Build Cross-Platform Prediction Market Arbitrage Tools

Getting the prices is the easy part

Anyone can pull three endpoints. The work that actually takes the time is deciding whether two markets are the same question — because until you know that, two prices are just two numbers.

Every platform disagrees with the others on nearly everything structural:

And that mapping is not write-once. It drifts every time a platform renames a market, splits one question into several, or retires a contract mid-cycle.

What we do before you see the data

What each response carries

Only fields that are consistently present. Note what is not here — no order book, no bid/ask, no liquidity figures. That shapes what you can honestly build, and the caveats section below is direct about it.

One event, three platforms, one object

This is the shape that makes a scanner simple. Polymarket at +112, Kalshi at −102, Novig at +108 — the same outcome on the same question, already lined up. Comparing them is a field lookup, not a matching exercise.

curl "https://api.predictionmarketapi.com/v2/events\
  ?apiKey=YOUR_API_KEY\
  &bookmakerID=polymarket,kalshi,novig\
  &oddsAvailable=true\
  &limit=10"

{
  "success": true,
  "data": [
    {
      "eventID": "US_PRESIDENTIAL_ELECTION_2028",
      "leagueID": "US_POLITICS",
      "status": { "started": false, "ended": false },
      "odds": {
        "probability-home-game-ml-yes": {
          "oddID": "probability-home-game-ml-yes",
          "fairOdds": "+105",
          "bookOdds": "+108",
          "byBookmaker": {
            "polymarket": { "odds": "+112", "available": true },
            "kalshi":     { "odds": "-102", "available": true },
            "novig":      { "odds": "+108", "available": true }
          }
        }
      }
    }
  ]
}

fairOdds is the no-vig reference across platforms, which is usually a better baseline to measure divergence against than any single venue. The schema reference lists every field.

What people build with it

A price difference is not an arbitrage opportunity

This is the section most pages like this leave out, and it is the one that will save you the most time. A gap in the data is a candidate. Whether it is tradeable depends on things the API does not know:

Matched on the question, not the wording

Title matching falls over immediately. “Will the Democratic nominee win in 2028?” and “2028 Presidential Election Winner — Democratic” are the same bet described two ways, while two markets with nearly identical titles can settle on different terms.

So matching works on the underlying question and returns one eventIDacross platforms. What that gives you is a reliable answer to “are these about the same thing?” — which is the hard part, and the part you would otherwise maintain forever.

What it is not is a guarantee that settlement terms are identical. Two venues can ask the same question and resolve it differently at the edges — an early call, a disputed result, a different cutoff. Before you trade a spread rather than just display it, read both sets of rules.

How fast the data moves

Update frequency matters more here than in almost any other use case. Gaps close as traders spot them, so a five-minute refresh will surface divergence that has already gone.

Poll near-term markets often and distant ones rarely — best practices covers how to spend a quota sensibly, and WebSocket streaming replaces polling entirely on custom plans.

Three integrations, or one

If you only ever watch one platform, go direct — the Polymarket API guide shows how. Arbitrage is by definition a two-platform problem, which is where this stops being true.

Which plan a scanner needs

Plans are metered in entities, and one entity is one event — however many outcome markets sit under it. So an election carrying twelve candidate markets counts once, not twelve times, which makes scanners cheaper to run than people expect.

Rough sizing: events watched × polls per day × days per month. Watching 200 events every five minutes is roughly 1.7 million entity reads a month, which is Alpha territory. Watching 200 events once an hour is around 144,000 — comfortably Edge.

Start on the free Signal plan to prove the logic, then buy update frequency when you need it. Full plan limits are on the pricing page.

Where to start

The quickstart gets a first call working in a few minutes. Setup and authentication covers keys, the schema reference covers the response, and pagination covers pulling a lot of markets at once.

Questions

Start building your scanner

Polymarket, Kalshi and Novig through one normalized integration, matched before you receive them. Spend the time on your strategy instead of your data pipeline.

Need WebSocket push for higher-frequency alerting? Talk to us about a custom plan.