured the position of the top three hits (\({\tilde{N}}\): mover) relative to the top six hits of the career, and compute \(P({\tilde{N}})\) for each of the three hits for artists, directors and scientists. We found a lack of predictive patterns for \(P({\tilde{N}})\) across the three domains, suggesting that the relative orders among the top six hits in real careers are random (Extended Data Fig. 7g, o, w). We tested hypotheses A–D systematically to describe real careers (Supplementary Information S6.3), and found that the hotstreak model was the only model whose predictions were consistent with real careers (Extended Data Fig. 7h–m, p–u, x–ac). As such, the hotstreak model also offers a superior fit to the data than the other models (Extended Data Fig. 7n, v, ad).
We also tested whether Markov models could account for our observations (Supplementary Information S6.2). We explored multiple variants of Markov models by introducing shortrange correlations between the impacts of adjacent works, correlations between the volatility of their impacts, and hidden Markov model with two states, finding again that the hotstreak model stood out in its ability to describe the observed patterns (Extended Data Fig. 8 and Supplementary Information S6.2). Together, these results demonstrate that none of these alternative hypotheses alone can account for the empirical observations in real careers.
Code availability
Code is available at https://luliu.github.io/hotstreaks/.
Data availability
The data are available at https://luliu.github.io/hotstreaks/.
References
 1.
Gilovich, T., Vallone, R. & Tversky, A. The hot hand in basketball: on the misperception of random sequences. Cognit. Psychol. 17, 295–314 (1985).
 2.
Miller, J. B. & Sanjurjo, A. Surprised by the gambler’s and hot hand fallacies? A truth in the law of small numbers. IGIER Working Paper No. 552 (2016).
 3.
Ayton, P. & Fischer, I. The hot hand fallacy and the gambler’s fallacy: two faces of subjective randomness? Mem. Cognit. 32, 1369–1378 (2004).
 4.
Rabin, M. & Vayanos, D. The gambler’s and hothand fallacies: theory and applications. Rev. Econ. Stud. 77, 730–778 (2010).
 5.
Xu, J. & Harvey, N. Carry on winning: the gamblers’ fallacy creates hot hand effects in online gambling. Cognition 131, 173–180 (2014).
 6.
Hendricks, D., Patel, J. & Zeckhauser, R. Hot hands in mutual funds: shortrun persistence of relative performance, 1974–1988. J. Finance 48, 93–130 (1993).
 7.
Kahneman, D. & Riepe, M. W. Aspects of investor psychology. J. Portfol. Manage. 24, 52–65 (1998).
 8.
Lehman, H. C. Age and Achievement (Princeton Univ. Press, Princeton, 1953).
 9.
Merton, R. K. The Matthew effect in science. Science 159, 56–63 (1968).
 10.
Simonton, D. K. Age and outstanding achievement: what do we know after a century of research? Psychol. Bull. 104, 251–267 (1988).
 11.
Jones, B. F. The burden of knowledge and the death of the renaissance man: is innovation getting harder? Rev. Econ. Stud. 76, 283–317 (2009).
 12.
Petersen, A. M. et al. Reputation and impact in academic careers. Proc. Natl Acad. Sci. USA 111, 15316–15321 (2014).
 13.
Stringer, M. J., SalesPardo, M. & Nunes Amaral, L. A. Effectiveness of journal ranking schemes as a tool for locating information. PLoS One 3, e1683 (2008).
 14.
Radicchi, F., Fortunato, S. & Castellano, C. Universality of citation distributions: toward an objective measure of scientific impact. Proc. Natl Acad. Sci. USA 105, 17268–17272 (2008).
 15.
Galenson, D. W. Old Masters and Young Geniuses: The Two Life Cycles of Artistic Creativity (Princeton Univ. Press, Princeton, 2011).
 16.
Wang, D., Song, C. & Barabási, A.L. Quantifying longterm scientific impact. Science 342, 127–132 (2013).
 17.
Way, S. F., Morgan, A. C., Clauset, A. & Larremore, D. B. The misleading narrative of the canonical faculty productivity trajectory. Proc. Natl Acad. Sci. USA 114, E9216–E9223 (2017).
 18.
Sinatra, R., Wang, D., Deville, P., Song, C. & Barabási, A.L. Quantifying the evolution of individual scientific impact. Science 354, aaf5239 (2016).
 19.
