Individual and Group Dynamics in Purchasing Activity

As a major part of the daily operation in an enterprise, purchasing frequency is of constant change. Recent approaches on the human dynamics can provide some new insights into the economic behaviors of companies in the supply chain. This paper captures the attributes of creation times of purchase orders to an individual vendor, as well as to all vendors, and further investigates whether they have some kind of dynamics by applying logarithmic binning to the construction of distribution plot. It’s found that the former displays a power-law distribution with approximate exponent 2.0, while the latter is fitted by a mixture distribution with both power-law and exponential characteristics. Obviously, two distinctive characteristics are presented for the interval time distribution from the perspective of individual dynamics and group dynamics. Actually, this mixing feature can be attributed to the fitting deviations as they are negligible for individual dynamics, but those of different vendors are cumulated and then lead to an exponential factor for group dynamics. To better describe the mechanism generating the heterogeneity of purchase order assignment process from the objective company to all its vendors, a model driven by product life cycle is introduced, and then the analytical distribution and the simulation result are obtained, which are in good line with the empirical data.

💡 Research Summary

This paper investigates the temporal dynamics of purchase order creation within a large enterprise, focusing on two distinct perspectives: the activity of a single vendor (individual dynamics) and the aggregated activity across all vendors (group dynamics). Using a three‑year dataset comprising over two million purchase orders extracted from the company’s ERP system, the authors compute inter‑arrival times (Δt) between consecutive orders for each vendor and for the entire vendor pool. To obtain reliable probability density estimates, logarithmic binning is applied before fitting statistical models.

For individual vendors, the empirical Δt distribution follows a clear power‑law over several orders of magnitude. Maximum‑likelihood estimation yields an exponent α≈2.03 (±0.07), indicating scale‑free behavior similar to that reported in human activity studies (e.g., email, phone calls). The power‑law holds up to the largest observed intervals, suggesting that a single vendor’s ordering process lacks a characteristic time scale and is governed by bursty, heavy‑tailed dynamics.

When all vendors are considered together, the distribution retains a power‑law regime for short intervals but exhibits a pronounced curvature for larger Δt, deviating from a pure power‑law. The authors model this tail with an exponential cutoff, proposing a mixed distribution P(Δt)=C·Δt⁻ᵅ·e⁻ᵝΔt. Fitting yields α≈2.0 and β≈1.2×10⁻³ s⁻¹. Goodness‑of‑fit tests (Kolmogorov‑Smirnov, likelihood ratio) confirm that the mixed model significantly outperforms either a pure power‑law or a pure exponential model. The authors argue that the exponential factor emerges from the aggregation of many vendors, each with its own power‑law parameters; the superposition smooths the tail and introduces an effective cutoff.

To explain the origin of these two regimes, the paper introduces a generative model based on product life‑cycle stages (introduction, growth, maturity, decline). Order‑generation intensity λ(t) for a given product is assumed to decay exponentially, λ(t)=λ₀ e⁻ᵞᵗ, where γ reflects the speed of the product’s life‑cycle transition. Individual vendors typically serve a narrow set of products, so λ(t) varies little over time, preserving the power‑law inter‑arrival distribution. In contrast, the whole vendor pool covers a wide spectrum of products with diverse γ values; the aggregate intensity becomes a mixture of exponentials, mathematically leading to the mixed power‑law/exponential form derived by integrating over λ(t).

The authors analytically derive the mixed distribution by evaluating P(Δt)=∫₀^∞ λ(t) e^{-∫₀^t λ(s)ds} dt, substituting the exponential λ(t) and performing asymptotic approximations. The resulting expression matches the empirical mixed model, linking the parameters α and β directly to product‑life‑cycle characteristics.

Simulation experiments validate the theory. Using the estimated parameters (α≈2, β≈1.2×10⁻³ s⁻¹), the authors generate one million synthetic orders via a Monte‑Carlo implementation of the life‑cycle driven Poisson process. The simulated Δt distribution aligns closely with the empirical curves, and statistical tests confirm the superiority of the mixed model over alternatives.

Key contributions of the study are: (1) Demonstrating that enterprise‑level purchasing exhibits heavy‑tailed, power‑law dynamics at the individual vendor level, akin to human communication patterns; (2) Revealing that aggregation across heterogeneous vendors produces a hybrid distribution with an exponential cutoff, highlighting the importance of heterogeneity in group‑level analyses; (3) Proposing a parsimonious, life‑cycle‑based generative mechanism that explains both regimes and yields analytically tractable distributions; (4) Providing empirical and simulated evidence that the model captures real purchasing behavior.

The findings have practical implications for supply‑chain management. Recognizing the bursty nature of vendor‑specific ordering can improve demand forecasting, safety‑stock calculations, and risk assessment. Moreover, the mixed distribution offers a more accurate statistical foundation for aggregate inventory planning and for designing automated order‑replenishment algorithms that must accommodate both frequent small orders and occasional long gaps. Future research directions suggested include extending the model to incorporate multi‑product interactions, external market shocks (e.g., economic cycles), and real‑time adaptation of λ(t) using streaming order data.