How AI Finds Edge in Prediction Markets
The exact model pipeline Polykit uses to surface mispriced bets — from vision to live news to fair-value estimation.
What 'Edge' Actually Means
In prediction markets, edge is the gap between your estimate of true probability and the market's price. If you think an outcome is 70% likely and it's trading at 55¢, you have 15 points of edge. Over enough trades, edge is the only thing that separates profitable traders from everyone else. Everything else — hot takes, conviction, gut feel — is noise.
Finding edge is hard because liquid markets are, most of the time, approximately right. The question isn't 'what do I think will happen' — it's 'what is the market getting wrong that I can measure?' AI is extremely well-suited to this question because it can process news, historical base rates, and order-book context at a speed no human can match.
Why Markets Misprice
Three structural biases create most of the edge in prediction markets. First, recency bias — traders overreact to the latest headline and underweight slower-moving fundamentals. Second, lazy consensus — many markets drift toward round numbers like 50¢, 75¢, or 25¢ because that's where casual money anchors. Third, thin liquidity — small markets get pushed around by a handful of traders, leaving prices that don't reflect real probability.
Layer on event-specific inefficiencies: resolution ambiguity, unfamiliar jurisdictions, and markets that combine several variables (a 'both A and B' contract, for example) almost always trade slightly wrong because humans are bad at combining probabilities.
Polykit's Pipeline: Vision First
The Analyzer starts with a screenshot. You paste or drop an image of any Polymarket or Kalshi market and our vision model — a fine-tuned GPT-4o variant — extracts the contract title, current YES/NO price, 24-hour volume, and resolution date. This takes under two seconds and works with cropped screenshots, mobile screenshots, and partially obscured images.
We chose vision-first intentionally. Scraping APIs would restrict us to the platforms we'd integrated, whereas a screenshot pipeline means you can analyze any market on any prediction venue — including new ones we haven't seen yet. This is why Kalshi support worked day one.
Live Context from Perplexity Sonar
Once the market is parsed, we query Perplexity's Sonar API for live news relevant to the contract title. Sonar returns up-to-the-minute web results — news articles, official statements, polling updates — with source citations. This is the step that gives Polykit its informational edge over a vanilla LLM: we're not relying on training-data knowledge that might be months stale.
The returned context is filtered, deduplicated, and ranked by relevance and recency. A market about 'Fed rate cut in March' gets FOMC minutes and fresh inflation prints; a market about 'AI model to hit benchmark X' gets the latest model announcements. Garbage-in, garbage-out applies — so we invested heavily in this retrieval layer.
Fair-Value Estimation with GPT-4o
The parsed market plus curated news context feeds into GPT-4o running a structured pricing prompt. The model is instructed to reason through base rates, direct evidence, counter-evidence, and common biases, then output a point estimate of true probability plus a confidence range. It also produces a written thesis — the 'why' behind the number — so you can sanity-check its logic instead of trusting a black box.
We benchmarked this pipeline against closing prices on 2,000+ resolved markets. The AI's fair-value estimates beat the entry price on a majority of flagged edges, with the biggest wins coming in medium-liquidity political and sports markets where news moves faster than the crowd updates.
An Example: The 15¢ Gap
During the 2024 election cycle, a down-ballot Senate race was trading at 42¢ YES. The Analyzer read the screenshot, pulled live polling from three pollsters plus a fresh fundraising disclosure, and returned a fair value of 57¢ with a clear thesis: the market hadn't repriced after a Tuesday fundraising filing that revealed a 3-to-1 cash advantage. That's 15¢ of edge on a $500 position — $75 of expected value, excluding variance, if the model is right.
The contract resolved YES. But more importantly, the Analyzer flagged the mispricing within minutes of the filing hitting the FEC database, while the market took another 36 hours to drift to 55¢. Speed plus structured reasoning is the product.
Sizing from Edge
A recommendation without a position size is half a product. Polykit outputs a suggested stake based on a fractional-Kelly calculation: the bigger your edge and the smaller the variance, the larger the suggested allocation, capped at a user-defined percentage of bankroll. Most users run at one-quarter Kelly, which trades slightly slower growth for dramatically lower drawdowns.
The result is a complete loop: screenshot, read, research, estimate, recommend, size. Everything a human analyst would do, compressed into about 30 seconds, on every market you care about. That's the edge.