Do the books time the market, or did the backtest get lucky?
Twelve books. Seven are daily stock and ETF books. Five are spot crypto: Bitcoin on a daily bar, and Bitcoin, Ethereum, Chainlink, and Solana on an eight-hour bar. Each one has a fixed rule. This page says what that rule is, then whether the trades beat a random reshuffle of those same trades, whether costs wipe that out, and whether new tape agrees. The reshuffle mostly holds. The new-tape record is mixed, and the misses are on the page.
Every book against every test
One row per book. Left to right, the tests get harder: a random reshuffle of the same trades, then those same trades laid on a plain 50/200 trend, then costs, then tape the rules had not seen. A dagger on a timing chip means the Calmar pass came from the return, not from a shallower drop. Click a book for its record.
Two families, tested apart
A signal forms on the close. The book is then either in a position or out. The daily books trade one stock or ETF. Six of the seven can be short as well as long. They are in a position on roughly half of days, and a position lasts one to three weeks.
Daily books
Crypto books
Five books, all long only, all spot. Bitcoin is here twice: once on a daily bar, once on an eight-hour bar. Ethereum, Chainlink, and Solana are eight-hour only. None of them can short, so a falling market helps only if the book is already in cash.
What each book does
Every book buys when price has stretched with the trend and the tape is not chopping, and it sells when that stretch fails or a stop is hit. It is in cash the rest of the time. Six of the daily books can also short. TQQQ and all five crypto books are long only. The verdicts are in the scorecard above; the full record for each book is under Sleeves.
Daily
Crypto
Three ways a good-looking backtest fools you
Why timing could exist at all. These assets tend to move in bursts with chop in between. A daily-reset leveraged fund loses most to compounding in the chop, and spot crypto gives back a lot of a burst in the chop that follows it. A book that takes the bursts and stands aside in the chop has a coherent reason to do better than holding. That is a hypothesis. Beating a random reshuffle mostly shows the book caught a trend. The harder check is whether the same trades, laid on a plain 50/200 moving-average rule, do worse. That column is on the scorecard.
Is the timing better than luck of placement?
Take one daily book's own positions: how many, how long, and which are long or short. Keep all of that. Move them to random places on the same price history, 5,000 times. Every random version is invested the same share of the time and pays the same costs. The only thing that changed is when the positions happened. If the live book is no better than those versions, its edge is luck of placement. The five crypto books get the same test in the next section. They are not in these tables.
The p-values
Three columns are the reshuffle. Calmar p is family-corrected in the Holm column. Return p and drawdown p are the same 5,000 draws, not family-corrected. The last two columns are the separate 50/200 test.
Does it hold if we only use recent years?
The same test, on shorter windows. "2022 onward" is the window behind the 5 of 7 count at the top of the page: five daily books still clear a corrected p of 0.05, and NAIL and ERX do not. A shorter window has fewer positions, so a miss there is weaker evidence than a miss on the full history. The p-values are corrected across the seven daily books, separately in each window.
Bitcoin daily, and the four eight-hour books
Same test as the daily stock books. Each crypto book is long only. Bitcoin daily uses one bar a day. The other four use an eight-hour bar, seven days a week. The five are judged together.
The p-values
Same split as the daily table. Holm is across the five crypto books, separately from the daily family.
Does it hold if we only use recent years?
Same test on shorter windows, still judged across the five crypto books. Solana's tape starts in 2021, so its full history and its 2022 window are almost the same years. Bitcoin daily's typical hold is about a month on the full history and about ten days from 2022 on, so the recent window is a different book of trades.
The five daily books that clear, held together
How they move together
Correlation of daily returns on the common window, after the same 5 bps and 5% borrow the timing test uses. A sleeve in cash that day contributes about zero.
Does it survive trading costs?
The tests above charge 5 basis points per side. The daily books also pay 5% a year on short positions. Here the same books are re-run with costs up to 50 basis points per side. The comparison is to buy-and-hold, which pays no turnover cost. Costs cut the numbers. The question is whether the result stays above buy-and-hold.
Daily books
Calmar after costs, full history. “vs B&H” is the book at 50 bps per side against buy-and-hold's Calmar. Short borrow at 20% a year, with 5 bps per side, is in the last column.
Eight-hour crypto
Long only, so there is no borrow. The same cost ladder, on each book's full history.
Does it hold on tape the rules had not seen?
This is the test that separates skill from fit. Parts A to C are research freezes on single sleeves. Part E scores the same live books the Performance page shows: the equity V7 book from go-live, and the crypto V6 book from its forward floor. Corrections sits next to this page and covers SPX 10%+ drops. Read E with those curves, not instead of them.
A · Rules frozen at the end of 2023, scored from 2024
Rule parameters fit only on data through 2023, then traded on 2024 to mid-2026, net of costs. This uses the frozen “core” versions of five sleeves, not the exact live knobs. NAIL and ERX are left out because their 2023 fit did not clear the screening floors, so they have no clean frozen version.
About 2.5 years. Net of costs. The p column tests mean daily excess return over the ETF in 10-day blocks. It is the weak spot, covered in the ledger.
B · Re-fit from scratch
If the method is real, running the same search with no hindsight at an earlier cutoff should land on something that works. It sometimes does and sometimes does not.
