Pareto-Optimal and Strategyproof Data-Sharing Between Competitors

Morgan Colin-Mourre, Felipe Garrido-Lucero, Patrick LoiseauarXiv 2026 · arXiv

TL;DR

This paper introduces a mathematical framework for "coopetitive" data-sharing, proving that competitors can mutually increase profits by sharing training data without shifting their relative market power. The authors propose a specific "equal-gain" contract that is both Pareto-optimal and strategyproof, ensuring that firms have no incentive to lie about their data holdings.

What Problem Does This Paper Address?

In the age of deep learning, data is the ultimate competitive advantage. However, this creates a "data-sharing dilemma": while multiple firms could build better models by pooling their datasets, they are often direct competitors in the same market. If Firm A shares data with Firm B, Firm B’s model improves, potentially allowing them to steal Firm A’s customers.

Most existing literature on data-sharing ignores this competitive friction or relies on complex privacy-preserving architectures. This paper tackles the economic core of the problem: How can we design a data-sharing agreement that incentivizes rational, self-interested competitors to collaborate?

Methods

The authors model a market where firms compete on model quality (precision). A firm's profit is decomposed into two parts:

  1. Precision: The probability of closing a sale once a customer visits (depends on the firm's own data).
  2. Market Share: The probability a customer chooses to visit that firm (depends on the relative precision of all firms in the market).

The core contribution is a parametric family of equal-gain data-sharing contracts. These contracts are designed so that the relative market shares of all participating firms remain constant. By "freezing" the competitive landscape, the negative effect of a rival getting better is perfectly offset by the firm's own improvement.

The researchers specifically highlight a contract, $H$, which represents the limit of this family where at least one firm receives the maximum possible amount of data. They prove that if the precision function is concave (exhibiting diminishing returns), contract $H$ is strategyproof—meaning a firm cannot gain an advantage by under-reporting the size of its initial dataset to a central coordinator.

Datasets

The authors validated their theoretical model using the Expedia hotel recommendations dataset (from a 2016 Kaggle competition).

  • Scale: They used a subset of 500,000 random observations.
  • Task: Predicting which hotel cluster a user will book based on ~30 features (search location, device type, dates, etc.).
  • Model: Random Forest classifiers were used to empirically derive the "precision function" $P(n)$ by measuring top-$k$ accuracy across different training set sizes.

Key Insights and Ideas

  • Neutralizing the Competitive Threat: The "aha moment" is that you don't need to hide data to protect market share; you just need to calibrate the amount shared so that everyone's "attractiveness" grows proportionally.
  • Harder Tasks, Bigger Gains: Data-sharing is most beneficial when the machine learning task is difficult (e.g., low top-$k$ accuracy). In these regimes, the marginal utility of additional data is highest.
  • Homogeneity Breeds Cooperation: The gains from sharing are significantly higher when firms start with similar amounts of data. When one firm is a "data giant" and others are "dwarfs," the giant has less to gain because it is already in the region of diminishing returns.
  • Robustness to Non-Concavity: While the strategyproofness proof requires a concave precision function, empirical tests showed that even when real-world models aren't perfectly concave, the potential profit gain from lying is negligible (often less than 0.04%).

Key Results

  • Significant Profit Boosts: In the Expedia simulations, firms saw profit increases ranging from 8% (in highly heterogeneous duopolies) to 68% (in harder prediction tasks with many similar firms).
  • Pareto-Optimality: The proposed contract $H$ is shown to be Pareto-optimal, meaning no firm can be made better off without making another firm worse off.
  • Relative Share Stability: The math ensures that if Firm A had twice the market share of Firm B before sharing, it will still have exactly twice the share after sharing, even though both firms are now more profitable.

Limitations and Future Work

  • Data Volume vs. Content: The model assumes all data points are i.i.d. (independent and identically distributed). In reality, firms might have "niche" data that is more valuable than raw volume. Future work could explore heterogeneous data content.
  • Over-Declaration: The current strategyproofness proof covers under-reporting. While a "central pool" implementation (where you must physically provide the data you declare) prevents over-reporting, other decentralized implementations might be vulnerable.
  • Computational Cost: Determining the exact precision function $P(n)$ requires training many models at different scales, which is computationally expensive for very large-scale applications.
  • Static Market: The study looks at a snapshot of the market. It doesn't account for how data-sharing today might change a firm's ability to collect more data in the future.

Summary updated

Abstract

We study data-sharing among firms competing in the same market through the quality of their predictive models. We model firms' profit through two main components, precision and market share, where the former depends on the firm's own number of datapoints and the latter depends on the endowments of all firms. We characterize Pareto-optimal contracts as those in which at least one firm receives all of the other firms' datapoints. We then propose a parametric family of data-sharing contracts and show that, under any of them, the relative market shares of all firms remain constant, eliminating the negative effect of sharing data with competitors and equalizing all firms' gains relative to the no-sharing scenario. We further show that one contract within this family satisfies both Pareto-optimality and strategyproofness against firms attempting to under-declare the size of their initial endowment. Finally, using a real-world dataset, we empirically test the gain in profit for all firms under the contract described above, and show that data-sharing is particularly beneficial when the machine learning task is harder or when firms' initial endowments are closer to homogeneous.
BibTeX
@misc{colinmourre2026pareto,
  author = {Morgan Colin-Mourre and Felipe Garrido-Lucero and Patrick Loiseau},
  title = {{Pareto-Optimal and Strategyproof Data-Sharing Between Competitors}},
  year = {2026},
  eprint = {2611.26108},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.26108},
}
Figure