Irrational Tape
DEEN
REGISTER

What's fixed in advance can't lie afterwards.

Every test is registered in advance: rules, costs, checkpoints, timestamp. After that, we measure. Below: our own verdicts.

A hash is a fingerprint: anyone who has the text can recompute it.

Key points

Families checked
24
passed
0
on probation
1

Verdicts

This is where the strategies lie that failed our own tests. Every card has the same size, whatever the result. We show it because a tool that wants to separate luck from skill has to show on ourselves how often “looked good” was just luck.

Score

43 cards · 24 pre-registered families checked · 0 passed

  • 26 registered, each with a hash before any run
  • 19 failed
  • 1 undecided
  • 2 not testable
  • 2 running forward
  • 2 not computed yet

We fixed every idea (“family”) before testing it and sealed it with a fingerprint (hash). Each family was allowed at most a 1 % false-alarm rate. “Checked” means computed or classified (not testable, undecided, still running). Two families are waiting for data. One family half failed and is half still open; it counts as undecided. Two plain measurements do not count as a family. “0 passed” does not mean no edge exists. Source: Result reports of the last two rounds, as of 18 Oct 2026. As of 2026-10-19 12:30.

How to read a card

Our own strategy research. All results in R (result divided by the risk at entry) after costs unless stated otherwise. No price data, no firm names. “U” next to a number: it is in no result report, solely in an older list.

  • Development (dev): data the rule could be built on. Validation (val): later data, looked at once. Holdout: the most recent years, opened exactly once, with a log line before.
  • n = number of trades. Mean R = average result per trade after costs. t and p: how far that is from chance; our bar was 1 % per family (0.5 % per rule with two rules).
  • Cost stress = the same computation with 1.5× costs.
  • “Base stream” = our own running stream. We do not publish its rules; we do publish its results.

How the tests run: How we test (6 steps). How we measure trading records against chance: Methods paper ZSB-1.

Pre-registered (forecast log, error rate 1 % per family)

Round 1 · 3 cards · 0 passed
R1-F3 Multi-day drift after scheduled releases · Failed · 2026-10-15

R1-F3

Multi-day drift after scheduled releases

Failed · Rest in variance.

Hypothesis
Slow participants carry the direction of the event day (rate decision, inflation, jobs) over one to two days. Enter at the close, exit at the close two days later, when the day moved ≥ 0.5 %.
Data
Index future (A) and precious-metal micros (B), minute data, development 06.2010–2019; validation and holdout not opened
Result
A: n 130, −60 $ per trade, t −0.55. B: n 163, +105 $ per trade, t 1.26 (cost stress +96 $).
Lesson
With 13–16 dates a year, the test lacks power, not the effect alone.
Cause
Too few events for a test
Record
Prereg R1, 2026-10-15 12:17, sha256 708e99ca…0480
family round 1 result report (section L5)
R1-F4B Drift before the rate decision in high volatility · Failed · 2026-10-15

R1-F4B

Drift before the rate decision in high volatility

Failed · Rest in variance.

Hypothesis
A premium for resolving uncertainty, when the volatility index closes ≥ 20 the day before; long 14:00 the day before to 13:55 on decision day.
Data
Index future, minute data, development 2010–2019
Result
n 14, −46 $ per trade, t −0.22, cost stress −61 $.
Lesson
One extra condition turns 80 events into 14. That is no longer a test, it is an anecdote.
Cause
Too few events for a test
Record
Prereg R1, 2026-10-15 12:17, sha256 708e99ca…0480
family round 1 result report (L5). Part A of the same family is undecided (see below).
R1-F5 Same-day options: iron fly with filter · Failed · 2026-10-15

R1-F5

Same-day options: iron fly with filter

Failed · Rest in variance.

