Prediction Markets in Decentralized Finance: An Agent-Based Model with Heterogeneous Traders

While prediction markets in decentralized finance aggregate dispersed information into probabilistic prices, their accuracy strongly dependents on design, incentives, and environment. Under suitable incentives, markets can theoretically approximate the likelihood of events (Wolfers & Zitzewitz, 2004). However, recent agent-based research suggests that in prediction markets with order-based matching, biased traders with large capital endowments (so-called “whales”) can move prices away from fundamentals. Such distortions may persist when updating beliefs is slow and herding reinforces such deviations (Smart et al., 2026). Such distortion dynamics thus depend not only on trader heterogeneity but also on the mechanisms translating demand into prices and spreads.

We analyze this challenge at the micro-level by simulating an automated market maker building on the principles of logarithmic market scoring rule (LMSR). LMSR derives prices from a convex cost function over outstanding shares and secures continuous liquidity with bounded worst-case loss (Hanson, 2003; Chen & Pennock, 2007). Unlike price-impact systems in which demand moves prices without relevant lag (Smart et al., 2026), LMSR embeds liquidity in the curvature of the cost function. This structural difference alters how capital accumulation and biased trading affect price formation. Our contribution lies not just in the replication of an order-book market but rather in an analysis of distortion mechanisms under scoring-rule-based liquidity provision.

Our model approximates a binary event market with heterogeneous traders, including informed traders with high-precision signals, noise traders generating random order flows, arbitrage-oriented traders reacting to publicly available information, strategic traders with biased valuations and larger budgets, and conviction-driven traders who update beliefs slowly. A latent “true” probability evolves stochastically with occasional shocks. Agents update beliefs in log-odds space using heterogeneous learning weights and submit trades proportional to the gap between belief and LMSR price, scaled by both risk aversion and trading intensity.

Performance is evaluated dynamically and at settlement. During trading we trace mispricing (price - true probability), volatility, volume, and wealth distribution across traders. At settlement, forecasting accuracy is measured using both Brier score and log loss. The experimental design focuses on three questions: (i) how the LMSR liquidity parameter affects the trade-off between responsiveness and stability; (ii) how large and persistent distortions become under high-budget biased trading; and (iii) how liquidity design interacts with stubborn belief updating.

In our preliminary analyses, we test whether distortion mechanisms identified in price-impact order-matching environments (Smart et al., 2026) also emerge under scoring-rule-based automated market makers and to which extent liquidity curvature mitigates persistence. This work’s limitations emerge primarily from generalized trading strategies and abstracted order-book dynamics. It is further limited by the exogeneity of true probabilities.

References

• Chen, Y., & Pennock, D.M. (2007). A Utility Framework for Bounded-Loss Market Makers. Conference on Uncertainty in Artificial Intelligence. 49-56.
• Hanson, R. (2003). Combinatorial Information Market Design. Information Systems Frontiers, 5(1), 107-119.
• Smart, J., et al. (2026). Manipulation in Prediction Markets: An Agent-Based Modeling Experiment. ArXiv.
• Wolfers, J., & Zitzewitz, E. (2004). Prediction Markets. Journal of Economic Perspectives, 18(2), 107–126.