The Lab

What changed, and when

The live studies' own record: when each was added, and every change to how they are worked out. New pages, write-ups among them, are on what's new.

• 12 Oct 2026 · First study run: does the order book's imbalance predict the next move? On 29 hours of record from ten coins, heavier bids came before falls, not rises.
• 12 Oct 2026 · The years added, from Binance's archive back to 2022. A timing error was found and fixed before they were published: Binance stamps each market-order bucket with its start, and counting it in the hour before let a signal see part of the move it was meant to predict. It had turned one answer around.
• 12 Oct 2026 · Binance's history re-collected on a corrected clock (its archive's times ran five minutes early); every study from the years worked out again on it.
• 12 Oct 2026 · The order-book answer turned around as 23 newer coins joined the record, so it is now marked too early: under two weeks of record.
• 12 Oct 2026 · The Lab rebuilt as research: each idea, its exact rule, a verdict and the same tests for all. The settings tested were fixed. The new bounce test showed about half of the leverage build-up result is the price bouncing back from its own move.
• 12 Oct 2026 · Crowded funding added, from Binance's funding history back to 2022. Its first answer: partly. High funding came before slightly weaker days, but much of that is the price bouncing back, and it is too small to pay its costs.
• 13 Oct 2026 · The real test begins: from today, each study is also judged on record that no one had seen when its settings were fixed.
• 13 Oct 2026 · Liquidation cascades added, from our own record of liquidations. Its settings were fixed before any of its numbers were seen; it waits for two weeks of record.
• 13 Oct 2026 · The hour's habit added, from Binance's archive back to 2022: does the time of day carry its own drift? Its settings were fixed before any of its numbers were seen.
• 13 Oct 2026 · Gold (XAUT, tokenised gold) taken out of every study: it is not a cryptocurrency, and a coin that moves apart from the rest skews a test of whether most coins agree. Decided before its effect was seen; every study worked out again on the 32 other coins. Two verdicts rose from partly to holds up, both on the edge of a test: fade the crowd (23 of 33 coins agreed, just under 70%; 23 of 32 is just over) and leverage build-up (with the bounce taken out, t −2.6 with gold, −3.2 without). Read both as borderline until the new data speaks.
• 13 Oct 2026 · Into the funding payment added: does the price move into Binance's funding payments the way the payers would push it? Suggested by a first look at study 8's longer horizons; its settings were fixed before any of its own numbers were seen.
• 13 Oct 2026 · Funding apart added: when Hyperliquid's funding runs above Binance's, does the price fall? From Hyperliquid's funding history back to May 2023. Its settings were fixed before any of its numbers were seen.
• 13 Oct 2026 · Memes behind Bitcoin added: does Bitcoin's last hour lead the meme coins' next? On the nine meme coins, apart from the Lab's fixed list. Its settings were fixed before any of its numbers were seen.
• 16 Oct 2026 · Every page now names its authors, PerpGremlin and Claude, and ends with its Credits: the published work each rests on. A new group, Theory (T1 to T9), explains the ideas behind the tests.
• 16 Oct 2026 · Edits to published pages, each checked again: a "Last checked" date on every write-up, from a weekly check that its finding still stands (the live studies, 1 to 11, are worked out again at every update, and show its time instead); momentum's swings stated as "90% or more in a week"; the grid test's bar of 15 of 20 shuffled twins, with what luck alone gives; and on the carry trade in one account, an update from a test on an exchange demo account. No verdict changed.
• 16 Oct 2026 · Added where the idea comes from, on 21 write-ups, each pointing to its theory page; and one credit on five of them: rebalancing (Fernholz, 2002), steady buying (Statman, 1995), momentum (Levy, 1967), the carry trade (Working, 1949) and survivorship (Wald, 1943). No finding changed.
• 16 Oct 2026 · Corrected the data source line on Shannon's demon (T1): futures prices stood in for spot in its first two tests.
