PREDICTION MARKETS · LIVE CAPITAL · PUBLIC RESULTS

I run real-money experiments in prediction markets,
and publish every result — including the failures.

Every line is preregistered before it runs: the pass/fail criteria are frozen in writing, timestamped, and the result is read exactly once at expiry. When a line dies, it gets a headstone — cause of death and tuition paid, in public. No equity curves here. Just what was tried, what it cost, and what it settled. 守型不改,満期再断 — hold the form, change nothing mid-window, judge only at the end.
静中有算
13
Experiments run
TOTAL
4
Under judgment
SEALED
9
Killed
KILLED

入門Three things firstSTART HERE

Written for someone who has never touched a prediction market. Answer these three and the rest of the site is readable.

Q 1

What is a prediction market?

You are not buying a stock. You are buying a claim. If it happens, each share pays $1. If it doesn't, it pays zero.

Which means the price is the probability. A market at 0.62 says the crowd thinks there's a 62% chance.

"Team X wins today" at 0.62
→ pay $62 for 100 shares
→ win: $100 · lose: $0
Q 2

What is Polymarket?

The largest prediction market today. On-chain settlement, USDC, and a central limit order book.

The key difference: the house does not take the other side. Your counterparty is another trader; the platform only charges the taker. That makes it an exchange, not a sportsbook — and it changes completely how money is made here.

taker fee = 0.07 × p × (1−p) × shares
makers pay zero
most expensive near 0.50
Q 3

Where does the money come from?

There are only four sources. Almost everything you'll read is one of these four, repackaged:

① Being right more often than the market (direction)
② Getting paid to quote (market making)
③ Structural mispricing (arbitrage)
④ Platform rebates (subsidy)

I have tried all four with real money.
Most of them I killed — the numbers are below.

四階From zero to running your own systemFOUR DEPTHS

The same subject read at four depths. Enter at whichever one you're standing on.

SEE

See it

How the market works, how fees are computed, who decides the outcome, how subsidies are paid. Skip this and everything after is guesswork.

FREE · FULLY PUBLIC
JUDGE

Judge it

Which plays are real and which are illusions. Ecosystem censuses, winner autopsies, and my own kill file.

FREE · FULLY PUBLIC
DO

Do it

Parameter ranges, market screening, cost calculators, pre-launch checklists. The layer you can actually act on.

TOOLS
BUILD

Build it

How to write a preregistration, how to validate your measuring stick, how to build a replay harness, where the data comes from.

METHOD · CODE · DATASETS

心法Seven principlesSEVEN PRINCIPLES

All seven were paid for. Each one names where it came from.

1

Freeze the ruler before you look

Pass/fail criteria are written and timestamped before any data is seen. After that, addenda only — never a revised verdict.

2

The ruler itself can break

The price that settles the market is not the price you're watching. On the 5-minute knife-edge window the two disagreed by 10.1% — and the error leaned my way. The "edge" I found was entirely an artifact of a broken label.

3

Losses can hide

A losing leg expires worthless and leaves no trace in the activity feed. Of 825 unpaired shares only 236 left a record; 589 evaporated silently. My first pass computed +$187. The true figure was +$23.

4

Rebates are a multiplier, not an addend

Both rebate programs maxed out only move the required gross edge from 1.7pp to 1.5pp. They cannot rescue a strategy that is already negative. Measured: 371 windows, −$496.51, rebates received $0.

5

Autopsy before you copy

July's star wallet was already dead by the time I took it apart in August. Leaderboards are lagging indicators — by the time you copy, the window has closed.

6

Capacity dies before edge does

The best wallet in the entire up/down market-making ecosystem earns roughly $95/day. Whether the method works is one question; how much money the lane can hold is a different one.

7

Claim your failure modes in advance

Before a line runs, write down how it will die. When it dies, claim one from the list — no inventing new explanations after the fact.

The Ledger
every line that ran — the losing ones stay on the page

Lines still under judgment show a count only, never a name. Their criteria are already frozen and hash-anchored on-chain; on judgment day the full preregistration is published alongside the result. That is the rule, not discretion.

