Research scanner
Find the gap before risking money.
Scans markets, reads resolution rules, builds a probability with evidence and ranks the gaps against the book. No orders, no exposure.
Market scanResolution rulesEvidenceRanking
Prediction markets / AI systems
AI for prediction markets,
built as a system.
We build the harness around the model: data, calibration, edge detection, execution and risk. For Polymarket and other event markets.
01Data
02Model
03Edge
04Execution
05Evaluation
01 / The failure
The model is not the strategy.
The demo works. The live account does not. The missing layer is the harness.
01
A forecast is not an edge
The market already prices the obvious. A model that agrees with the crowd has no trade, no matter how good the prose is.
02
A signal is not a trade
Fees, spread, liquidity and size decide whether the edge survives contact with the order book.
03
A backtest is not a system
Live feeds break, orders miss, calibration drifts. Someone has to watch the thing at 3 a.m.
Most teams start by plugging an LLM into an API key. That is the first week, not the product. The market does not care about your demo.
02 / The harness
Five layers the model does not write for you.
The model is a commodity. The system around it is the advantage.
01 / Data
Market feeds, order books, resolution rules, external data.
Snapshots you can replay, not a stream you hope to catch.
02 / Model
A probability, not a paragraph.
LLM, statistical or ensemble. One calibration layer turns confidence into a number.
03 / Edge
Model probability minus price, fees and fill cost.
Ranked opportunities, not a watchlist of everything.
04 / Execution
Sizing, limits, order types, cancellation, kill switch.
Paper before live. Human approval where the size justifies it.
05 / Evaluation
Backtests, live paper, traces, P&L attribution.
Calibration drift is a first-class metric, not an afterthought.
More on the thesis: the harness is the moat
03 / The work
Three ways to start.
You do not have to begin with live money. Most teams should not.
Execution agent
The model proposes. The harness disposes.
Places orders inside explicit limits: size, price, liquidity, daily loss. Paper trading first, human approval and a kill switch.
SizingLimitsPaper tradingKill switch
Risk and monitoring
See the failure before the loss.
Exposure, calibration drift, failed orders and fee drag in one place. A short daily brief instead of another dashboard nobody opens.
ExposureCalibration driftFee dragDaily brief
04 / From production
What we learned building one.
A prediction market system we built and run. No client names, no P&L promises, just the parts that mattered.
Weather, priced against an ensemble
The edge was not the prompt. It was the data pipeline: an open weather ensemble, a calibrated mapping to the market question, and a filter for markets where the book was thin. The model was the easy part.
Lesson: the data pipeline carried the edge.
An LLM forecaster that was too confident
Raw model probabilities were overconfident at the tails. A calibration layer and a market filter made the output usable. The lesson: an uncalibrated forecast is a liability, not an edge.
Lesson: calibrate before you size.
Fees and liquidity ate the first strategy
The naive signal looked profitable until fees, spread and size were included. A fee-aware sizer and a liquidity filter turned it into something that could survive live.
Lesson: the book is part of the model.
05 / Beyond Polymarket
Event markets first. Then the rest.
Prediction markets are a good starting point: clean questions, fast feedback, public data. The same five layers apply to trading, hedging and arbitrage on other venues.
Trading
Signals, sizing and execution on liquid markets. The harness decides what reaches the book.
Hedging
You cannot hedge what you cannot measure. Limits, scenarios and a clear audit trail.
Arbitrage
Two prices, one position. Latency, fees and fill probability are the whole game.
06 / How we start
Diagnostic, prototype, production.
No live money on day one. The first deliverable is a decision, not a trade.
01
Diagnostic
1 to 2 weeksWe audit the data, the model and the execution path. You get a short report: what is missing, what to build first, what not to build.
02
Prototype
4 to 6 weeksA scanner or a paper-trading agent on your data. No live money until the evaluation is clear and the failure modes are named.
03
Production
OngoingLive system, monitoring, calibration and a retainer. We stick around for the failure modes, not just the launch.
07 / From the blog
The research behind the system.
The Five Things Polymarket's Winners Do and 87% Don't
What 112,000 wallets say about sizing, category focus and exits.
The Weather Bots Quietly Making Millions on Polymarket
Why the boring markets attracted the best data pipelines.
How an LLM Turned $1,000 into $14,000 on Polymarket in 48 Hours
The viral demo, and what it leaves out about execution and risk.
AI Agents and Prediction Markets
How forecasting agents change research, and where they still fail.
They Didn't Hack Polymarket. They Hijacked Its Speed Test
The trust boundary behind a market is the browser, not the contract.
Start with a diagnostic
Bring us the market. We will build the harness.
Tell us what you are trying to trade and what you have already tried. We will answer with a point of view, not a sales deck.
alex@vegaforge.dev ↗VegaForge / Prediction markets