Fortunato, S. et al. Science of science. Science 359, eaao0185 (2018).
 20.
Zeng, A. et al. The science of science: From the perspective of complex systems. Phys. Rep. 714–715, 1–73 (2017).
 21.
Simonton, D. K. Scientific Genius: A Psychology of Science (Cambridge Univ. Press, Cambridge, 1988).
 22.
Duch, J. et al. The possible role of resource requirements and academic careerchoice risk on gender differences in publication rate and impact. PLoS One 7, e51332 (2012).
 23.
Price, D. S. A general theory of bibliometric and other cumulative advantage processes. J. Assoc. Inf. Sci. Technol. 27, 292–306 (1976).
 24.
Barabási, A.L. & Albert, R. Emergence of scaling in random networks. Science 286, 509–512 (1999).
 25.
Wasserman, M., Zeng, X. H. T. & Amaral, L. A. N. Crossevaluation of metrics to estimate the significance of creative works. Proc. Natl Acad. Sci. USA 112, 1281–1286 (2015).
 26.
Uzzi, B., Mukherjee, S., Stringer, M. & Jones, B. Atypical combinations and scientific impact. Science 342, 468–472 (2013).
 27.
Wuchty, S., Jones, B. F. & Uzzi, B. The increasing dominance of teams in production of knowledge. Science 316, 1036–1039 (2007).
 28.
Shockley, W. On the statistics of individual variations of productivity in research laboratories. Proc. IRE 45, 279–290 (1957).
 29.
Redner, S. Citation statistics from 110 years of physical review. Phys. Today 58, 49 (2005).
 30.
Moreira, J. A., Zeng, X. H. T. & Amaral, L. A. N. The distribution of the asymptotic number of citations to sets of publications by a researcher or from an academic department are consistent with a discrete lognormal model. PLoS One 10, e0143108 (2015).
 31.
Pan, R. K., Petersen, A. M., Pammolli, F. & Fortunato, S. The memory of science: inflation, myopia, and the knowledge network. Preprint at https://arxiv.org/abs/1607.05606 (2016).
 32.
Palla, G., Barabási, A.L. & Vicsek, T. Quantifying social group evolution. Nature 446, 664–667 (2007).
Acknowledgements
We thank A.L. Barabasi, R. Burt, J. Chown, J. Evans, C. Jin, L. Nordgren, W. Ocasio, B. Uzzi, Y. Yin, the Kellogg Insights and all members of NICO for invaluable comments. This work is supported by the Air Force Office of Scientific Research (AFOSR) under award number FA95501510162 and FA95501710089, and Northwestern University’s Data Science Initiative. R.S. acknowledges support from AFOSR grant FA95501510364 and from the Central European University Intellectual Themes Initiative ‘Just Data’.
Reviewer information
Nature thanks J. Walsh, J. West and the other anonymous reviewer(s) for their contribution to the peer review of this work.
Author information
Affiliations
Contributions
D.W. conceived the project and designed the experiments; L.L. and Y.W. collected data and performed empirical analyses with help from R.S., C.L.G., C.S. and D.W.; L.L., Y.W., C.S. and D.W. did theoretical calculations; all authors discussed and interpreted results; D.W. and L.L. wrote the manuscript; all authors edited the manuscript.
Corresponding author
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Extended data figures and tables
Extended Data Fig. 1 Additional results on hot streaks in artistic, cultural, and scientific careers.
a–c, The cumulative distribution P(≥N^{i}/N) for the order of the top three highest impact works within a career for artists (a), directors (b) and scientists (c). N^{i} denotes the order of the i^{th} highestimpact work within a career. The colours denote different hit works, and the dashed grey line denotes P(≥N^{i}/N) for a uniform distribution. d–f, ϕ(N**, N***) for the second and thirdhighestimpact works within a career. ϕ(N**, N***) is also overrepresented along the diagonal. g–i, ϕ(N*, N***) for the first and thirdhighestimpact works within a career. j–r, We shuffled the order of each work in a career while keeping their impact intact. The diagonal patterns in d–i and Fig. 1a–c disappeared for shuffled careers. s–u, ϕ(N*, N**) predicted by the hotstreak model successfully recovered the diagonal patterns observed in a–c. For d–u and Fig. 1a–c, we applied the same binning procedure to data, using bins that ranged from 0 to 1 with increments of 0.1.