Hypothesis
Buyers of far out-of-the-money options and hedgers pay a daily premium. Short straddle with wings at ±1.0 × expected daily move, entry 10:00 (A) or 13:00 (B), exit 15:45.
Data
Same-day options on the index future, minute data, development 01.2022–2025; fixed cost 0.25 points per leg
Result
A: n 632, net −2.37 points per day, t −7.44. B: n 632, −2.02 points, t −11.0. Gross B +0.34 points against costs of 2.36 points.
Lesson
Four legs cost a multiple of the premium. A premium that exists gross is out of reach net for small accounts.
Cause
Costs larger than the gross effect
Record
Prereg R1, 2026-10-15 12:17, sha256 708e99ca…0480
family round 1 result report (L3)
Round 2 · 1 card · 0 passed
R2-F1 Drift after the weekly oil inventory report · Failed · 2026-10-16

R2-F1

Drift after the weekly oil inventory report

Failed · Rest in variance.

Hypothesis
The direction of the 15 minutes after the report continues until 14:29, threshold 0.6 % (A) or 1.0 % (B).
Data
Oil future, minute data, micro-based costs, development 2010–2019
Result
A: n 155, mean −0.009 R, t −0.09. B: n 74, mean +0.044 R, t 0.44, cost stress +0.019 R.
Lesson
The 15 minutes after the report hold all of the processing; what follows is noise.
Cause
Price pattern without a mechanism
Record
Prereg R2, 2026-10-16 06:50, sha256 240d7e98…16dd
family round 2 result report (part 2)
Round 3 · 3 cards · 0 passed
R3-F2 Bid/ask spread filter for the base stream · Failed · 2026-10-16

R3-F2

Bid/ask spread filter for the base stream

Failed · Rest in variance.

Hypothesis
A wide spread in the decision minute signals a thin market; trade when the spread is ≤ P80 (A) or ≤ P90 (B) of the same minute over the previous 60 days.
Data
Base stream (development 518, validation 261, holdout 348 trades), per-minute spreads 2010–2026
Result
A: difference kept − all: development +0.010, validation −0.038, holdout +0.014 R (p 0.34). B: validation −0.019, holdout −0.004 R (p 0.55).
Lesson
A filter on a positive strategy has to win paired, not in absolute terms. A spread that is almost always 1 tick filters nothing.
Cause
Not better than the base, paired
Record
Prereg R3, 2026-10-16 19:40, sha256 6b229f5d…d267
round 3 result report (16 Oct 2026 23:58)
R3-F4 Shorter profit target for evaluation accounts · Failed · 2026-10-16

R3-F4

Shorter profit target for evaluation accounts

Failed · Rest in variance.

Hypothesis
On a 50K-type evaluation account (target 3,000 $, max loss 2,000 $) a shorter profit target yields more net payout over 12 months (median), even though the per-trade mean falls.
Data
real trade sequences of the base stream, real trading days, 3,000 paired paths (block bootstrap of days)
Result
Validation: median +296 $. Full holdout: −1,384 $ (p 0.70), losing year 49 → 60 %. Clean holdout: +49 $ (p 0.76).
Lesson
An idea formed from holdout numbers already seen is half a piece of evidence.
Cause
Idea formed after looking at the data · Different in development and validation
Record
Prereg R3, 2026-10-16 19:40, sha256 6b229f5d…d267
round 3 result report; re-run window 2 (17 Oct 2026, 10,000 paths)
R3-F5 Market weighting by rolling 12/24-month return · Failed · 2026-10-16

R3-F5

Market weighting by rolling 12/24-month return

Failed · Rest in variance.

Hypothesis
Markets have phases: size 1.5× if the market's rolling mean is > 0, else 0.5×.
Data
Base stream 1,127 trades 2013–2026; windows 12 months (A) and 24 months (B)
Result
A: validation −0.006, holdout +0.050 R (p 0.056 at threshold 0.5 %). B: validation +0.003, holdout −0.034 R (p 0.87).
Lesson
With 6–38 trades per market and year, the sign of the annual mean is wrong 36–42 % of the time. The forecast “noise” was in the log beforehand and came true.
Cause
Too few events for a test · Not better than the base, paired
Record
Prereg R3, 2026-10-16 19:40, sha256 6b229f5d…d267
round 3 result report (16 Oct 2026 23:58)
Round 4 · 5 cards · 0 passed
R4-K1 Reversal after the FX fixing (16:00 London) · Failed · 2026-10-17

R4-K1

Reversal after the FX fixing (16:00 London)

Failed · Rest in variance.

