Building a Reproducible Simulation Stack for Population-Level Prediction Market Research

Abstract

This manuscript describes the engineering design behind a reusable simulation stack for studying population-level behavior in prediction markets. The current system combines a microstructure-oriented limit order book engine with a vectorized PyTorch batch-auction engine that scales to one million agents on local CPU hardware. The research goal is not only to produce one manuscript, but to create an instrument that can support repeated paper generation across market mechanisms, agent populations, and intervention designs. We describe the requirements, architecture, agent APIs, experiment orchestration pattern, reproducibility contract, and claim-audit workflow. We also document current gaps: the code is functional research software, but it still needs a package boundary, formal config schemas, stronger provenance manifests, systematic tests, and an interactive simulation layer before it can serve as a durable research platform.

Keywords

Read Full Text

The full paper is not yet available.