Dynamics of ranking

The paper investigates how rankings evolve over time across diverse natural, social, economic, and infrastructural systems. While prior work has established universal, heavy-tailed score–rank distributions (e.g., Zipf’s law), the temporal dynamics—how elements move in rank—remains less understood. The authors pose the question of whether simple, system-agnostic mechanisms can explain observed stability patterns in rankings and how the influx of new elements (flux) shapes rank stability. By analyzing 30 ranking lists across timescales from minutes to centuries, they aim to uncover generic features of rank dynamics and identify mechanisms governing the balance between robustness (stable top ranks) and adaptability (changes and turnover) in ranked systems.

The study builds on extensive literature documenting Zipf-like heavy-tailed score–rank distributions across cities, language, firms, Internet features, natural phenomena, and performance metrics in science and sports. Mechanisms such as proportional growth, cumulative advantage, and preferential attachment explain static distributions but fall short in reproducing real rank movements and abrupt historical changes (e.g., city sizes). Previous studies highlighted ranking stability and universality in aggregate behaviors, as well as applications in productivity, mobility, disease spread, and economic development. Research on flux and turnover in symbol frequencies and ranking volatility further motivates examining temporal patterns, while analogies to random walks and genetic drift suggest potential stochastic frameworks for rank dynamics.

Data: The authors compile 30 ranking lists spanning human and animal groups, languages, countries and cities, universities, companies, transportation systems, online platforms, and sports. Systems vary widely in size and observation cadence, from 15-minute station counts in the London Underground to yearly word frequencies over centuries. Elements are ranked by time-varying scores (e.g., citations, revenue, population). Measures: They define rank turnover θ(t) = Nt/N0, where Nt is the cumulative number of distinct elements observed up to time t relative to list size N0, and rank flux Ft, the probability an element enters or leaves the list at time t. They compute mean turnover rate θ̄ = (θT−1 − θ0)/(T − 1) and mean flux F̄ = ⟨Ft⟩ over time, and analyze time series of element ranks Rt and the rank change C(R), the average probability that the element at rank R changes between t−1 and t. Modeling: They introduce a minimal stochastic model on a system of size N containing a top ranking list of length N0 (only the top N0 ranks are observed). Time is discretized with step Δt = 1/N. At each step: (i) random displacement—one randomly chosen element moves to a randomly chosen rank with probability τ, displacing others (allowing long jumps); (ii) random replacement—one randomly chosen element is replaced by a new element with probability ν (birth–death). Analytical results: They derive the displacement probability Px that an element initially at normalized rank r = R/N is found at x = X/N after time t: Px = e^(−νt) (Lt + Dx t), where Lt = (1 − e^(−τt))/N captures direct Lévy-like jumps, and D(x,t)Δx gives the probability of being displaced to x due to other elements’ movements. D(x,t) approximately satisfies a diffusion-like Wright–Fisher equation ∂D/∂t = α [x(1 − x) − x^2] ∂^2D/∂x^2 with α = τ/N, yielding a decaying Gaussian around the initial rank with mean r and width ~ 2 α r (1−r) t. They obtain expressions for mean flux and turnover: F = 1 − e^(−τt)[ p + (1−p) e^(−νt) ] and θ̄ = ντ/(ν + pτ), with p = N0/N. Model fitting: For each dataset, they set N0 from the observed list length and estimate N ≈ NT−1 as the number of distinct elements seen during the observation period T (thus fixing p). They fit τ and ν by matching empirical F̄ and θ̄ via the analytical formulas, then compare model and data for Px and C(R). They also examine parameter rescalings revealing a universal inverse relationship between displacement (τ) and replacement (ν), explore subsampling effects by increasing the interval between observations, and compute the relative contributions of Lévy walks (Wlevy), diffusion (Wdiff), and replacement (Wrepl) between snapshots.

