KALSHEEP
A true prediction market where money is the meme.
Mission: make paid prediction markets useless by proving, in public and with a pre-registered method, that a free crowd with nothing at stake forecasts as well as or better than Kalshi and Polymarket. Everything on this site either produces that proof, keeps the proof honest, or keeps people here long enough to produce it.
No real money. Ever. Wool (W$) has no value.
Prior art, so we do not pretend otherwise: Manifold, Metaculus, Fatebook. None of them run a live, methodologically clean scoreboard against the paid markets. That scoreboard is the product. The rest is a forecasting site with jokes.
Mission metrics
| Offset | n | W / L / T | Rival Brier | Crowd Brier | Diff |
|---|---|---|---|---|---|
| n=0. Nothing eligible has resolved yet. | |||||
How scoring works
Constants: NEW_ACCOUNT_DAYS = 7, MIN_RESOLVED_FOR_BOARD = 10, MIN_CROWD_N = 20, SHRINK_K = 10. Raw mean Brier is shown on profiles for the curious but never ranks anything.
| Metric | Definition | Notes |
|---|---|---|
| Brier | (p - outcome)^2, outcome in {0,1} | 0 perfect, 0.25 coin flip, 1 confidently wrong. |
| Crowd probability at time t | Median of the latest forecast per user at t, excluding accounts younger than 7 days at the time of the forecast | Median resists trolls. This is the number compared against rivals. |
| User window Brier | Time-weighted average of Brier from question open to close. Before your first forecast, and during any gap, you are scored as if you held the crowd probability | Late joiners cannot beat early forecasters on an obvious outcome. Being right early is the only way to gain. |
| Crowd window Brier | Same time-weighted average using the crowd probability | The baseline every user is measured against. |
| Peer score per question | Crowd window Brier minus user window Brier | Positive means you beat the crowd. Easy questions give ~0. Collusion on easy questions gains nothing. |
| Oracle score | Shrunk mean peer score: sum(peer scores) / (n + 10), over resolved, non-voided questions you forecast on, excluding questions you created | Ten lucky calls do not outrank two hundred good ones. Board needs n >= 10. |
| Calibration | 10% buckets, predicted vs realised frequency, on your last forecast per question | Profile chart. |
| Degen board | Net winnings = balance minus lifetime top-ups plus lifetime burned fees; ROI = net winnings / lifetime staked | Bankruptcy does not reset this history. |
The Us vs Them protocol
Pre-registered, public before launch, reproduced verbatim.
| Item | Rule |
|---|---|
| Universe | Every Polymarket and Kalshi binary market whose volume at mirror time exceeds MIRROR_MIN_VOLUME (launch value 50,000 USD) and whose close is at least 48 hours after mirror time. Mirrored automatically by cron every 6 hours. No human picks. |
| Snapshots | Rival price and our crowd probability recorded together every hour, and additionally at fixed offsets before close: T-30d, T-7d, T-24h, T-1h. |
| Eligibility per question | Counts at an offset only if our crowd had at least MIN_CROWD_N distinct eligible forecasters at that offset. Questions voided by the rival, or where their resolution was overturned by our dispute process, are reported separately and excluded from the headline. |
| Scores | At each offset: Brier of rival price vs Brier of crowd probability. Over the window: time-averaged Brier of each from mirror time to close. |
| Headline | Wins, losses and ties (difference under 0.005) per offset and for the window average, with pooled mean Brier for each side and a bootstrap 95% confidence interval on the difference. Sample size printed next to every number. |
| Publication | Tally page updated nightly. Full CSV of every forecast (pseudonymous user id, question id, timestamp, probability) and every snapshot, nightly, public, versioned. The About page carries this table verbatim. |
| Commitment | Published whether we are winning or losing. Early numbers at n=30 will probably lose; the story "lost at n=30, won at n=3,000" is worth more than any single number, and hiding a bad tally is the exact behaviour we accuse them of. |
Data
Every forecast (pseudonymous user id, question id, timestamp, probability) and every snapshot. Never email or name. Cached one hour. Check our work.
Legal, briefly (not legal advice)
Gambling law regulates activity, not the word: consideration, chance, prize. Nothing is paid in, W$ cannot be bought, sold, transferred or cashed out, and there is no prize of value, which places the site with free-to-play social games. 18+ self-declaration at signup.
| Rule | Why |
|---|---|
| Never sell W$ in any form, including donations that grant W$ | Big Fish Casino (Washington) held purchasable chips to be a thing of value. |
| No prizes, merch, gift cards or crypto tied to W$ or rank | Rank converting to value triggers sweepstakes and lottery rules. |
| 18+ declaration and "No real money. Ever. Wool (W$) has no value." in footer and ToS | Advertising regulators care about gambling-themed content reaching minors. |
| No Kalshi or Polymarket logos or implied affiliation | Naming them in comparisons is fine. |
| Creator tier, if ever, sells slots only | Consideration buys publishing capacity, never a chance at a prize. |
Privacy
- Jurisdictions: GDPR (Netherlands) and DPDP Act 2023 (India).
- Delete = pseudonymise. Forecasts stay in the ledger as [deleted user] with a random id so scoring and the CSV dump keep integrity; email, name, avatar, follows, notifications and sessions are wiped. Delete your account.
- Public dump contains pseudonymous user id, question id, timestamp, probability, stake. Never email or name.
- Every forecast you make is public under your username, forever, including edit history. That is the point.
- Analytics: EU-hosted product analytics only. No third-party ad trackers.