Agentic Research Wins Where Markets Don't Exist

Why our AI couldn't beat Polymarket — and what that proves about where it does win

1. Thesis

We spent real engineering effort trying to use an AI agent swarm to make money on Polymarket. It didn't work, and we now understand why with enough precision to state it as a finding: a general-purpose AI reading public information has no durable, after-cost edge betting against a liquid prediction market — and the choice of model (Gemma, Haiku, Opus, anyone's) does not change that. It is the same reason index funds beat stock-pickers: the price already contains what we know.

The more useful half of the finding is the inverse. The thing that makes MSR's agentic research valuable — deep, structured, continuously-collected intelligence on local and institutional topics — is valuable precisely because no liquid market prices it. Our edge and efficient markets are orthogonal by construction.

2. Evidence

- Outcome. 21 closed trades, 28.6% win rate, -$9.44 realized. Order fill rate 2.3% (819 of 840 attempts failed). The most recent trades were 100% losses. Every profit/canary readiness gate stayed blocked the entire time.

- The markets we could find were the markets we had no edge on. A live scan of 5,000 markets surfaced top books on the 2026 World Cup, MicroStrategy's Bitcoin holdings, the Peruvian and Colombian presidential elections, and Roland-Garros tennis. None intersect what MSR researches.

- Backtest comfort vs. live reality. Backtested consensus accuracy was 70-75%, but live lost money — and the sub-evals were weak where it matters (evidence_citation 43%, edge_estimate 67%). Calibration is not edge when the market is already calibrated.

3. Framework: the Orthogonality of Agentic Research and Prediction Markets

Three properties decide whether any researcher can profit from a prediction market:

1. A market must exist and be liquid — enough volume to take a position without moving the price.

2. You must hold information the marginal trader doesn't — better, earlier, or more structured.

3. The edge must survive costs — spread, slippage, fees, and fill rate.

Mapped to MSR's footprint: Texas county / municipal / grants / procurement -> no market (deepest edge, no surface); AI-education / FERPA -> no market; Tech-Scout / AI ecosystem -> public, real-time, faster crowd (edge <= 0); US political discourse -> most-efficient category (edge 0); public health (CDC NWSS wastewater) -> rare, thin (the one candidate); geospatial -> no market.

The pattern: liquid markets form around mass attention; MSR's edge lives in institutional, local, slow-moving domains that attract no betting market. The intersection is nearly empty, and the one non-empty cell collapses under scrutiny.

4. Case study: killing the one real candidate

Public-health markets were the only place MSR ingests a genuine leading indicator the betting crowd doesn't systematically watch — CDC NWSS wastewater data leads case counts by 1-2 weeks. We tested it directly (Gamma API, 2026-06-02): health markets are ~1% of the active book, and the only one listed was "Hantavirus pandemic in 2026?" It fails twice: hantavirus is not tracked by NWSS wastewater surveillance, and "pandemic" resolves on a WHO/CDC declaration — a bureaucratic event — not a data threshold our feed could front-run.

5. Implications

1. Don't buy "AI alpha" on liquid markets. If a vendor claims an LLM beats prediction markets or public equities at scale, the burden of proof is enormous, and our own honest attempt returned the expected negative.

2. Point agentic research where markets are absent — local-government and grant-outcome forecasting, where structured intelligence is scarce and no efficient price competes with it.

3. A negative result, published honestly, is an asset — more credible than another hype claim, and it argues directly for what MSR actually sells.

6. Limitations

Single operator, small capital, ~10 weeks. We did not exhaustively test illiquid long-tail markets (unsizable by design) or ultra-low-latency news trading (a game retail loses structurally). The bounded claim: no durable, scalable, after-cost edge for a public-information AI against liquid consensus.