Hypothesis
The move into the fixing is order pressure that reverts afterwards. Euro and pound futures pooled; B month-end alone.
Data
Minute data, development 2010–2019, validation 2020–2022, holdout 2023–09.2026 once
Result
A holdout: n 432, mean −0.114 R, p 0.9994. B holdout: n 76, mean −0.114 R, p 0.948. Even with zero fees A would stay at −0.041 R gross.
Lesson
Even without costs the reversal after the fixing stays negative (gross −0.041 R per trade): it does not exist.
Cause
No effect even before costs
Record
Prereg R4, 2026-10-17 09:15, sha256 c533abac…3133 · OpenTimestamps
round 4 result report (17 Oct 2026 09:40, §2)
R4-K2 Reversal after the afternoon gold auction · Failed · 2026-10-17

R4-K2

Reversal after the afternoon gold auction

Failed · Rest in variance.

Hypothesis
The move into the auction (15:00 London) reverts afterwards; A short window, B 60 minutes.
Data
Gold future, minute data, development/validation/holdout as R4-K1
Result
A holdout: n 332, mean −0.097 R, p 0.996. B holdout: n 318, mean −0.102 R, p 0.998.
Lesson
A control window that is even worse does not make the rule positive.
Cause
No effect even before costs
Record
Prereg R4, 2026-10-17 09:15, sha256 c533abac…3133 · OpenTimestamps
round 4 result report (§2)
R4-K3 Reversal after the settlement window (oil, gold) · Failed · 2026-10-17

R4-K3

Reversal after the settlement window (oil, gold)

Failed · Rest in variance.

Hypothesis
Pressure in the settlement window reverts afterwards; A oil 14:30, B gold 13:30 (New York).
Data
Minute data, development/validation/holdout as R4-K1; overlaps with the base stream netted
Result
A holdout: n 291, mean −0.086 R, p 0.9975. B holdout: n 310, mean −0.069 R, p 0.935.
Lesson
Negative in development, validation and holdout, even before costs: no effect.
Cause
No effect even before costs
Record
Prereg R4, 2026-10-17 09:15, sha256 c533abac…3133 · OpenTimestamps
round 4 result report (§2)
R4-A1 Hedge-flow momentum into the close · Failed · 2026-10-17

R4-A1

Hedge-flow momentum into the close

Failed · Rest in variance.

Hypothesis
On days with negative dealer gamma and a large daily move, the move continues from 15:01 to 15:58.
Data
Index future, minute data, daily gamma series from 2016, rate-decision days excluded
Result
Validation: n 110, mean −0.017 R. Holdout: n 96, mean −0.049 R, p 0.71. Control group (gamma ≥ 0) holdout −0.157 R.
Lesson
The condition separates (better than the control group), but the rule itself earns nothing net.
Cause
No effect even before costs
Record
Prereg R4, 2026-10-17 09:15, sha256 c533abac…3133 · OpenTimestamps
round 4 result report (§3)
R4-S1 Rank threshold for the base stream · Failed · 2026-10-17

R4-S1

Rank threshold for the base stream

Failed · Rest in variance.

Hypothesis
Trade the strongest candidates of a rolling 20-day window (C1), variant with a larger pool (C6); paired against the base.
Data
Base stream, development/validation, clean holdout 01.2023–09.2025
Result
C1: validation −0.076 R per day against the base; clean holdout +0.012 (p 0.44), annual sum +7 % instead of the required +10 %. C6: validation −0.119, holdout −0.064.
Lesson
The rule smooths frequency but adds no gain on average.
Cause
Not better than the base, paired
Record
Prereg R4, 2026-10-17 09:15, sha256 c533abac…3133 · OpenTimestamps
round 4 result report (§3)
Round 5 · 1 card · 0 passed
R5-DB Reversal on a thin order book · Failed · 2026-10-17

R5-DB

Reversal on a thin order book

Failed · Rest in variance.