• 16 Oct 2026 · Two new write-ups: when a carry trade on a loan should stand down (#48), and the QFL signal in a TradingView accumulator (#49).
• 17 Oct 2026 · Corrections from an outside check of every page, each a dated "Update, 16 October 2026" part on its page, the text above it left as it was. No verdict changed. Funding carry (theory): figures from an unpublished test, since withdrawn, restated from published pages; "survived every crash" narrowed to Bitcoin and Ether. Mean reversion (theory): the drop bounce was bigger, but not reliably so; Lo and MacKinlay restated. Luck and controls (theory): Harry Roberts credited.
• 17 Oct 2026 · More of the same corrections. Funding carry and the hedged basket: its two years were one up and one down. Momentum on spot: the yearly figures compounded (−16%, +30%, +21%), and the blue-chip result noted. The carry trade in one account: at a 5% loan, borrowing earned a little more; "safe leverage" defined; its outside source named. Carry stand-down: the demo book held about 5.7 times its money, not 10; Solana's gap corrected. QFL: duplicate runs counted once, 11 of 216 against about 11 by luck; and the two scripts described without "open source".
• 17 Oct 2026 · More of the same corrections. CPI follow-up: the low was at 02:28 UTC; Bachelier credited for the doubling rule. CPI and PPI levels: a note on the uneven odds. Grid bots on 14 coins: a caution that the coins were picked after the window. Grid follow-ups: the with-stop counts labelled. How a grid works: Credits added. Small corrections and labels on liquidation levels, the S&P 500 check, touch odds, Sharpe, the drop bounce, the bracket bot, rungs, models, market moods and pairs. And study titles given in full wherever pages refer to them (11 pages).
• 17 Oct 2026 · A new write-up: the touch map in calm and wild markets (#50). A table of odds for each mood, read from the volatility rank or from the hidden Markov model, against today's single table; neither cleared the bar set in advance. And a new page, Calm or wild, shows each coin's volatility rank, hourly.
• 18 Oct 2026 · Three new pages. The volume clock (#51): counted in coins traded, returns sit much nearer a bell curve, and the touch map's odds gain from an hour to a day but not at a week, so it misses the bar set in advance. The BDS test (#52): returns have memory, which a standard volatility model explains for days but not at four hours. And a theory page, quantum finance (T10). The first charts drawn in a write-up, on the volume clock.
• 18 Oct 2026 · Charts added to 17 write-ups, most with a short Update part saying what they draw: First passage; The touch map in calm and wild markets; How far to trust a Sharpe ratio; Hidden order: the BDS test; Buying steadily or all at once; Shannon's demon; Grid bots: more rungs or fewer?; Momentum across coins, and tested again on 2018 to 2020; Pairs trading; The carry trade in one account; When a levered carry book should stand down; Do coins have a clock?; Ending up with more Bitcoin; and the theory pages on luck and random twins, Kelly, and Shannon's demon.
• 18 Oct 2026 · A new write-up: coin-flip traders (#53), the first of a series testing published edges against the coin. 90,000 random traders a setting on Bitcoin, Ether and Solana; about 1 in 100 beat the fee, and fewer than half beat even no fee.
• 21 Oct 2026 · A new series, Tokenised assets. Stock perps while the market is shut (#54): on weekend hours alone the perp moved about half of Monday's gap, and by the open it has followed the stock's own early trading. Gold on crypto exchanges (#55): every gold token and perp tracks gold within 2 to 5% a year; no grid beat chance.
• 21 Oct 2026 · Correction: Binance's readings of 15 Oct 2026, 11:00 to 14:55 UTC, were wrong (frozen, halved or zero on every coin, in Binance's own archive too). They are now left out as missing, and the nine studies from the years were worked out again. No verdict changed. Some figures moved slightly, the most: #5's gap from 19.4 to 19.1 bps; #9 from 28 to 27 of 32 coins pointing the same way (the bar is 70%); #4's gap from +0.1 bps to about none. The "so far" figures on the new record moved more, as 15 Oct is one of its eight days.