BTC 5-minute · taker

net −$496.51
~496 windows · funded $1,138.02 · closing cash $641.51 · rebates $0
Preregistered single read returned FAIL; line closed the same day. The ledger's own window count was a stale mid-run figure — corrected here.

Weather · liquidity provision

≈ −$210
Hard stop on day zero → redesigned → stopped early → replay returned FAIL → archived.

Up/down market making 5m / 15m

KILLED
no capital committed
Median across 47 same-shape wallets: −0.74¢/share. Several stayed net negative while collecting $100–380 in rebates.

Complementary-set arbitrage

KILLED
no capital committed
Census of 694 participants: live opportunity inventory was zero, and 39% of the lane loses money.

Ladder arbitrage

KILLED
no capital committed
Falsified twice, independently, through two unrelated data paths.

TWAP tail lock

KILLED
no capital committed
0 of 4 tenors passed. The information was real; the cost assumption was not.

Under judgment

4 LINES
Criteria frozen before launch, hashes anchored on-chain. On judgment day the full preregistration is published with the result, pass or fail. Until then: no names, no markets, no tenors.
The Kill File
verified not to work — so you don't pay the tuition twice

This is the most valuable section on the site. Everyone else writes "three tricks to arbitrage." Here are nine things I tried and why they failed.

ExperimentVerdictBasisCost
BTC 5-minute takerFinal kill of its familyPreregistered read returned FAIL — the edge was absent, not eaten by fees~496 win · −$496.51
Weather liquidity provisionStopped early, archivedReplay returned FAIL; the star wallet being copied was itself already dead≈ −$210
Up/down market making 5m / 15mQuantitatively excludedMedian of 47 same-shape wallets: −0.74¢/sharenone
Complementary-set arbitrageLane closedLive opportunity inventory zero; 39% of the lane loses; only 22% of it is genuinely risklessnone
Ladder arbitrageFalsified twiceTwo independent data paths, both netting to zeronone
TWAP tail lockWhole family closed0 of 4 tenors; information real, cost assumption falsenone
NFL player propsKilled three waysMy own proposal, overturned by adversarial verificationnone
Public forecasts vs weather pricesAlready priced inThe 12–24 hour band is fully priced; nothing left
BTC direction agent fleetArchivedNo edge against the real book
A newer venue's outcome marketsDid not enterZero market-making rewards; total spread pie across the venue was $600/day
← scroll for the full table →
By strategy
follow the money to its source

Ordered shallow to deep. If you're not sure where you stand, read A through H in order.

All entries
54 of them, every number checkable

Every entry states where its numbers came from and how to verify them. Knowledge is free here; only the engineering to run the whole thing is separate.

54ENTRIES
7SECTIONS
9KILLS

AMechanics & ground truth10/10

How fees work, who resolves, how subsidies pay, how often the rules changed

BMarket making & rebates10/10

Does quoting for rebates actually pay — adverse selection versus the rebate

CStructural arbitrage4/4

Is "riskless arbitrage" real, and how much of it is left

DSports5/5

Where the money is: vig, devigging, odds sources, real winner archetypes

ECrypto up/down6/6

Where I lost the most, and the market-by-market verdicts

FResearch method11/11

How to tell whether a strategy is real. The most valuable section here

GEngineering8/8

API semantics, data traps, circuit breakers, reconciliation

By market
one-line verdict on each
Crypto up/downsix kills
ecosystem ceiling ≈ $95/day
Sportswhere the money is
but the winners aren't who you think
Weatherthe reward pool is real, the ladder isn't
12–24h band fully priced
Politics & long tailzero fee rate, pure judgment
no structural edge
Esportsfee rates measured
the odds feed is the gap

道具ToolsTOOLS

Things you can actually run. Not an ebook.

Rebate ladder calculatorSIX TIERS · CATEGORY WEIGHTS
Taker fee calculator0.07·p(1−p)·SHARES
True arbitrage cost calculatorMEASURED, NOT ASSUMED
Pre-launch checklistFROM A REAL DEFECT LOG