Data-Driven Knowledge-Aware Inference of Private Information in Continuous Double Auctions

Lvye Cui, Haoran YuAAAI 2024 · DOI

Introduction

In strategic market environments, "private information"—such as a seller's production cost or a buyer's maximum willingness to pay—is the cornerstone of economic analysis. Knowing these hidden values allows market designers to quantify efficiency and develop better policies. However, in Continuous Double Auctions (CDAs), inferring this information is notoriously difficult. CDAs are fast-paced, dynamic markets where multiple participants submit bids and asks simultaneously, making traditional economic equilibrium models mathematically intractable.

The Challenge of Inference

Traditional "Inverse Game Theory" fails in CDAs because it assumes participants play perfectly according to an equilibrium that cannot be easily calculated. Conversely, pure machine learning (ML) approaches struggle because private costs are almost never recorded in real-world data, precluding standard supervised learning. To address this, researchers have developed KATE (Knowledge-Aware Cost Inference), a framework that combines neural networks with population-level statistical knowledge to "peek under the hood" of these complex auctions.

The mechanics of a Continuous Double Auction, showing the flow of bids, asks, and trades over time.

The KATE Framework

KATE moves away from rigid economic theory without falling into the trap of "black-box" ML. It employs a Gated Recurrent Unit (GRU) to model seller behavior, mapping private costs and market states to predicted "ask" prices. The framework is guided by three key pillars:

  1. Behavior Modeling: Predicting the next move based on current market dynamics.
  2. Statistical Knowledge: The authors observed that while individual costs are hidden, the ask-cost ratio (the markup) across a population follows a predictable shifted exponential distribution. KATE uses this to anchor its inferences.
  3. Economic Constraints: A specialized loss function enforces economic logic, such as the monotonicity of marginal costs, ensuring the model doesn't produce "hallucinations" that violate basic market principles.

Experimental Results

The study utilized a massive experimental dataset from MobLab, encompassing over 9,000 markets and 300,000 data points. Unlike real-world exchanges, these experiments provided "ground truth" private costs, allowing for rigorous verification.

Comparison of Mean Squared Error (MSE) convergence and testing performance across KATE and various baseline models.

As shown in the performance metrics, KATE achieved a significant leap in accuracy. With a Mean Squared Error (MSE) of ~888, it vastly outperformed the best "knowledge-unaware" ML baselines. The integration of domain knowledge not only lowered the error rate but also allowed the model to converge much faster during the training process.

Conclusion

KATE demonstrates that a perfect equilibrium model isn't necessary for accurate inference. By bridging the gap between statistical "hints" about a population and the flexible power of neural networks, researchers can effectively reconstruct hidden economic variables in even the most chaotic trading environments. Future work aims to extend this seller-centric model to buyer behavior and more complex multi-asset markets.

Summary updated

BibTeX
@inproceedings{cui2024data,
  author = {Lvye Cui and Haoran Yu},
  title = {{Data-Driven Knowledge-Aware Inference of Private Information in Continuous Double Auctions}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i9.28864},
}

Loading related content...

Figure