Author Correction: Quantifying the dynamics of failure across science, startups and security

Yian Yin, Yang Wang, James A. Evans, Dashun WangNature 2020 · DOI · OpenAlex

Abstract

Failure is often perceived as a random setback or a necessary stepping stone. However, research into failure streaks across science, startups, and security reveals that failure follows predictable mathematical dynamics. By analyzing decades of data, researchers identified a "tipping point" that separates eventual success from a cycle of stagnation. This separation is driven by how effectively individuals learn from previous attempts, a process reflected in the timing between consecutive failures. This article summarizes the core findings and details a recent author correction regarding the graphical representation of these dynamics.

Introduction

Conventional wisdom suggests that success comes either through sheer persistence ("try, try again") or the "Matthew Effect" (success breeds success). Yet, these narratives fail to explain the underlying mechanisms of failure. This research aims to provide a quantitative framework to predict whether a string of failures will lead to a breakthrough or if an individual is merely "treading water."

Methodology

The researchers developed a model-driven approach to analyze "failure streaks"—sequences of attempts by the same entity ending in failure, potentially culminating in success. The model posits that each attempt consists of a "new" component and a "reused" component from previous attempts. A key "learning parameter" represents how much an individual incorporates from their best previous attempt. The study analyzed the time interval between consecutive attempts ($T_n$) as a proxy for learning efficiency across three massive datasets: NIH grant applications (Science), VentureXpert investment records (Startups), and the Global Terrorism Database (Security).

Key Insights and the Tipping Point

The research identified a critical threshold in learning. Above this "tipping point," individuals incorporate enough information from past failures to eventually succeed. Below it, they fall into a "stagnant regime" where they make many attempts but never improve enough to cross the finish line.

The primary indicator of these regimes is the timing of subsequent attempts. In the successful group, the time between failures ($T_n$) decreases following a power law. In the unsuccessful group, the time between attempts stays flat. This suggests that those who succeed work "smarter" by systematically building on components of previous failed attempts.

Comparison of original and corrected Figure 3d–f, showing the divergence in timing between successful (blue) and unsuccessful (orange) groups.

Author Correction

The 2019 publication contained a graphical error in Figure 3d–f. The correction specifically fixes the y-axis plots showing the average time between attempts ($T_n$). In the corrected figures, the "Success" group (blue) shows a clear downward trend in $T_n$ as the number of attempts ($n$) increases, while the "Failure" group (orange) remains relatively flat. This correction addresses the misplacement of $10^0$ markers and the omission of $T_n$ labels, ensuring the visual data accurately reflects the study's conclusions. This was a typographic/formatting error and does not change the scientific conclusions of the original paper.

Conclusion and Limitations

Success is not merely a product of luck or persistence but of the efficiency of the "search" for a solution. By the second or third attempt, the change in time between attempts can predict with high accuracy whether an entity is in a success or failure regime. While the model treats learning as a "black box" based on output timing and does not account for external shocks, these insights offer potential for mentors and investors to identify stagnant streaks early and provide necessary interventions.

Summary updated

BibTeX
@inproceedings{yin2020author,
  author = {Yian Yin and Yang Wang and James A. Evans and Dashun Wang},
  title = {{Author Correction: Quantifying the dynamics of failure across science, startups and security}},
  booktitle = {Nature},
  year = {2020},
  doi = {10.1038/s41586-020-2325-6},
}
Figure