Extended Data Fig. 2 Measuring the length of streaks using different thresholds.
a, The distribution of auction price P(Price) for artists. Blue dots denote data, and the red line is a lognormal distribution with average μ = 7.9 and standard deviation σ = 1.5. b, The distribution of film rating P(Rating) for directors. The red line is a normal distribution with average μ = 7.1 and standard deviation σ = 1.2. c, The distribution of raw and rescaled C_{10} (inset) for scientists. The red line is a lognormal distribution, with μ = 2.3 and σ = 1.3 for c and μ = −0.4 and σ = 0.8 for the inset. d–f, Definitions of the longest streak L for artists (d), directors (e) and scientists (f). Dots are coloured orange above the threshold, blue otherwise. The lower panel highlights the longest streak in a career. g–i, P(L) for real careers and P(L_{s}) for shuffled careers using the mean impact within a career as the threshold. j–l, As in g–i, but using the top 10% impact as the threshold to calculate L and L_{s}. In all cases, P(L) has a wider tail than P(L_{s}), indicating that highimpact works in real careers tend to cluster together.
Extended Data Fig. 3 Varying career length.
To test the robustness of our results, we repeated our measurements by controlling for the career length of individuals. a–i, Artists and directors with careers of at least 20 years and scientists with careers of at least 30 years. a–c, P(≥N^{i}/N) of the top three highestimpact works within a career. d–f, \(R(\frac{{\rm{\Delta }}N}{N})\) among the top three highestimpact works in a career. g–i, P(L) for real careers and P(L_{s}) for shuffled careers. j–r, As in a–i but for artists and directors with careers of at least 30 years and scientists with careers of at least 40 years. These results demonstrate that the patterns observed in Fig. 1 hold for individuals with different career lengths.
Extended Data Fig. 4 Artistic careers from different eras.
a, \(R(\frac{{\rm{\Delta }}N}{N})\) for artists who started their careers before 1850. b, P(L) for real careers and P(L_{s}) for shuffled careers for artists who started their careers before 1850. c, \(R(\frac{{\rm{\Delta }}N}{N})\) for artists who started careers between 1850 and 1900. d, P(L) for real careers and P(L_{s}) for shuffled careers for artists who started their careers between 1850 and 1900. These results demonstrate that the patterns observed in Fig. 1 hold for artists from different eras.
Extended Data Fig. 5 Additional properties of hot streaks.
a–c, Correlations between Γ_{H} and Γ_{0} for artists (a; n = 3,166), directors (b; n = 5,098) and scientists (c; n = 18,121). The blue background denotes the kernel density of data, dots represent binning results of data, and the red lines depict a linear fit. Inset, the relationship between ΔΓ ( = Γ_{H} − Γ_{0}) and Γ_{0}. d–f, The distribution of τ_{H}/T, representing the duration of hot streaks over total career lengths. The temporally localized nature of a hot streak is also captured by its proportion over career length τ_{H}/T.
Extended Data Fig. 6 Comparison of g(t) between the null model and the hotstreak model.
a–c, g(t) of three scientists in our data set with midcareer (a), latecareer (b) and earlycareer (c) onset of hot streaks. Red dots denote data, the blue line is the null model’s prediction based on early performance, and the red line captures the predictions from the hotstreak model, with dashed grey lines denoting the start and end of hot streaks. d, The difference between our hotstreak model and the null model for each individual, Δg(t). Dashed lines with corresponding colours denote the start of the hot streak. d illustrates the discrepancies in estimating an individual’s future impact if we ignore the uncovered hot streaks. e, The distribution of the BIC measure, P(BIC), showing that the hotstreak model outperforms the null model in describing g(t) after accounting for model complexity. f, The distribution of the MAPE measure, P(MAPE), showing that the hotstreak model outperforms the null model in describing g(t). g, The uncertainty envelope of g(t) for an individual in our data set. Blue dots denote data, and the red line is the fitting result of equation (4). Shaded area illustrates predicted uncertainty (one standard deviation). h, The fraction of g(t) falling within the envelope for the null model (blue) and our hotstreak model (red). Fraction = 1.0 indicates that the entire g(t) trajectory falls within the envelope. i, Average MAPE of our hotstreak model and the null model for individuals with earlycareer, midcareer and latecareer onset of hot streaks. The difference is largest for individuals with earlyonset hot streaks and smallest for those with lateonset ones.