C · Sleeve forward clocks
A clock freezes one live sleeve on a date and scores it on the sessions after, against buy-and-hold. The gate was set before scoring: at least 20 sessions, and the sleeve has to beat buy-and-hold. These are sleeve tests. The book test is part E.
D · Crypto Pine lock
Do the books on the Performance page clear unused tape?
Same gate as the sleeve clocks: at least 20 sessions, and the book must beat equal-weight buy-and-hold on return. Numbers come from the live Holocron performance APIs, the same source as the Performance page. Leave-one-out asks whether one sleeve carries the beat.
Leave one sleeve out
Equity V7
Crypto V6
What the evidence does not support
What would settle it
More time on the live books, especially equity past a full year from 2026-07-01, with leave-one-out still clean and sleeve clocks no longer failing their own gates. Until then the right reading is: the live books beat buy-and-hold on unused tape, the timing is not luck of placement on the books that clear the line, and sleeve-level generalisation is still open. Solana, NAIL, and ERX remain the weak sleeves on the in-sample recent windows.
One book at a time
Daily
Eight-hour crypto
Capacity is a liquidity limit
Daily stock and ETF books only. The spot crypto books are not in this table. Soft limit is 1% of median daily dollar volume for that sleeve. Hard limit is 5%. Above the hard limit, participation is large enough to impair fills. These are participation ceilings. They are not return estimates.
An equal-weight seven-sleeve book is soft near $1.6M and hard near $8.0M, because every sleeve has to fit the tightest name. The binder is ERX.
How the numbers were made
Timing test (Tests 1 and 2)
- Books: the live webhook configuration of each sleeve, run through the engine, then rebuilt as a daily position series. The position is taken from the bar after the signal, and daily return is position times close-to-close return (an approximation to next-open fills), less 5 bps per side on turnover and 5% a year on short bars.
- Null: the book's own positions (count, lengths, long/short mix) placed at random, without overlap, on the same tape. 5,000 draws. Daily seed . Eight-hour seed .
- Statistic: Calmar, annualised return over maximum drawdown on daily marks. p is the share of random placements at least as good as the live book, counting the live book as one draw. One-sided.
- Correction: Holm inside each family, separately for each window. Seven daily books together. Five crypto books together. Windows start at each book's first position, or the stated year, and run through the last bar of that book's tape. Daily stock tapes run through . Crypto tapes end on the dates in the crypto table, in April or May 2026.
- Crypto calendar: Bitcoin daily uses 365.25 bars a year. Eight-hour books use 365.25 × 3 bars a year. A hold is written in days. Daily stock books still use 252 sessions a year.
- Fidelity: the rebuilt equity is checked against the engine's own trade-level equity. Correlation is shown on each sleeve record. It is lowest on NAIL and ERX.
- Buy-and-hold uses the same price series and statistic.
- Naive trend: a 50-bar and 200-bar average. Long when price is above the slow average and the fast average is above the slow one. Short, on books that short, when both are below. Eight-hour books use 150 and 600 bars, the same two horizons in clock time. Each of 5,000 draws keeps the book's own trades and starts each one at a random date where at least half its bars agree with that side, with no overlap. A trade that cannot find such a date fills the next free gap. p is the share of those draws whose Calmar is at least the live Calmar, with the same +1 correction as the reshuffle, then Holm inside the family. “Calmar, raw 50/200” is that moving-average rule run as its own book, with its own time in the market. It is not the exposure-matched number.
- The five-book portfolio equal-weights TQQQ, TSLA, GBTC, IONQ, and MSTR on the dates all five have a bar. Its reshuffle places each sleeve's trades at random on that sleeve's own tape, independently, 5,000 times. One p, no Holm.
- GBTC long-only before 2024 sets every short to cash before 2024-01-11 and leaves the later shorts in place. Same costs.
- Return p and drawdown p are the same 5,000 reshuffles as the Calmar p, without a family correction. A Calmar pass with a weak drawdown p came from the return, not from a shallower worst drop.
- Drawdowns are daily mark-to-market, including open positions. They are deeper than trade-closed drawdowns, which only update when a position exits.
Out-of-sample pieces (Test 3)
- Frozen-2023, re-fit and sleeve clock results are taken from the research runs named on each row. They were run on earlier data cutoffs than the timing test and use each study's own cost settings.
- Sleeve forward clocks use Yahoo daily bars filled to the scoring date, mark to market, rebased at the freeze.
- Live book clocks pull the equity and crypto portfolio series from the same Holocron performance APIs the Performance page uses. Equity freeze is V7 go-live 2026-07-01. Crypto freeze is the V6 forward floor 2025-10-13. Gate: ≥20 sessions and book return above equal-weight buy-and-hold.
- Leave-one-out rebuilds equal-weight daily returns after dropping one sleeve. Equity uses sleeve NAVs through the sleeve cache tip. Crypto leave-one-out uses the static daily sleeve series in crypto_performance.json; the live crypto book can differ because unstarted weight sits in BTC until each sleeve's first trade.
- Published trade counts elsewhere on the bridge come from the historic simulations and can differ from the position counts on this page, which come from the rebuilt position series.