Hypothesis
After a large 30-minute move on a thin order book, price reverts; four index futures.
Data
Minute data with book depth, development, validation 2020–2022, holdout 01.2023–09.2025 once (18 Oct 2026)
Result
Validation: n 424, mean −0.075 R, p 0.98. Holdout: n 301, mean −0.065 R, p 0.96. The normal book (control) did better in both.
Lesson
Before the real run we found a counting bug in our own code (Sunday rows) and corrected it openly, before any result was visible.
Cause
Price pattern without a mechanism
Record
Prereg R5, 2026-10-17 11:50, sha256 531a2b33…f733 · OpenTimestamps
round 5 result report (§2, §6)
Round 6 · 5 cards · 0 passed
R6-T3 Index rebalancing (block 1, three index futures) · Failed · 2026-10-18

R6-T3

Index rebalancing (block 1, three index futures)

Failed · Rest in variance.

Hypothesis
Around the index rebalancing date, volume and direction shift briefly; we trade the expected drift in the three large index futures.
Data
Minute data, development and validation through 2019, holdout after that, opened once (18 Oct 2026)
Result
Holdout n 176, net +0.063 R (p 0.21); validation −0.017 R; control −0.144 R (difference +0.207, p 0.022).
Lesson
No holdout p below 1 %; validation negative.
Cause
Different in development and validation
Record
Prereg R6, 2026-10-18 11:25, sha256 9027fef3…3afb · OpenTimestamps
round 6 result report (18 Oct 2026)
R6-KR 60/40 rebalancing (stock and bond futures) · Failed · 2026-10-18

R6-KR

60/40 rebalancing (stock and bond futures)

Failed · Rest in variance.

Hypothesis
Portfolio rebalancing between stock and bond futures creates short-term drift around the rebalance date.
Data
Minute data, development and validation through 2019, holdout after that, opened once (18 Oct 2026)
Result
Variant A: holdout n 570, +0.028 R (p 0.20), validation −0.113 R. Variant B: holdout n 105, +0.024 R (p 0.31).
Lesson
Both variants far from p < 1 %.
Cause
Different in development and validation
Record
Prereg R6, 2026-10-18 11:25, sha256 9027fef3…3afb · OpenTimestamps
round 6 result report (18 Oct 2026)
R6-T4 Month-end (index and bond future) · Failed · 2026-10-18

R6-T4

Month-end (index and bond future)

Failed · Rest in variance.

Hypothesis
Second attempt of family R1-F1 (still running forward there): at month-end, flows between index and bond futures briefly tilt direction.
Data
Minute data, development and validation through 2019, holdout after that, opened once (18 Oct 2026)
Result
Variant A: holdout n 46, −0.011 R (p 0.54). Variant B: n 40, −0.135 R (p 0.77). Cost stress negative.
Lesson
No gain in the second attempt either.
Cause
Different in development and validation
Record
Prereg R6, 2026-10-18 11:25, sha256 9027fef3…3afb · OpenTimestamps
round 6 result report (18 Oct 2026)
R6-K3 Auction cycle (bond futures) · Failed · 2026-10-18

R6-K3

Auction cycle (bond futures)

Failed · Rest in variance.

Hypothesis
Around the regular bond auction, price discovery in bond futures shifts briefly.
Data
Minute data, development and validation through 2019, holdout after that, opened once (18 Oct 2026)
Result
Variant A: holdout n 190, +0.014 R (p 0.42); lead over control p 0.026, from a weak control. Variant B: n 239, +0.054 R (p 0.18). One sub-market (n 55, +0.124 R) was selected after the fact and is not a candidate.
Lesson
A lead over a control is not a stand-alone gain.
Cause
Not better than the base, paired
Record
Prereg R6, 2026-10-18 11:25, sha256 9027fef3…3afb · OpenTimestamps
round 6 result report (18 Oct 2026)
R6-K5 Trend-following fund flow (CTA flow) · Failed · 2026-10-18

R6-K5

Trend-following fund flow (CTA flow)

Failed · Rest in variance.

