Swarm AI for Prediction Markets: Collective Intelligence Gets an Algorithm

Swarm AI applies the logic of collective intelligence—the idea that many independent reasoners outperform any single one—to algorithmic forecasting. MiroFish, an open-source engine that reached GitHub’s global trending list on March 7, 2026, operationalizes this by spinning up thousands of AI agents that argue, evolve opinions, and generate emergent predictions. The approach has demonstrated real promise, but its accuracy claims remain largely unvalidated.¹

What Is Swarm AI for Prediction?

The core premise draws on decades of research into collective intelligence. Individual forecasters, whether human or model, carry systematic biases. Aggregate enough independent judgments and those biases tend to cancel. Prediction markets exploit this through financial incentives; ensemble ML models exploit it through statistical averaging.

Swarm AI introduces a third model: agents that actively interact, influence each other, and produce emergent outcomes rather than simply averaging independent outputs.

The distinction matters. In a traditional ensemble, 12 models each produce a probability estimate and the outputs are averaged or weighted. In a swarm simulation, agents debate, shift positions based on social pressure, and generate dynamics that mirror how real human populations process information. The prediction emerges from the system’s behavior, not from a post-hoc aggregation step.

Unanimous AI, the longest-running commercial swarm intelligence company, operationalized this with human participants moderated by AI. Over a full NBA season, their swarm system reached 72% prediction accuracy against Vegas favorites, which hit 66.5% accuracy across the same games—a 5.5 percentage point edge that translated to a 57% ROI over the season.² In a separate Oxford University study, individual human forecasters averaged 55% accuracy; organized into AI-moderated swarms, those same people hit 72%—a 131% amplification in predictive performance.³

MiroFish attempts to achieve similar amplification using AI agents in place of human participants.

How MiroFish Works

MiroFish, developed by Guo Hangjiang—a senior undergraduate student in China—is described in its own documentation as “a simple and universal swarm intelligence engine, predicting anything.” The project attracted 30 million yuan in investment from Shanda Group founder Chen Tianqiao within days of going viral.⁴

The system runs a five-step pipeline:

1. Knowledge graph construction: Input documents—news reports, policy drafts, financial signals, even works of fiction—are parsed using GraphRAG, which extracts entities and relationships into a connected graph rather than treating text as a flat document.
2. Agent persona generation: The knowledge graph seeds thousands of AI agents with unique personalities, initial stances, long-term memory, and behavioral logic.
3. Dual-platform simulation: Agents interact across two simultaneous environments modeled on Twitter-like and Reddit-like social structures. The underlying simulation engine, OASIS (Open Agent Social Interaction Simulations) from CAMEL-AI, supports up to one million concurrent agents executing 23 distinct social actions—posting, commenting, reposting, following.
4. Emergent pattern synthesis: A dedicated ReportAgent analyzes opinion shifts, coalition formation, and behavioral clusters as they emerge from agent interactions.
5. Scenario testing: Users can inject new variables mid-simulation, query individual agents, and explore alternate outcomes.⁵

The technical stack is Python 3.11+ with a Vue.js frontend. Agent memory runs through Zep Cloud. Any OpenAI SDK-compatible model works as the underlying LLM; the project recommends Qwen-plus for cost efficiency.

Swarm vs. Ensemble: What the Research Actually Shows

The forecasting accuracy literature offers a more nuanced picture than MiroFish’s positioning suggests.

A 2024 study published in PMC tested an ensemble of 12 LLMs—including GPT-4, Claude 2, Llama-2-70B, and Mistral-7B—against human crowd predictions on 31 binary forecasting questions drawn from Metaculus. The LLM ensemble produced a mean Brier score of 0.20, compared to 0.19 for the human crowd. The difference was statistically indistinguishable (p = 0.850).⁶

That’s a significant result: a simple ensemble of off-the-shelf LLMs already matches human crowd intelligence in aggregate forecasting accuracy, without requiring any agent simulation at all.

