arbnx · prediction markets

Narratives
and models,
with the evidence attached.

arbnx reads what the news is claiming and what prediction markets are pricing. Every model we trade with publishes how it was tested — on what sample, on what date, and with what verdict, the failures included.

Get startedSee the models

Evidence ledger

as recorded
  1. 2026-09-10Germany · graded on elections it never saw60 held back · typical error 1.815 ptsholds
  2. 2026-09-26Germany · vs the zweitstimme.org benchmarkMAE 1.28 vs 1.67 · 2 electionsinconclusive
  3. 2026-10-04Israel · probabilities vs polls at face valueseat brackets 0.561 vs 0.880 Brierholds
  4. 2026-10-04US · 2022 midterms vs Polymarket's pricesBrier 0.082 vs 0.098 · one electioninconclusive
  5. 2026-08-24Football · model vs the market priceBrier 0.2174 vs 0.2017 · 3,677 gamesfailed
  6. 2026-10-04Contradictions · US midterm winner vs margin159 of 787 pairs · $222 fillablemeasuring

Each line links to the model page that carries the full finding — its method, its sample and the ones that did not hold.

Meet the arbnx desk.

30 seconds · Works without sound
Markets, model evidence, and bot controls in one desk.
Read the walkthrough
  1. Explore arbitrage, football, politics, and stocks from the arbnx desk.
  2. Compare equivalent Kalshi and Polymarket outcomes. Illustrative prices: 55¢ Yes plus 38¢ No costs 93¢ against a $1 paired payout, a 7¢ gross gap before fees and slippage. Execution and settlement carry risk.
  3. Inspect each model’s “How we know” record: method, sample and date, metric, result, and verdict. The video shows an illustrative inconclusive record, not measured performance.
  4. Configure one model per bot, a budget, and a threshold. Preview paper decisions and inspect why an opportunity passed. Live trading requires explicit approval, eligible released models, and passing safety gates.
  5. Markets. Models. Decisions. Explore the desk. All displayed data is illustrative; no real orders are placed.

01 · Claims · Radar by arbnx

What the news is claiming — and who disagrees.

Radar reads a reviewed set of sources — news feeds, Hacker News, GDELT, Reddit, and a set of Bluesky and X accounts that grows by itself — and groups what they say into stories and narratives. A narrative is a claim two or more outlets make; its headline is one of those claims, word for word, never a summary written for it.

buildingOpen Radar
“No counter-position found” — never “undisputed”.

Radar sees only what its sources say. Repetition is not confirmation, and the page says so in words.

01Stories
Events two or more outlets report, with copies not counted twice.
02Narratives
Claims two or more outlets make, quoted verbatim, with who disputes them.
03Markets
Open prediction markets on a narrative's subject — same subject, not the same question.
04Gaps
Where a narrative with attention is contested, unconfirmed, one-sided or one-language.

Meet Radar · 30 seconds · Part of arbnx

See the story. Question the narrative.

Sources, stories, claims, and the gaps between them. Works without sound; all displayed headlines and data are illustrative.
Read the Radar walkthrough
  1. Read headlines with their original sources and languages. Search the feed without losing provenance.
  2. Group reports into stories. Count original outlets separately from syndicated copies. Lexical grouping can split reports across languages.
  3. Follow a claim group’s member claims and inspect source quotes, counter-positions, and review status. Repetition is not confirmation; no counter-position found does not mean undisputed.
  4. Inspect contested coverage as a model’s reading, not a verified fact. The film’s example remains unreviewed.
  5. Explore publication links and modelled links to markets on the same subject. Same subject is not the same question. No market is a price on the claim.

02 · Models

Forecasts that have to earn a trade.

A model bets only where its probability is above the market price — and only after it has been graded on elections or games it never saw. Most of what we have built is on paper. That is the point of saying so.

Germany, the day before the vote: average miss per party, in points
arbnx de-71.28
zweitstimme.org1.67

Berlin and Mecklenburg-Vorpommern, 14 party results, forecasts of 2026-09-19 · lower is better · two elections, so: inconclusive.

Edges that need no forecast

Cross-venue arbitrageKalshiPolymarket

The same question priced differently on Kalshi and Polymarket: buy both sides for less than the $1 they pay. Hedged, rules checked before a spread is shown.

live, inside per-trade and daily caps
Contradictions

One venue disagreeing with itself — a winner market and its margin brackets that cost less together than they must pay. Real, persistent, and small.

scanner measuring · bot paper

The route to a live order

  1. 1Pre-registerThe test and its pass mark are committed before any score is computed.
  2. 2AuditHeld-out elections, the poll average and the market as baselines. Failures are published.
  3. 3ShadowA challenger records its forecasts beside the released version, on paper.
  4. 4ReleaseA person promotes it in the registry, with a reason — an override is recorded as one.
  5. 5Live, cappedOnly released versions, only on live portfolios, inside per-trade and daily limits.

The platform underneath

Portfolios
Capital per book, what each bot has claimed, what is still free.
Wallets
Kalshi and Polymarket keys, encrypted; balances read from the venues.
Live guards
Per-trade and daily spend caps, idempotent orders, an alert on a lone leg.
Bots
A step-by-step setup per model, paper by default, version pinning.
Pay per fill
One credit per executed trade. Data and the feeds are free.