Researcher ORCID Identifier

https://orcid.org/0009-0006-9952-7516

Graduation Year

2026

Date of Submission

4-2026

Document Type

Open Access Senior Thesis

Degree Name

Bachelor of Arts

Department

Economics

Reader 1

Fan Yu, Ph.D.

Terms of Use & License Information

Creative Commons Attribution-Noncommercial 4.0 License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License

Rights Information

© 2026 Saif Aldeen A.K. Agha

Abstract

This thesis documents the design, deployment, and forward-test evaluation of an evolutionary multi-agent algorithmic trading system on Polymarket, the largest decentralized prediction market. The system pairs a locally-hosted 72-billion-parameter language model with a gradient-boosted statistical filter and an evolutionary selection mechanism that maintains a population of approximately 500 autonomous trading agents. Each agent generates a probability estimate for an event, compares it to the prevailing market price, and trades the resulting disagreement.

The central empirical exercise estimates a panel regression of trade-level profit on the absolute disagreement between the agent's probability estimate and the market price, controlling for agent identity, calendar day, and trade-level characteristics. On 113,346 resolved trades placed by 3,157 unique agents over a 34-day window, a one-unit increase in absolute disagreement is associated with approximately $245 of additional per-trade profit (t = 24.5). A 1,000-iteration permutation test in which disagreement values are randomly reassigned within each agent's history fails to recover the observed coefficient in any iteration, placing the estimate 22.3 standard deviations outside the placebo distribution. Performance persists across each agent's trading history: the rank correlation of mean profit across the first and second halves of an agent's trades is 0.66, and 81 percent of top-decile agents in the first half remain in the top decile in the second.

A parallel live deployment using USDC (a dollar-pegged stablecoin) across 65 on-chain wallets (1,791 trades, 19 days) produces a 7.2 percent realized loss on initial deposits, while the mark-to-market value of open positions approximately offsets that loss at the experimental cutoff. The realized loss is concentrated in a single trade-direction code path that placed aggressive positions when the language model had not produced an actual probability estimate; following the documented engineering fix on April 20, 2026, the system produces an 82.5 percent win rate and positive realized profit on the subsequent sample. The decomposition isolates signal generation from execution: when the language model's signal reaches the order book as designed, the relationship documented in the paper sample carries through to live capital.

The asset class itself differs structurally from conventional arbitrage strategies. Fixed-income arbitrage is characterized by Duarte, Longstaff, and Yu (2007) as earning small, steady returns under normal conditions while suffering large, correlated losses when systematic factors move against the strategy. Prediction-market contracts resolve to idiosyncratic events, so a diversified portfolio across categories has no analogous systematic exposure. Residual risk is concentrated in the platform layer and the execution layer rather than in the underlying contracts, and that relocation of risk is itself a substantive feature of the new asset class.

Download

Share

COinS