Hypothesis
Large inflows or outflows at trend-following funds briefly pull futures in the same direction.
Data
Minute data, development and validation through 2019, holdout after that, opened once (18 Oct 2026)
Result
Holdout n 483, −0.004 R (p 0.57); cost stress negative.
Lesson
No gain in development, validation or holdout.
Cause
No effect even before costs
Record
Prereg R6, 2026-10-18 11:25, sha256 9027fef3…3afb · OpenTimestamps
round 6 result report (18 Oct 2026)
Round 7 · 4 cards · 0 passed
R7-F1 Closing-pressure reversal (12 markets, at most 3 trades per day) · Failed · 2026-10-18

R7-F1

Closing-pressure reversal (12 markets, at most 3 trades per day)

Failed · Rest in variance.

Hypothesis
Extreme closing pressure reverses by the next close; at most three signals per day across twelve markets.
Data
Development, validation, holdout opened once (18 Oct 2026)
Result
Holdout n 2,293, net +0.004 R (p 0.389), gross +0.021 R; at 1.5× costs −0.005 R; validation −0.017 R.
Lesson
The gross remainder is smaller than costs; close-to-close reversal is not tradable.
Cause
Costs larger than the gross effect
Record
Prereg R7, 2026-10-18 18:13, sha256 5ebaa20c…4ea3 · OpenTimestamps
round 7 result report (18 Oct 2026)
R7-F2 Positioning extreme from the weekly trader positioning report (17 markets, Mon–Fri) · Failed · 2026-10-18

R7-F2

Positioning extreme from the weekly trader positioning report (17 markets, Mon–Fri)

Failed · Rest in variance.

Hypothesis
Extreme net positions in the weekly report revert over the week; trading days Mon–Fri alone.
Data
Development, validation, holdout opened once (18 Oct 2026)
Result
Holdout n 1,477, net −0.001 R (p 0.530); validation +0.004 R.
Lesson
The family was dropped on 14 Oct; the Mon–Fri version confirms it. Do not revive.
Cause
Price pattern without a mechanism
Record
Prereg R7, 2026-10-18 18:13, sha256 5ebaa20c…4ea3 · OpenTimestamps
round 7 result report (18 Oct 2026)
R7-F3 Pair reversal (measurement, no error budget) · Measurement, negative · 2026-10-18

R7-F3

Pair reversal (measurement, no error budget)

Measurement, negative

Hypothesis
Short-term reversal between two correlated futures; measurement without a family error rate.
Data
Development, validation, holdout opened once (18 Oct 2026)
Result
Development −0.200 R, validation −0.212 R, holdout −0.082 R (n 679).
Lesson
The cost of two legs (≈ 0.1–0.19 R) overwhelms any signal.
Cause
No effect even before costs
Record
Prereg R7, 2026-10-18 18:13, sha256 5ebaa20c…4ea3 · OpenTimestamps
round 7 result report (18 Oct 2026)
R7-F4 Monthly seasonality (measurement, no error budget) · Measurement without verdict · 2026-10-18

R7-F4

Monthly seasonality (measurement, no error budget)

Measurement without verdict

Hypothesis
Calendar patterns per market and month; measurement without a family error rate.
Data
Development, validation, holdout opened once (18 Oct 2026)
Result
Holdout +0.085 R per market-month, n 832, 75 months, p 0.031. Under the null, 1 of 22 random seeds fell below 5 %, mean bias +0.013 R; 379 exits are expiry roll approximations.
Lesson
No finding. Holdout is spent; continue via a new prereg with a pure forward test.
Cause
Too few events for a test
Record
Prereg R7, 2026-10-18 18:13, sha256 5ebaa20c…4ea3 · OpenTimestamps
round 7 result report (18 Oct 2026)

Gate period (rule and code hash before the run, before the log existed)

G series (gate checks) · 17 cards · 0 passed
G-18A Multi-day mean reversion · Failed · 2026-10-11

G-18A

Multi-day mean reversion

Failed · Rest in variance.

