Accuracy

Inspiration

Prediction markets like Kalshi often misprice sports
outcomes relative to Vegas sportsbook consensus. Vegas
lines aggregate millions of dollars of informed money —
when Kalshi diverges, there's a systematic edge to capture.

## What it does
Accuracy is a fully automated paper-trading bot that runs against the Prophet Arena benchmark. Every 15 minutes it:

1. Pulls real-time Vegas moneylines and outright winner
   odds via The Odds API
   
2. Scans the Kalshi market universe for YES/NO contracts where the Kalshi price diverges from the Vegas implied
   probability by more than 3%
   
3. Sizes positions using fractional Kelly criterion (25%
   Kelly, capped at $400/trade)
   
4. Exits positions automatically when the edge compresses below 0.5%
   

Coverage spans NBA, MLB, NHL, EPL game markets plus World
Cup and EPL title outrights, over 1,200 tradeable markets per tick.

## How we built it
Pure Python bot using the ai-prophet-core SDK. Odds data comes from The Odds API with a SQLite cache to stay within the free-tier quota (~450 calls over 14 days). EPL title probabilities are derived from live standings via the ESPN public API. If the Odds API quota runs low, the bot falls back to LLM-estimated probabilities using ESPN schedules as grounding data. Deployed as a long-lived worker on Fly.io (Chicago region).

## Challenges The hardest part was market ID parsing, Kalshi tickers like KXNBAGAME-26MAY18CLEDET-CLE encode sport, date, team abbreviations, and the yes-side team in a single string. We built a full regex parser with abbreviation dictionaries for all four sports plus fuzzy matching for edge cases
(e.g. soccer draws encoded as TIE).

## What we learned
Vegas odds are an extremely clean signal, the main challenge isn't finding edge, it's managing position limits and API quotas across a 14-day window without burning
through your data budget on the first day.

Built With

• ai-prophet-core-sdk
• docker
• espn-public-api
• fly.io
• python
• sqlite
• the-odds-api