State of the Lab — Oct 18, 2026
Real money is back on, and speed turned out to be the problem. Plus two studies, and the night vs day test's verdict.
No. 12 · 18 Oct 2026 · All ponds
State of the Lab, week ending Oct 18, 2026
The Heron · 4 min read · bd3c5b68…97b90b 2026-10-18T04:56Z
Real money is back on. On Wednesday our bot started trading the last minute rule with real money, one contract at a time: the paper strategy that held up best through October, with its stops written down before the first order. It has placed 9 orders and settled 4 real trades. That is far too few to judge, so we won't show a per trade number yet. The verdict comes at 60 real trades, pass or fail, and members can watch every one of them on the live trading page.
The surprise so far is not the price. It is speed. On Thursday the bot got a faster engine that sends its order well under a tenth of a second after the signal. It still missed: four orders in a row found the offer already gone, because in the last minute of these markets the book clears in about a quarter of a second. On Saturday it got a small price allowance, so it can still trade when the price ticks up a cent or two on the way. We measured how often a slow bot misses, and which trades it misses, in this week's first study: our live Kalshi bot reacts in under a tenth of a second, and it still missed. Short version: the offers that vanish fastest are the ones on the side that goes on to win.
The second study started as a speed question too. Could a faster bot copy Kalshi's biggest traders? We tracked 1,855 trades of $1,000 or more, about $4 million in all, and the big money lost 4.9 percent of it. Following them a quarter of a second later landed in exactly the same place, because there was no move to catch. That study also contains our own mistake: the first run said whales made 14 cents a contract, until we found a bug that flipped the price of every No trade. It is fixed, and the whole thing is here: we tracked $4 million of Kalshi whale trades.
The night vs day test is over, and it failed. A viewer suggested our one winning strategy only died in US daytime, so we ran its full rule on paper at all hours with the pass mark written down first: night had to be positive, in both halves, and beat day by two cents. Night ran 311 honest fills at minus 1.09 cents a contract. Day ran 490 at plus 0.43. The hour split was hindsight. It also answers the question of how our winner would do if we switched it back on today: across all 801 fills, about flat. The edge it had in September is not there now.
The rest of the board, all paper. Perp pressure lost $33.03 over 1,638 honest fills, and the ten crypto taker rules lost $499.00 over 30,691 trades, about 1.6 cents each: paying the asking price keeps costing about what the edge is worth. We stopped the commodity favourite on Wednesday to free memory on the lab's server; it had already failed its September pass mark. Two new paper tests started on Wednesday: the last minute rule on eight other coins, +$1.76 over its first 181 fills and judged at 200, and an hourly version, 18 fills so far. A stop loss on late Bitcoin favourites is running as a forward test, 44 entries in, judged at 300.
Our AI forecaster, Claude with web search setting its own probabilities, has 123 settled. Its Brier score is 0.254, and Kalshi's own prices scored 0.159 on the same markets. Lower is better, so the market is still well ahead. It runs to mid November.
Still hunting. Fight Night, every real-money fight our bots have had, premieres on YouTube on Wednesday.
Real money this week
| Real money | Orders | Settled real trades | Result |
|---|---|---|---|
| Last-minute rule, Bitcoin (live since 14 Oct) | 9 | 4 | too few to judge yet |
This week, on paper
| Paper strategy | Trades | Net | ¢/trade |
|---|---|---|---|
| Perp pressure, honest fills | 1,638 | −$33.03 | −2.02¢ |
| Night vs day test (the winning rule, all hours) | 754 | −$1.60 | −0.21¢ |
| Late favourite sniper | 466 | −$0.61 | −0.13¢ |
| Commodity favourite (stopped 14 Oct) | 356 | +$14.97 | +4.21¢ |
| Last-minute rule, 8 other coins (new) | 181 | +$1.76 | +0.97¢ |
| Sniper on four coins | 177 | −$0.45 | −0.25¢ |
| Mention markets | 76 | −$1.80 | −2.37¢ |
| Speech markets | 31 | −$1.94 | −6.26¢ |
| Last-minute rule, Bitcoin | 29 | +$0.13 | +0.45¢ |
| Last-minute rule, hourly (new) | 18 | −$1.59 | −8.83¢ |
| Tennis tournament favourite, resting | 8 | +$1.08 | +13.50¢ |
| Rotten Tomatoes | 2 | +$0.36 | +18.00¢ |
| Ten crypto taker rules, 15 minute markets | 30,691 | −$499.00 | −1.63¢ |
Paper: simulated fills against Kalshi's real prices, one contract, fee included. From this week a paper fill only counts when a real trade went through the order's price, so these numbers are lower than last week's method would show. Paper is still a best case. Members get the all-time board, every rule in detail, and every real-money fill in the archive.
Every trade above was logged live, then sealed: each night the lab hashes its raw market archive and trade ledgers into an append-only chain. Rewriting any past trade would change every head after it, so check this one against last week's issue.
Chain head as of 2026-10-18T04:56Z: bd3c5b68d36c9c20ccc2ea74447b428d70291d7b88da26bfd00a03acab97b90b