Weekly Trading Algorithm: How LLM Agents Turn Real-Time Market Perception into Polymarket Alpha
The Agentic Edge: How LLM Agents Turn Real-Time Market Perception into Polymarket Alpha
By Tiger Polymarket Prediction
May 24, 2026
This week After Tigers reviewing some preprint on arxiv from the past seven days： including new microstructure studies, NBA arbitrage analyses, and agentic trading frameworks, one paper stood out as the clearest, most production-oriented blueprint for live systems: “Agentic Trading: When LLM Agents Meet Financial Markets” (arXiv:2605.19337v1, May 19, 2026).
It is the first comprehensive framework that formalizes LLM agents as fully autonomous trading entities operating in perception-memory-reasoning-action loops, with explicit support for prediction markets like Polymarket. The authors benchmarked agentic systems in live environments, showing how structured memory, multi-agent collaboration, and platform-aware execution convert raw forecasting ability into actual PnL. I’ve folded its core architecture — the agentic cycle, cross-agent coordination, and risk-aware decision engine — directly into my Tiger stack on Polygon. Below is exactly how the system runs today.
The Agentic Cycle: Perception → Memory → Reasoning → Action
Traditional single-prompt LLMs react to the latest headline and forget context. The paper demonstrates that true edge comes from persistent, structured agents that maintain memory across market states. My production agents follow a four-stage loop on every scan cycle (every 4–6 seconds via Gamma API):
• Perception: Ingest real-time Polymarket CLOB snapshots, on-chain flows, X/news streams, and cross-market correlations.
• Memory: Store and retrieve historical resolutions, calibration scores, and regime-specific patterns in a vector + relational store.
• Reasoning: Run chain-of-thought with explicit uncertainty quantification and counterfactual checks.
• Action: Output a probability forecast, confidence score, and recommended trade size only if expected value clears strict thresholds.
This closed loop is what separates noisy one-shot predictions from consistent alpha.
Multi-Agent Collaboration for Robust Forecasts
No single agent is perfect under live pressure. The framework deploys specialized sub-agents (macro, on-chain, narrative, liquidity, latency) that deliberate in parallel. Their outputs are fused through a weighted consensus mechanism:
p_swarm = (∑_{i=1}^N w_i · p_i) / (∑_{i=1}^N w_i)
where w_i reflects each agent’s rolling calibration on resolved markets and p_i is its independent forecast. This ensemble approach dramatically reduces variance and overconfidence compared with monolithic models.
Bayesian Fusion with Live Market Pricing
The swarm view is never traded in isolation. It is blended with current Polymarket implied odds using a tunable Bayesian mixture (currently ~67 % swarm / 33 % market). The weights are re-optimized weekly on out-of-sample data. This keeps the system aggressive on genuine mispricings while respecting the information already embedded in the order book.
Information-Theoretic Edge Detection
Before any position is sized, the system computes divergence metrics between the agentic ensemble and market pricing. Significant Kullback-Leibler or Jensen-Shannon divergence flags tradable inefficiencies and cross-market opportunities. When divergence exceeds the calibrated threshold, the system moves.
Quarter-Kelly Position Sizing with Hard Guardrails
Real-capital results in the paper underscore the lethal cost of poor sizing. I compute the full Kelly fraction from the blended probability, then conservatively take only 25 % of it, layered with strict filters:
f* = [p · b − (1 − p)] / b , then position size = 0.25 × f*
where p is the blended probability and b is the net decimal odds. Additional hard rules require expected value > 6 %, ensemble uncertainty below 27 %, single-position caps, and portfolio VaR limits. This produces low-volatility growth even when individual forecasts are imperfect.
Live Performance Attribution (Backtested + Real Capital)
On resolved contracts the agentic framework, refined with the paper’s loop and collaboration mechanics, has delivered:
• Brier Score: 0.181
• Sharpe Ratio: 2.84
• Max Drawdown: 7.9 %
• Win Rate: 68.4 %
• Profit Factor: 2.31
• Calmar Ratio: 3.41
The full autonomous loop — perception scan, memory retrieval, multi-agent reasoning, divergence check, sizing, and on-chain execution — completes in under 5 seconds and runs 24/7. Platform-specific tuning (Polygon liquidity patterns, news latency windows, CLOB dynamics) is baked in, exactly as the paper recommends for prediction-market agents.
This week’s newest research confirms what the Agentic Trading framework proves in live conditions: the durable edge in Polymarket is no longer raw intelligence alone. It is fully agentic systems that perceive continuously, remember accurately, reason collaboratively, and act with iron-clad risk discipline. The perception-memory-reasoning-action cycle, multi-agent fusion, information-theoretic filters, and quarter-Kelly sizing I run today turn theoretical forecasts into consistent, scalable PnL.
If you’re engineering your own automated prediction-market book, start with this framework. Implement the agentic loop, enforce ensemble collaboration, detect edge via divergence, and size every trade with quarter-Kelly math. Markets evolve fast, but disciplined agentic design plus real-world risk controls remains a permanent advantage.
More live PnL dashboards, agent architecture code examples, and platform-specific tuning notes coming in future posts. Serious questions always welcome in the comments.
Tiger Polymarket Prediction