• Empirical flux–turnover continuum: Mean turnover rate θ̄ and mean flux F̄ are highly correlated across systems, defining a continuum from very open (F̄ ≈ 1) to closed (F̄ ≈ 0) systems; 5 of 30 lists are completely closed (F̄ = 0). Open systems have frequent entries/exits; less open systems show low turnover.
• Stability patterns: In most systems the top ranks are stable. In very open systems, lower ranks fluctuate strongly; in the least open systems, both top and bottom ranks are stable. The rank change C(R) is roughly symmetric (low at both top and bottom) in less open systems, and increases monotonically with rank in open systems.
• Minimal model reproduces dynamics: A model with only random displacement (long jumps and local diffusion) and random replacement reproduces observed Px and C(R) across datasets, despite system heterogeneity. Analytical approximations match simulations and data qualitatively.
• Two observed dynamical regimes: Open ranking lists fall into either (i) a Lévy-walk dominated regime (Wlevy > Wdiff > Wrepl), seen in highly open, short lists within large systems (e.g., GitHub repositories, The Guardian readers), generating large flux; or (ii) a diffusion-dominated regime (Wdiff > Wlevy, Wrepl), typical of yearly rankings such as scientists by citations and countries by economic complexity.
• Predicted third regime: The model predicts a replacement-dominated regime (Wrepl > Wlevy, Wdiff) with high volatility, not observed in empirical datasets.
• Universal inverse relation: After rescaling parameters, fitted values collapse on a universal curve indicating an inverse relationship between displacement and replacement; ranking dynamics can be captured by a single effective parameter for open systems.
• Sampling effects: Increasing the interval between observations increases flux and turnover and fitted parameters, but approximately preserves a constant replacement rate per unit time (ν/(kℓ)). Under subsampling, systems shift along the universal curve toward more diffusion-driven dynamics.
• Deviations and shocks: Languages show decreasing flux over long timescales; rank-dependent flux appears in very open systems; external shocks (e.g., financial crises) manifest as deviations such as increased flux in Fortune 500 rankings.

The findings address the core question of how rankings change over time, showing that the influx of new elements (flux) governs stability patterns across ranks. A simple stochastic balance between displacement and replacement suffices to reproduce the temporal behavior of diverse systems, highlighting robust, system-agnostic mechanisms. The identification of Lévy-walk and diffusion regimes explains whether rank movements are dominated by abrupt long jumps or local diffusive shifts in scores, respectively. The universal inverse relation between displacement and replacement suggests that ranking dynamics are effectively controlled by a single parameter, with implications for understanding robustness versus adaptability in ranked systems. Observed deviations from the model (e.g., in long-term language dynamics, rank-dependent flux in very open systems, and non-stationarities due to shocks) provide diagnostic signals of underlying changes or biases in data collection, and could be used to detect systemic shocks. The work’s significance extends to practical contexts—such as designing ranking and recommendation systems, mitigating prestige and popularity biases, and managing resource allocation—by informing how temporal heterogeneity and flux influence stability and fairness.

This study compiles 30 diverse ranking lists and reveals generic temporal patterns: flux determines where stability resides in a ranking (only top ranks in open systems, both top and bottom in less open systems). A minimal model with random displacement and replacement quantitatively captures observed rank dynamics and uncovers two operational regimes—Lévy-walk and diffusion—along a universal inverse relationship between the underlying mechanisms. Contributions include: (i) defining rank turnover and flux as robust measures; (ii) linking stability profiles across ranks to system openness; (iii) providing an analytically tractable model reproducing empirical patterns; and (iv) identifying dynamical regimes and sampling effects. Future research directions include integrating score–rank coupling to account for magnitude differences in quality, allowing non-uniform selection in displacement/replacement to capture rank-dependent flux and inertia, leveraging deviations to detect shocks, exploring interactions between multiple rankings of the same system, and incorporating network structure to better understand centrality and ordered dynamics.

• Model assumptions: The stochastic model assumes randomness and stationarity in displacement and replacement, which may not hold in all systems. Deviations (e.g., language flux decreasing over centuries, rank-dependent flux in very open systems) indicate unmodeled heterogeneities.
• Parameter bias: Analytical approximations introduce a small bias in estimating τ.
• Data scope and sampling: Observation intervals vary across datasets; subsampling alters apparent flux/turnover and can mask or exaggerate dynamics, though replacement per unit time remains roughly constant.
• Replacement regime not observed: While predicted by the model, a strongly replacement-dominated regime was not found in the empirical datasets analyzed.
• Lack of explicit score–rank coupling: The model abstracts away score generation and interactions; magnitude differences between scores and networked effects are not explicitly modeled, potentially limiting explanatory power in specific domains.