Statistical Models for Company Growth

Matthieu Wyart, Jean‐Philippe BouchaudSSRN 2003 · DOI

TL;DR

This paper explores statistical models for company growth, aiming to explain why larger companies' growth rates fluctuate less than smaller ones. The authors analyze Sutton's "microcanonical" model and propose an alternative based on power-law distributions of sub-sector sizes, offering testable predictions to differentiate between the models.

What Problem Does This Paper Address?

The paper tackles the "scaling puzzle" in economics: empirical observations show that the standard deviation of a company's growth rate decreases with its size ($S$) as $\sigma(S) \sim S^{-\beta}$, with $\beta \approx 0.18$. This is counterintuitive, as the Central Limit Theorem would suggest $\beta = 1/2$ if companies were simply aggregations of independent, weakly correlated sub-units. The paper investigates existing models and proposes a new one to explain this observed deviation from standard statistical physics predictions.

Methods

The authors employ methods from statistical physics, particularly Fourier-Laplace transforms and saddle-point approximations, to analyze their models.

  • Sutton's Model Analysis: They revisit Sutton's "microcanonical" model, which assumes all partitions of a company's size into sub-pieces are equiprobable. They derive predictions for the distribution of growth rates and sub-sector sizes.
  • Alternative Model Proposal: They introduce a new model where the sizes of sub-entities follow a power-law distribution (Pareto distribution). This is motivated by empirical observations of company size distributions and a proposed dynamical model for sector growth involving resource redistribution.
  • Analytical Derivations: Both models are analyzed mathematically to derive scaling exponents ($\beta$) and the shape of the growth rate distributions.

Datasets

Not discussed in the paper. The authors refer to empirical data from Stanley et al. [1, 2] and Sutton [4] for comparison but do not use specific datasets for their model derivations or evaluations.

Key Insights and Ideas

  • Sutton's Model vs. Empirics: Sutton's model, while predicting a $\beta = 1/4$ scaling for growth rate variance, leads to an asymptotically Gaussian distribution of rescaled growth rates. This contradicts empirical findings that suggest non-Gaussian tails, especially for larger companies.
  • Power-Law Sub-Sector Sizes: The proposed alternative model, assuming power-law distributed sub-sector sizes, can naturally produce non-Gaussian growth rate distributions and a $\beta$ value that interpolates between $1/2$ and $0$.
  • Testable Predictions: The paper offers distinct, testable predictions for the distribution of sub-sector sizes: Sutton's model predicts a Bose-Einstein distribution, while the new model predicts a power-law distribution for smaller sub-sectors.
  • Aggregation Stability: The proposed power-law model is shown to be stable upon aggregation, meaning that the scaling properties observed at the company level are preserved at the country level (GNP growth), which is consistent with empirical observations. Sutton's model, however, is not stable in this regard.

Key Results

  • Sutton's Model: Predicts $\beta = 1/4$ for the variance of growth rate decay with size. The rescaled growth rate distribution is asymptotically Gaussian, though finite-size corrections can lead to non-Gaussian behavior for moderate company sizes.
  • Alternative Model:
    • For $1 < \mu < 2$ (where $\mu$ is the exponent of the power-law sub-sector size distribution), the variance of the relative growth rate decays as $S^{-\beta}$ with $\beta = (\mu-1)/2$.
  • The rescaled growth rate distribution is a symmetric Lévy stable distribution of index $\mu$.
  • The empirical value $\beta \approx 0.18$ corresponds to $\mu \approx 1.36$.
  • The model predicts a power-law distribution for sub-sector sizes ($s^{-1-\mu}$ for $s \ll S$).

Limitations and Future Work

  • Gaussian Sector Growth Assumption: Both models, in their simplest forms, assume Gaussian growth rates at the sub-sector level. The authors acknowledge this is likely unrealistic and influences the detailed shape of the growth rate distributions for finite $S$.
  • Finite Size Effects: The asymptotic results for the alternative model are valid when $S^{(\mu-1)/\mu} \gg 1$. For finite $S$ and $\mu$ close to 1, strong finite-size effects and slow convergence are expected, necessitating numerical simulations.
  • Empirical Verification: The authors emphasize the need for further empirical research to test the falsifiable predictions of the two models, particularly regarding the distribution of sub-sector sizes and the quantitative description of rescaled growth rates.
  • Dynamics of Sector Growth: The paper suggests that a more natural dynamics for sector growth might lead to power-law distributions, contrasting with Sutton's "microcanonical" assumption.

Summary updated

BibTeX
@misc{wyart2003statistical,
  author = {Matthieu Wyart and Jean‐Philippe Bouchaud},
  title = {{Statistical Models for Company Growth}},
  year = {2003},
  doi = {10.2139/ssrn.391860},
  howpublished = {SSRN},
}
Figure