Arena

Backtesting.Paper Trading.Evaluation.Routing.

Get the CLI
npm install -g arena-prediction-cli

Backtesting

Early access

In early access: write a rule and Arena will run it over markets that have settled, then run it forward so the record is real. No look-ahead, no cherry-picked windows, and a verdict that is allowed to be no.

fade NBA favourites when the line moves 3¢ or more in the last hour before tip-off

Illustrative sample, not a real backtest · 640 settled games

Game 1Game 640

Key moments

  • +5.8¢

    Games 24 to 54, 22 wins

    Rose from −10.7¢ to +5.8¢.

  • −4.9¢

    Games 55 to 155, 58 losses

    Fell from +5.8¢ to −4.9¢.

  • +1.8¢

    Games 156 to 640, 282 wins

    Rose from −4.9¢ to +1.8¢.

Settled games
640
56% won
Edge per contract
+1.8¢
mean, after fees
95% interval
−2.0¢ to +5.6¢
crosses zero
Verdict
Not an edge yet
run it forward
  1. 01

    Describe the rule

    A JSON spec, or your own probabilities for each game. Game-winner markets in seven leagues.

  2. 02

    Test it on the settled record

    Markets that have resolved, with the edge reported as an interval, not a win rate.

  3. 03

    Run it forward

    Trade the same picks in the sandbox at live prices, and the eval board grades each as it settles. It goes on the public board only if you ask.

Design

How a run reads

A saved backtest or eval becomes a workflow you can read left to right. It shows where the forecast came from, the model, the rule, what it did and how it scored. The workflow below is an example.

Source

Home win

Forecast

62%

Model

Your model

Stated probability

Rule

At least 55%

Take the side

55%

Action

Picks

Two markets

Allocate
$250.00
Take
+35%
Stop
-15%
  • KC ML
  • BOS ML
Result

Result

Recorded

P&L
+$14.20
Brier
0.22

The line

What moved it

One line per outcome through a game. Each play that moves the price is pinned on the line of the team it concerns, and the note under the chart names the play, when it happened and what that price did.

  • Home99%
  • Away1%
Illustrative sample, not a live game
Q1Q2Q3Q4
7:15 PMFinal 10:21 PM

Key moments

  • Home69%

    Touchdown, Home

    Q1 6:187:40 PM

    Home 7, Away 0.

    Rose from 55% to 69%.

  • Away37%

    Field goal, Away

    Q2 9:318:13 PM

    Home 7, Away 3.

    Rose from 30% to 37%.

  • Away58%

    Touchdown, Away

    Q3 4:079:24 PM

    Home 7, Away 10.

    Rose from 41% to 58%.

  • Home84%

    Touchdown, Home

    Q4 4:3610:08 PM

    Home 14, Away 10.

    Rose from 44% to 84%.

Historical data

Point-in-time

Arena keeps the history a backtest needs: the markets it lists, recorded as they trade, the facts those markets settle on, and the AI models’ graded picks. Everything is stored point-in-time: each row keeps the time it became known, taken from the source’s own dates for older history, so a backtest will only see what could have been known then. Backtesting is in early access, and none of this is offered as a download.

What Arena holds, one row per kind of data, as checked on September 29, 2026: whether it is recorded or still being added, and the date a finished stretch starts. A row that is only being added names the start of that work.
DataCoverageFrom
Markets
Market listingsRecorded from July 2021.Jul 2021
SettlementsRecorded from July 2021.Jul 2021
Price candlesBeing added from July 2021 across the whole window. No finished stretch yet.Jul 2021
Imported trades and fillsRecorded from November 2022 through March 4, 2026.Nov 2022
Earlier trade historyRecorded from September 17, 2026. Earlier days, back to June 30, 2021, are being added.Sep 2026
Daily exchange trade filesRecorded from August 2026, for every market on the exchange, not only Arena’s games.Aug 2026
Order books and trades on the games Arena listsRecorded from September 28, 2026.Sep 2026
Graded record
AI model picks, gradedRecorded from June 2026; stated probabilities scored from September 2026.Jun 2026
Economics
Weekly gas pricesRecorded from 1990.1990
US economic series, with revisionsRecorded from 2018, each revision kept.2018
Sports
International football resultsRecorded from 1872 to 1970; the years since are being added.1872
NFL scores and closing linesThe 1999 to 2002 seasons and this season recorded; the seasons in between are being added.1999
Club soccer resultsThis season recorded from August 2026, and some leagues’ 2010–11 season; the seasons in between are being added.Aug 2026
College football scores and linesThe 2013 season and this season recorded; the seasons in between are being added.2013
Weather
Weather at settlement stationsDaily reports, forecasts and observations recorded from September 2026.Sep 2026
Elections
2026 elections: polls, race ratings, early votePolls, race ratings and early vote recorded from September 2026; campaign finance filings from 2025.Sep 2026
As checked on September 29, 2026. Not to scale: the years before 2010 are squeezed and 2026 is stretched.

Planned, not held yet: baseball; Treasury yield, air-travel and weather archives; page-view counts; past election results and election-night counts.

Recorded around each game
Order books and trades on the games Arena lists, from six hours before the start until two hours after the final. Where an exchange’s book can only be polled, busy days are sampled.
Kept as it was known
Economic series keep every revision since 2018, and market rules keep each wording from September 2026 on, so a test will see what was published on the day, not the later correction.
Graded against the result
The AI models’ picks are graded against the real outcome, and since September 2026 a pick can state a probability, which is scored too. Your own paper trades held to the end are graded the same way, and stay private unless you make your profile public.

Agents

Early access

Tell an agent what you want. It builds the strategy, tests it against Arena’s markets, and runs it around the clock on infrastructure we host. What exists today is the CLI and the MCP server; hosting is what early access is for.

    1. 01

      Say what you want

      In plain English, to Claude Code, Cursor or any agent with the Arena CLI or MCP server.

    2. 02

      It builds and tests the strategy

      Against every settled market on record, and tells you straight whether the idea is an edge.

    3. 03

      It runs it, hosted

      Around the clock on infrastructure we host, on paper. The record goes on the public board only if you ask.

    Signals

    Noticeable orders · Unusual

    Large trades on Kalshi and Polymarket as they print, within seconds. Follow one on paper, or trade against it.

    Sample printsNot live data
    YesChiefs ML$62,40063¢
    YesBTC ≥ $65.5k$90,20057¢
    NoLakers ML$148,10041¢
    NoFed cut in Sept$57,50033¢
    YesYankees ML$71,80052¢
    NoETH ≥ $1.9k$81,90047¢