Approach	Mechanism	Demonstrated Accuracy	Validation Status
Traditional ensemble (ML)	Averaged independent model outputs	Baseline	Well-established
Human swarm (Unanimous AI)	AI-moderated real-time human consensus	72% NBA / 131% uplift	Multi-year production data
LLM ensemble	Averaged outputs across 12 models	Statistically ≈ human crowd (Brier 0.20 vs 0.19)	Academic study, 31 questions
Agent-based simulation (MiroFish)	Emergent dynamics from interacting AI agents	No published benchmarks	Early-stage; no outcome validation

The comparison table reveals the honest state of the field: the methods with the most compelling accuracy data (Unanimous AI’s human swarms) involve real human judgment augmented by AI, not fully synthetic agent populations. MiroFish’s fully automated approach is architecturally closer to the LLM ensemble than to Unanimous AI’s validated methodology—but without the benchmarking that would let practitioners assess its predictive value.

AI Agents Are Already Trading Prediction Markets

While MiroFish focuses on simulation and scenario generation, autonomous AI agents have already entered live prediction markets as active participants.

Polystrat, an AI trading agent deployed on Polymarket in February 2026, executed more than 4,200 trades within its first month of operation, achieving returns as high as 376% on individual positions, according to CoinDesk reporting on March 15, 2026.⁷ The agent trades on behalf of self-custody users and operates continuously.

The Olas network hosts similar autonomous agents operating on prediction market platforms, where 24/7 machine-speed position-taking creates dynamics that human-only markets never experience. This is the live edge of what swarm-oriented forecasting looks like in deployment—not simulated agents in a Python environment, but algorithms placing real bets with real capital based on continuously updated market signals.

The gap between simulation (MiroFish’s domain) and live trading (Polystrat’s domain) is significant. A system that generates plausible emergent narratives from thousands of agents and a system that executes profitable trades in real time require fundamentally different validation standards.

Where Agent-Based Simulation Actually Adds Value

The limitations of MiroFish’s current iteration don’t invalidate the underlying approach. Agent-based simulation has a well-established track record in specific domains:

Policy impact modeling: Running thousands of synthetic stakeholders through a policy change and observing coalition dynamics, friction points, and unintended consequences is genuinely useful analytical work—even without ground-truth accuracy validation.

Narrative stress-testing: Organizations can seed MiroFish with a draft announcement and watch how simulated agent populations react across different social structures. The output is qualitative intelligence, not probability estimates.

Creative and speculative reasoning: The system’s ability to accept fiction as seed material—and simulate how agent populations would interpret or extend a narrative—opens non-forecasting use cases in game design, scenario planning, and strategic foresight.

The herd behavior problem is real and worth flagging here: LLM agents in social simulations tend toward faster consensus and more extreme polarization than real human populations exhibit. Agent populations built on similar training data share systematic biases, which can make simulated societies behave more uniformly than actual ones. This isn’t a fatal flaw, but it’s a calibration challenge the field hasn’t solved.⁸

The Broader Market Signal

The swarm AI and autonomous agent infrastructure market reached $2.01 billion in 2025 and is projected to expand to $5.98 billion by 2030 at a 24.3% CAGR, according to GlobalNewsWire market data.⁹ MiroFish’s viral trajectory—from open-source release to 18,000+ GitHub stars to institutional investment in under three months—reflects genuine practitioner interest in collective AI reasoning frameworks.

Google Research’s 2025 work on scaling agent systems offers a useful counterpoint: adding more agents doesn’t linearly improve performance, and systems can degrade if agent architecture isn’t matched to task structure. Sequential dependencies, tool density, and coordination overhead all create ceilings that raw agent count doesn’t overcome.¹⁰

The same principle applies to prediction quality. MiroFish can spawn a million agents. Whether those agents produce reliable forecasts depends on validation methodology—and that work remains ahead.

