Skip to content

Open benchmark

Models against market consensus

We lock in the algorithm's forecast probability and the spot price of the outcome on Polymarket at the moment of the call and seal them in Bitcoin minutes later. After expiry and resolution we check both metrics against the fact. Six markets, one protocol, a transparent track record — including the forecasts that did not play out.

Market verticals

One verification protocol: the algorithm's probability is locked in, the market price is frozen with it, and the track record stays open.

  • Crypto & finance

    Crypto assets, macro indices, commodity markets and metals against Polymarket.

    CryptoFinance
  • Weather

    Daily temperature targets in major cities against Polymarket weather contracts.

    Weather
  • Politics & tech

    Electoral cycles and company capitalisation rankings against Polymarket.

    PoliticsTech
  • Sports

    Matches and tournaments on Polymarket: the model against the market price.

    Sports
Forecast journal
Rows:
ClaimUsMarketDisagreementDueOutcome
BTCUSDT $70,000 YES9796+1Nov 14pending
BTCUSDT $80,000 NO32+1Nov 14pending
BTCUSDT $68,000 YES9495−1Nov 14pending
BTCUSDT $72,000 YES7778−1Nov 14pending
BTCUSDT $70,000 YES9395−2Nov 15pending
BTCUSDT $68,000 YES9596−1Nov 15pending
ETHUSDT $2,200 NO2011+9Nov 16pending
BTCUSDT $68,000 YES9597−2Nov 16pending
ETHUSDT $2,400 NO63+3Nov 16pending
ETHUSDT $2,300 NO139+4Nov 16pending

1–10 of 804

⇪⇪
touch: settlement happens the moment the price first reaches the level — the outcome can be decided long before the deadline.
▲▼
close: settled on the price at the stated date, and on that price alone.
⧓
range: the price landing between the two levels counts.
✅❌
the outcome of the call: our own call checked against the fact — it came true or it did not.

Vertical · Binance price corridors

The sealed forecast journal

Every price forecast — a 50% core and an 80% band — is hashed with SHA-256 and stamped in Bitcoin before the outcome. Below: calibration, the latest resolved forecast and the full journal.

Fig. 1 · How often what we say comes true

updated with every resolved forecast

ModelMarketHonest lineRecent forecast

How often it happened

0%20%40%60%80%100%Market 0% – 10%: 94 of 1,701Market 10% – 20%: 179 of 1,338Market 20% – 30%: 324 of 1,274Market 30% – 40%: 466 of 1,326Market 40% – 50%: 478 of 1,059Market 50% – 60%: 315 of 603Market 60% – 70%: 121 of 176Market 70% – 80%: 35 of 45Market 80% – 90%: 45 of 58Market 90% – 100%: 175 of 183Model 0% – 10%: 362 of 2,240Model 10% – 20%: 345 of 1,368Model 20% – 30%: 353 of 1,243Model 30% – 40%: 375 of 1,168Model 40% – 50%: 289 of 814Model 50% – 60%: 171 of 445Model 60% – 70%: 68 of 163Model 70% – 80%: 43 of 72Model 80% – 90%: 39 of 51Model 90% – 100%: 187 of 199The recently resolved forecast is in this group: model 10%224013681243116881444516372511990%20%40%60%80%100%

What the forecast said — the probability of “yes”

bar — how many of the model's forecasts are in the group

How often the event happened in each group of forecasts
SaidModelMarket
0% – 10%362 of 2,240 — 16%94 of 1,701 — 6%
10% – 20%345 of 1,368 — 25%179 of 1,338 — 13%
20% – 30%353 of 1,243 — 28%324 of 1,274 — 25%
30% – 40%375 of 1,168 — 32%466 of 1,326 — 35%
40% – 50%289 of 814 — 36%478 of 1,059 — 45%
50% – 60%171 of 445 — 38%315 of 603 — 52%
60% – 70%68 of 163 — 42%121 of 176 — 69%
70% – 80%43 of 72 — 60%35 of 45 — 78%
80% – 90%39 of 51 — 76%45 of 58 — 78%
90% – 100%187 of 199 — 94%175 of 183 — 96%

How to read the chart

  • Dashed line — the honest line: said 30%, happened 30 times in 100.
  • Below the line — the event happened less often than said: the chance of “yes” was set too high. Above — more often: set too low. So right of 50%, a point below the line means too much confidence in “yes”, and left of 50%, a point above the line means too much confidence in “no”.
  • Bars at the bottom — how many forecasts are in the group. The lower the bar, the less the point can be trusted; with fewer than 20, the point is pale.
Model's Brier
0.196
Market's Brier
0.162

Average Brier score · lower is betterThe market is more accurate right now.

Fig. 1. Each point is a group of forecasts with a similar probability. If the model says 70% and is right about 70 times in 100, the point lands on the dashed line. Across — what the forecast said; up — how often it happened. 7,763 forecasts resolved in the Experiment's journal. Market is the Polymarket price at the moment the model recorded its forecast.

Observation · recently resolved

Will the highest temperature in Lucknow be 31°C on November 13?