Extended Data Fig. 7 Testing alternative hotstreak dynamics.
a–f, Illustrative examples of Γ(N) for the hotstreak model (a), right trapezoid function (b), isosceles trapezoid function (c), quadratic function (d), tent function (e) and left trapezoid function (f). g, The distribution of the relative position \(P({\tilde{N}})\) of the three highestimpact works among the six highestimpact works within a career for artists, where \({\tilde{N}}\) denotes the relative order among the top six hits. h–m, \(P({\tilde{N}})\) predicted by corresponding models shown in a–f, respectively, according to artists’ real productivity profiles. To test whether data agree with model predictions, we measured their statistical difference using the P value of the Kolmogorov–Smirnov test for discrete distributions. We colour the distributions green if we cannot reject the hypothesis that the data and the model predictions come from the same distributions, and red otherwise. Among the six models, the hotstreak model is the only model whose predictions are consistent with the data in terms of the relative ordering among the six highestimpact works observed in real careers. n, The proportion of real careers that are captured by the model with the smallest BIC among different hypotheses. The hotstreak model again stands out as the best model to describe real careers. We repeated the analyses for directors (p–v) and scientists (x–ad), the conclusions remained the same across all three domains.
Extended Data Fig. 8 Testing Markovian hypotheses.
Here we test whether the observed patterns can be explained by Markovian dynamics that introduce correlations between neighbouring data points. We first test the assumptions of the Markovian hypothesis from the data (a–f). a–c, The distribution of N, N + 1 differences between adjacent data points observed in real careers for artists (a, n = 3,480), directors (b, n = 6,233) and scientists (c, n = 20,040). d–f, The autocorrelation measured in real careers for artists (d, n = 3,480), directors (e, n = 6,233), and scientists (f, n = 20,040). a–f suggest that there is little shortrange correlation in data across the three domains. We test three variants of Markovian models (g–l). The details of these models are outlined in Supplementary Information S6.2. g–i, ϕ(N*, N**) of the top two highestimpact works within a career for three Markovian models using scientists’ profiles as input. j–l, The distribution of the longest streak length P(L) and P(L_{s}) using median impact within a career as threshold for the three Markovian models. g–l demonstrate that the three Markovian models failed to capture the observed colocations among hits.
Extended Data Fig. 9 Additional examples of Γ.
a–c, Each subplot denotes the fitting result on Γ sequence for a randomly selected career for artists (a), directors (b) and scientists (c). Blue dots denote the moving average Γ(N) from data and red lines denote the best fitting result of the hotstreak model for each individual.
Extended Data Fig. 10 Individuals with one or more hot streaks.
a–c, The distribution of average impacts for individuals with one or more than one hot streaks for artists (a), directors (b) and scientists (c). Blue dots denote individuals with one hot streak, and red dots denote individuals with at least two hot streaks. d–f, The distribution of the number of works P(N) within a career for individuals with one or more than one hot streak for for artists (d), directors (e) and scientists (f). g–i, The distribution of career length P(τ) for individuals with one or more than one hot streaks for artists (g), directors (h) and scientists (i). j–l, The distribution of P(Γ_{H}) for individuals with one or more than one hot streaks for artists (j), directors (k) and scientists (l). Between those who have one or two hot streaks, there is no detectable difference in terms of typical performance metrics, including impact, productivity and career length, suggesting that the hot streak captures an orthogonal dimension to current metrics characterizing individual careers.
Supplementary information
Supplementary Information
This file contains supplementary information S1S6; which includes supplementary tables S1 and S2.
Rights and permissions
About this article
Cite this article
Liu, L., Wang, Y., Sinatra, R. et al. Hot streaks in artistic, cultural, and scientific careers. Nature 559, 396–399 (2018). https://doi.org/10.1038/s4158601803158
Received:
Accepted:
Published:
Issue Date:
Further reading

The Gene of Scientific Success
ACM Transactions on Knowledge Discovery from Data (2020)

Investigating the quantity–quality relationship in scientific creativity: an empirical examination of expected residual variance and the tilted funnel hypothesis
Scientometrics (2020)

Recency predicts bursts in the evolution of author citations
Quantitative Science Studies (2020)

Renormalizing individual performance metrics for cultural heritage management of sports records
Chaos, Solitons & Fractals (2020)

The Relation of Scientific Creativity and Evaluation of Scientific Impact to Scientific Reasoning and General Intelligence
Journal of Intelligence (2020)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.