---

Frequently Asked Questions

Q: Does MiroFish actually outperform traditional prediction methods? A: No validated evidence supports this claim as of March 2026. MiroFish has not published benchmarks comparing its simulation outputs to historical outcomes. Its own documentation positions it as a scenario exploration tool, not a precision forecasting system.

Q: How does swarm AI differ from ensemble machine learning? A: Ensemble ML averages independent model outputs after the fact. Swarm AI—whether using humans or agents—simulates active interaction between participants, allowing emergent dynamics like opinion drift, coalition formation, and social pressure to influence the final prediction. MiroFish is an agent-based simulation system; ensemble methods are aggregation systems.

Q: What has Unanimous AI demonstrated about swarm prediction accuracy? A: Over a full NBA season, Unanimous AI’s human swarm system hit 72% accuracy against Vegas odds (66.5% baseline), producing a 57% ROI. Over four NFL seasons, it posted a 62.5% win-loss record against the spread. These results involve human participants moderated by AI, not fully synthetic agent populations.

Q: Can I use MiroFish in a production forecasting system today? A: Not reliably. MiroFish v0.1.0 is a prototype that lacks accuracy validation, runs at high API cost for large simulations (the documentation recommends starting under 40 rounds), and exhibits known LLM herd behavior biases. It’s useful for exploratory scenario planning, not production prediction pipelines.

Q: Are AI agents already active in prediction markets? A: Yes. Autonomous agents like Polystrat launched on Polymarket in February 2026 and executed more than 4,200 trades in its first month, according to CoinDesk reporting from March 15, 2026. These are live trading agents operating on real markets, distinct from simulation systems like MiroFish.

---

Footnotes

1. GitHub. “666ghj/MiroFish: A Simple and Universal Swarm Intelligence Engine, Predicting Anything.” https://github.com/666ghj/MiroFish ↩

2. Unanimous AI. “Season-Long NBA Study Shows Swarm AI Beats Vegas, Produces 57% ROI.” https://unanimous.ai/nba-study-cicn-2019/ ↩

3. Unanimous AI. “How does Swarm work?” https://unanimous.ai/what-is-si/ ↩

4. TMTPOST. “AI Product, Developed by a Chinese Young Prodigy and Invested by Chen Tianqiao, Tops GitHub.” https://en.tmtpost.com/post/7905996 ↩

5. DEV Community. “MiroFish: The Open-Source AI Engine That Builds Digital Worlds to Predict the Future.” https://dev.to/arshtechpro/mirofish-the-open-source-ai-engine-that-builds-digital-worlds-to-predict-the-future-ki8 ↩

6. PMC / NIH. “Wisdom of the silicon crowd: LLM ensemble prediction capabilities rival human crowd accuracy.” https://pmc.ncbi.nlm.nih.gov/articles/PMC11800985/ ↩

7. CoinDesk. “AI agents are quietly rewriting prediction market trading.” March 15, 2026. https://www.coindesk.com/tech/2026/03/15/ai-agents-are-quietly-rewriting-prediction-market-trading ↩

8. BreezyScroll. “MiroFish AI: The Open-Source Platform Simulates Human Opinion and Markets Using Thousands of Agents.” https://www.breezyscroll.com/technology-news/mirofish-ai-the-open-source-platform-simulates-human-opinion-and-markets/ ↩

9. GlobeNewsWire. “Artificial Intelligence (AI) Swarm Control Station Research Report 2026.” January 29, 2026. https://www.globenewswire.com/news-release/2026/01/29/3228433/28124/en/Artificial-Intelligence-AI-Swarm-Control-Station-Research-Report-2026-5-98-Bn-Market-Opportunities-Trends-Competitive-Analysis-Strategies-and-Forecasts-2020-2025-2025-2030F-2035F.html ↩

10. Google Research. “Towards a science of scaling agent systems: When and why agent systems work.” https://research.google/blog/towards-a-science-of-scaling-agent-systems-when-and-why-agent-systems-work/ ↩