The lab
Everything we have tried on these markets, written up as lab reports: the hypothesis, the test, the measured result, the conclusion. 29 experiments — most of them failures, which is the point. Nearly every "edge" in circulation dies at one of three steps: executable prices (can you actually get the fill the backtest assumed?), fees and spread (the toll peaks exactly where the signals live), or the luck test (was the win rate ever distinguishable from chance?). The entries below show which step killed what. As of the latest update, nothing we have tested holds a confirmed edge at real prices — two strategies we once marked confirmed were downgraded when later real-money results came in.
Entries for strategies we still run omit some live parameters. Everything else is disclosed in full.
Open experiments (3)
Lab note
Still testing
The dormant wallet that wakes up
A wallet that traded, went silent for months, then returns with one large bet. Past its 30-trade gate it is still positive — but not by more than luck could produce.
Lab note
Still testing
Mid-band mega prints: conviction money at coin-flip prices
Across 112,000 resolved whale bets, six-figure buys lose at nearly every price — except between 45¢ and 70¢, where they ran +23%. The prints stopped for a month, then came back: six markets, six wins so far.
Lab note
Still testing
Liquidity-reward farming
The platform pays daily rewards for resting orders near the mid on designated markets. Quote both sides as buys, stay inside the reward band, never improve the touch — and learn why the lonely market maker is the sniper's favorite target.
Tested clean — nothing there (10)
Lab note
No edge
Racing the oracle at the final whistle
The game just ended; the market hasn't resolved yet. Buy the winner for 97¢ and collect a riskless 3¢? We watched every final. There is never anything to buy.
Lab note
No edge
Fading consistently bad traders
Find wallets that reliably lose, bet against everything they buy. On paper it returned +38%. At real order books, the trade you'd actually place earns almost nothing.
Lab note
No edge
Take-profits and stop-losses: every exit rule lost to holding
Lock in the double? Cut the loser? We tested both against our own real positions. Holding to resolution beat every mechanical exit we tried — including in the case built to favor it.
Lab note
No edge
Following "proven winners": trader skill does not persist
Screen for wallets with the best lifetime records, then copy only them. The cleanest-sounding strategy on the board — and the data killed it twice.
Lab note
No edge
Copying whale bets in real time
The founding strategy: watch large bets hit the tape, copy them within seconds at small size. On real money it made +2.1% — and all of it came in one World Cup month.
Lab note
No edge
Scanning for markets that are already decided
Somewhere in the long tail, a market's outcome has already happened and the price hasn't noticed. An LLM read the rules and the news for thousands of markets, hunting for exactly that. The bounty didn't cover the ammunition.
Lab note
No edge
Logical-implication arbitrage, with an LLM finding the logic
If market A happening logically requires market B, then A can never price above B — any violation is riskless profit. An LLM found over a thousand implication pairs. The books never once mispriced them.
Lab note
No edge
Crypto-candle arbitrage: Up + Down < $1 never happens
If Up and Down ever sum below a dollar, buying both locks a riskless profit. We scanned for it 4,195 times, then tracked every market's entire life. Total opportunities: effectively zero.
Lab note
No edge
Options-implied probabilities vs crypto digitals
"Will BTC be above $K on date X" is a digital option, and the options market prices those with billions in open interest. Compare the two venues and trade the gap. There is no gap.
Lab note
No edge
Sum-to-one and ladder arbitrage across whole events
Multi-outcome events must sum to $1; probability ladders must be monotonic. Scanners find "violations" all day long. Then you check the live book, and nearly every one is a mirage.
Where the strategy lost money (16)
Lab note
Negative
Weather tails: non-modal buckets and sub-cent lottery tickets
Born from our audit of the "weather legends" article: test both the lottery-ticket story and the one real winner's style, at executable asks. Both lanes lost — the lottery lane catastrophically.
Lab note
Negative
Mean reversion at maker prices
Crypto candles genuinely mean-revert — we measured it across 118,000 candles before building anything. The pattern clears maker breakeven and nothing else. Then 17,491 live maker fills answered the only open question: adverse selection ate it.
Lab note
Negative
Fading fresh listings
A 27,000-listing calibration study found brand-new markets systematically mispriced by band: mid-priced listings too rich, favorites too cheap. The executable follow-up looked like the best thing on the board at 29 settlements (+53.6%) — and read −15.2% at 289.
Lab note
Negative
When a whale dumps a favorite
A whale selling a 70–97¢ favorite should know something, and historically the dumped favorites did fail more than their price implied. Buying the other side at live prices lost money in both lanes.
Lab note
Negative
An LLM judging violent repricings
When a market moves 20¢ in an hour, can an LLM tell real news from panic before the book settles? Early on it looked like it could. Past its 30-trade gate, the high-confidence calls lost money.
Lab note
Negative
The GPT weather forecaster
An AI bot combining rule-based temperature models, GPT-4o, and two weather APIs, betting daily temperature markets. Lifetime score: −$138, plus the API bill.
Lab note
Negative
Numeric forecasts vs temperature-bucket prices
No LLM this time: pure meteorology against market asks, at seven different lead times, with three forecast models and per-station bias correction. The asks won every single round.
Lab note
Negative
Buying near-certain favorites before resolution
We expected 90–98¢ favorites close to resolution to be slightly underpriced. With real money they won less often than their prices implied — and each loss cost about fifteen wins.
Lab note
Negative
Order-flow confirmation on candle leaders
Follow the leading side when exchange flow confirms it; fade it when flow contradicts it. A widely-circulated rule, implemented exactly as published — with the entry price the publication forgot.
Lab note
Negative
The passive dust-bid grid
A real wallet grinds out six figures resting cheap bids on both sides of every crypto candle, fee-free. We simulated the same grid honestly for a week. It lost 31.6%.
Lab note
Negative
Trading the observed temperature maximum
Forget forecasting — wait until the day's high has already happened, then buy the winning bucket. The purest information-latency trade we could design, and it still went 1-for-6.
Lab note
Negative
Sniping "stale" quotes after a spot move
Bitcoin jumps on the exchange; the prediction market's quote hasn't moved yet. Buy the lagging side before it reprices — the classic latency trade. The quote wasn't stale. It was right.
Lab note
Negative
The staircase: buying underdogs as they climb
A 27-million-trade backtest said underdogs that climb keep climbing — +15% per cycle. At real order books the same rule lost 13.5%. The difference was exactly the spread.
Lab note
Negative
The "fresh wallet insider" signal
A brand-new wallet funds up and immediately places one big, focused bet. Surely that's informed money. Five thousand paper copies say: no.
Lab note
Negative
An LLM vs the market on breaking news
Give a language model the resolution rules and fresh headlines, hide the price, and ask for a fair probability. Where it disagreed with the market, the market was right.
Lab note
Negative
Fading miscalibrated price bands
Historical data showed some price bands win less often than their price implies — so sell into them. The calibration was real; the trade was not.