Understanding the onset of hot streaks across artistic, cultural, and scientific careers

Creative careers often exhibit periods of intense productivity known as "hot streaks," characterized by a cluster of high-impact works. While hot streaks are observed across various creative fields, the mechanisms behind their onset remain unclear. This study addresses this gap by investigating the relationship between creative strategies (exploration and exploitation) and the timing of hot streaks. Exploration involves diversification and experimentation, while exploitation focuses on refining existing skills and knowledge. The researchers hypothesize that the transition from exploration to exploitation might be a critical factor in the initiation of hot streaks. Understanding the emergence of hot streaks is significant for both theoretical understanding of creative life cycles and practical applications in talent identification and development. The unpredictability of hot streak timing further underscores the need for a systematic investigation to uncover underlying patterns.

Existing literature highlights the importance of exploration and exploitation in various fields, including organizational learning and innovation. Exploitation allows for deepening expertise in a specific area, potentially leading to sustained success. However, excessive exploitation can stifle originality. Exploration, though riskier with varied outcomes, increases the likelihood of discovering groundbreaking ideas through unexpected combinations. Prior research has studied these strategies separately or in combination, but less attention has been paid to their sequential relationship in driving career-level success, particularly concerning the emergence of hot streaks. This study builds upon previous work by quantitatively examining the temporal dynamics of exploration and exploitation in relation to hot streaks across diverse creative domains.

The researchers employed computational methods combining deep learning and network science to analyze large-scale datasets from three creative domains: visual arts, film directing, and scientific research. **Visual Arts:** Over 800,000 images were analyzed, covering the careers of 2128 artists. A deep neural network (based on a pre-trained VGGNet model) was used to generate a 200-dimensional embedding for each artwork, capturing both low-level features (e.g., brushstrokes) and high-level features (e.g., themes). Art styles were identified through clustering in this embedding space. **Film Directing:** Data comprised 79,000 films by 4337 directors from the IMDB database. A 200-dimensional representation for each film was created by combining word embeddings from plot descriptions and node embeddings (using DeepWalk) from a co-casting network. Film styles were identified via clustering. **Scientific Research:** The dataset included 20,040 scientists, tracking publication and citation data. Research topics were identified using a community detection algorithm applied to a co-citation network of publications. A node embedding method was also used as a robustness check. In all three domains, exploration and exploitation were quantified using topic/style entropy (H). H=0 represents pure exploitation (focus on a single style/topic), while H=log(n) represents pure exploration (even distribution across styles/topics), with n being the number of works. Entropy was calculated for periods before and during hot streaks, and compared against randomized null models. Hot streaks were defined using impact metrics (auction price for art, IMDB rating for films, and citations for publications). Temporal dynamics of entropy around hot streak onsets and ends were analyzed by aligning careers based on these events. For scientists, collaborations patterns were also investigated. Finally, predictive modelling was conducted to investigate the relationship between the exploration-exploitation sequence and the likelihood of a hot streak.

The key findings reveal consistent patterns across all three domains: 1. **Pre-Hot Streak Exploration:** Before a hot streak, individuals exhibit significantly higher entropy (H) than expected by chance, indicating a diverse range of explorations. Z-scores showed significant deviation from randomized models (e.g., 4.24 for artists before hot streaks). 2. **Hot Streak Exploitation:** During a hot streaks, entropy (H) drops significantly below expectation, reflecting a focused exploitation strategy. Z-scores for real careers during hot streaks consistently showed strong negative deviations from randomized careers (e.g., -2.42 for artists during hot streaks). 3. **Sequential Dynamics:** The transition from exploration to exploitation is strongly associated with the onset of hot streaks. Analyzing temporal changes in entropy shows a clear shift from high exploration to low exploitation around the hot streak onset. This sequence consistently outperforms other combinations (exploration alone, exploitation alone, exploitation-then-exploration) across all three domains. The probability of initiating a hot streak with the exploration-exploitation transition showed a considerable increase over baseline values (20.5%, 13.8%, and 19.2% for artists, directors, and scientists respectively). 4. **Post-Hot Streak Return to Baseline:** After a hot streak concludes, entropy returns to levels indistinguishable from those in randomized careers, indicating a cessation of distinct exploration or exploitation patterns. 5. **Collaboration Patterns in Science:** In scientific careers, team size decreases before hot streaks and increases afterwards, suggesting a shift from smaller, exploratory teams to larger, exploitative ones. 6. **Predictive Power:** Analysis of scientists' topic choices reveals that the subsequently exploited topic is not necessarily the most recently explored, highly cited, or popular one. Predictive modeling based on exploration phase features (team size and topic properties) accurately predicts the choice of exploited topic (accuracy of 0.89 and AUC of 0.83).

These findings demonstrate consistent patterns in the emergence of hot streaks across vastly different creative domains, challenging the notion of hot streaks as purely random events. The observed sequence of exploration followed by exploitation suggests a strategic interplay between diversification and refinement in achieving significant creative success. The moderate effect size indicates potential for further investigation of mediating variables such as collaboration patterns (particularly evident in science). The results support the idea that both exploration and exploitation are necessary, but their optimal sequence significantly enhances the likelihood of a hot streak. This sequence may increase creative possibilities by allowing new insights to be channeled into focused projects, while filtering less promising directions. The research contributes to a nuanced understanding of creative strategies and their role in career success.

This research uncovers a robust pattern of exploration followed by exploitation preceding hot streaks across artistic, cultural, and scientific careers. This sequence, rather than exploration or exploitation alone, appears critical for producing high-impact work. Future studies could explore domain-specific variations in this pattern and investigate the influence of additional factors such as external pressures and feedback mechanisms. The methodology used could also be applied to examine other creative domains and deepen our understanding of creativity.

The study is observational, limiting causal inferences. The datasets focus on individuals with sufficiently long careers, potentially excluding less prolific individuals. While the study uses large datasets, other factors influencing career trajectories might remain unidentified. Future research could address these limitations through causal designs and more comprehensive datasets.