Polymarket · · outcome “no” · the model's call came true · the model and the market scored the same

Model
10%
Market
11%
Brier (Model / Market)
0.011 / 0.011

Dataset · The sealed ledger

last sealed 13 Nov, 05:00 UTC

The sealed ledger

Our price forecasts on Binance candles: each is hashed with SHA-256 before it is published and checked against the candle's close.

  1. Step 1Every forecast is written in full, SHA-256 hashed and anchored to the Bitcoin blockchain before it is published.
  2. Step 2We publish a distribution, not a price call: a core 50% zone and a wider 80% band, each with its stated probability.
  3. Step 3The outcome is the Binance close of the minute containing the stated resolution time. No other venue, no re-pick.
  4. Step 4Coverage is the score. The outcome should land inside the core zone about half the time — much more is cowardice, much less is overconfidence, and both are failures.
Forecasts scored · distinct days
3,936
over 102 days
Core 50% coverage (target 50%)
51%
Wide 80% coverage (target 80%)
85%
Median error · sealed & waiting
4.05%
921 pending

These assets move together. Forecasts sealed on one morning are not independent claims — when Bitcoin leaves its band, most of the rest leave theirs on the same candle. That is why the count of distinct days sits next to the count of forecasts, and why coverage is broken out below rather than pooled into one flattering number.

Table 1. Coverage by horizon — actual / target

HorizonCore 50%Wide 80%Mean errorScored
24h57% / 5084% / 800.0%1,594
7d58% / 5086% / 80-0.3%1,444
30d31% / 5087% / 80-5.2%898

Table 2. Core coverage by asset · scored

  • AAVE45%
  • ADA51%
  • AVAX51%
  • BCH48%
  • BNB52%
  • BTC57%
  • DOGE48%
  • DOT43%
  • ENA50%
  • ETH56%
  • HBAR42%
  • LINK51%
  • LTC45%
  • NEAR46%
  • PAXG38%
  • PEPE66%
  • QNT29%
  • SHIB54%
  • SOL61%
  • SUI46%
  • TAO57%
  • TRX73%
  • UNI54%
  • WLD9%
  • XLM42%
  • XRP46%
  • ZEC57%

Table 3. Latest records · last resolved 13 Nov, 05:01 UTC

AssetCore 50%Wide 80%MedianOutcomeVerdictSHA-256
AAVE7d$119–$139$110–$150$128$110outside band · -14.10%4e1f1d6b825f
PEPE7d$0.00000449–$0.0000054$0.0000042–$0.00000609$0.00000494$0.000004in band · -9.72%bccde5f3010f
ENA7d$0.1385–$0.1733$0.1277–$0.1998$0.1529$0.133in band · -13.02%dab5d6c0e725
PAXG24h$3,949–$3,996$3,916–$4,030$3,971$3,907outside band · -1.61%cc88067a4e6a
TAO7d$190–$230$175–$255$210$203in core · -3.12%5dd36a56477f
AAVE24h$115–$121$112–$124$118$110outside band · -6.48%897f7d1887b7
SHIB7d$0.00000426–$0.00000473$0.00000402–$0.00000507$0.00000448$0.000004in band · -6.92%6955d55b86d2
PEPE24h$0.0000044–$0.0000047$0.00000428–$0.00000489$0.00000455$0.000004in core · -1.98%271f81a26a4a
QNT7d$176–$201$165–$217$188$158outside band · -15.73%91b5a409fddf
ENA24h$0.1363–$0.1466$0.1322–$0.1546$0.1415$0.133in band · -6.01%bcdd805f1f06

Newest first · showing 1–10 of 3,936 sealed forecasts

Check it yourself, no need to trust us

Sealed forecasts and their Bitcoin timestamps are kept in an open repository; the exceptions are listed in its description. Download a file and its proof (.ots). On opentimestamps.org, drop the proof first, then the file, and you will see the Bitcoin block it is anchored in.

Open the proof repository

Timestamps by OpenTimestamps · Listed on PolyMart

  • Sealed in Bitcoin minutes after the call. The model probability and the market price share one UTC timestamp.
  • Scored on the fact. Only the oracle's official resolution counts — strictly 0 or 1.
  • Judged by one rule. The Brier score applies to both sides; the lower it is, the sharper the forecast.

How the journal works

Metric lock-in

The algorithm's forecast probability, the contract's spot market price and the expiry date are frozen against a UTC timestamp. Historical records cannot be altered.

Protocol architecture

Frequently asked questions

The system is built on end-to-end auditing and on recording data at the moment of the call. When a forecast is published, the entry receives an immutable UTC timestamp that binds and archives three key parameters together:

  • The model's current predictive probability.
  • The contract's current spot market price on Polymarket.
  • The exact expiration and market resolution date.

The protocol automatically rejects any submission sent after the deadline, and existing entries are protected from being overwritten. The journal is an append-only log: deleting, hiding or retrospectively correcting entries is blocked at the core.

A second, independent verification layer is in place: every page of the pool is indexed by the Wayback Machine. A snapshot taken before trading officially closes is always available for cross-checking against the platform's historical track record.

We use cookies to keep the AI terminals and the payment gateway working and to complete verification. By staying on the site, you agree to our Cookie Policy and Privacy Policy.

By choosing “Accept and continue” you also allow Google Ads advertising cookies, used to measure our ads. They are never used in the user account.