Hypothesis
After short-term weakness, indices revert (four classic signals, holding period days).
Data
four index futures, development 2010–2019; cross-check on 22 markets
Result
0 of 4 main lines above the bar (best t 1.68). On 22 markets all negative.
Lesson
Gross, the reversion was there; net and across markets it was not. It lived on index drift.
Cause
Costs larger than the gross effect
Record
rule and code hash before the run, 2026-10-11, sha256 dd6106f0…8d3a
Test no. 18 A (11 Oct 2026)
G-18A-ID Mean reversion as an intraday leg · Failed · 2026-10-11

G-18A-ID

Mean reversion as an intraday leg

Failed · Rest in variance.

Hypothesis
The same signals, traded the next day 09:31–15:50.
Data
as G-18A
Result
Best Sharpe ratio 0.30 (t 1.00); negative across all markets.
Lesson
What does not hold over several days holds even less as an intraday leg.
Cause
Costs larger than the gross effect
Record
rule and code hash before the run, 2026-10-11, sha256 dd6106f0…8d3a
Test no. 18 A (11 Oct 2026)
G-14 Volatility premium with index options · Failed · 2026-10-11

G-14

Volatility premium with index options

Failed · Rest in variance.

Hypothesis
Sellers of put spreads and iron condors (7/30 days) collect a premium above realized volatility.
Data
Index options, development 2016–2020, two runs (after a data repair)
Result
0 of 32 variants above the bar. Best line: 46 trades, 91 % hits, 0.4 % a year; with double costs Sharpe 0.08.
Lesson
91 % hits and one crisis month that eats years of premium: a high hit rate is not an edge.
Cause
Costs larger than the gross effect
Record
rule and code hash before the run, 2026-10-11, sha256 64b87c08… (Auszug / excerpt)
Tests no. 14 (11 Oct) and 17 (12 Oct 2026)
G-23 Treasury auctions intraday · Failed · 2026-10-12

G-23

Treasury auctions intraday

Failed · Rest in variance.

Hypothesis
Dealers push the price down before the auction (short before) and cover afterwards (long after).
Data
1,288 auctions 2010–2026 (official calendar), minute data, development 2010–2019
Result
Long after: n 631, mean −0.103 R. Short before: n 631, mean −0.207 R. Control on non-auction days even worse.
Lesson
An effect relative to a control is not yet a return; in absolute terms it has to cover costs.
Cause
Costs larger than the gross effect
Record
rule and code hash before the run, 2026-10-12, sha256 d858718d…ac3f
Test no. 23 (12 Oct 2026)
G-28 Base stream in the European session · Failed · 2026-10-12

G-28

Base stream in the European session

Failed · Rest in variance.

Hypothesis
The base stream's method also works 00:00–09:29 New York, on 6 or 10 markets.
Data
Minute data, development 2013–2019
Result
6 markets: n 550, mean −0.023 R, 3 of 7 years positive. 10 markets: n 828, mean +0.091 R, t 1.81, 4 of 7 years.
Lesson
A model trained on the US session is a different model in the European session.
Cause
Different in development and validation
Record
rule and code hash before the run, 2026-10-12, sha256 0fb1d1b8…1aec
Test no. 28 (12/13 Oct 2026)
G-29 Base stream at the edges of the session · Failed · 2026-10-12

G-29

Base stream at the edges of the session

Failed · Rest in variance.

Hypothesis
The opening and last minutes of the day hold extra opportunities.
Data
six markets, minute data, development 2013–2019
Result
n 577, mean +0.004 R, t 0.06, 2 of 7 years positive. Random selection −0.180 R.
Lesson
The model picks better than chance but earns nothing there after costs.
Cause
Costs larger than the gross effect
Record
rule and code hash before the run, 2026-10-12, sha256 be967fab…1693
Test no. 29 (12/13 Oct 2026)
G-30 Holding the base stream longer · Failed · 2026-10-12

G-30

Holding the base stream longer

Failed · Rest in variance.

Hypothesis
The profit target is hit more often if the trade stays open past the time exit (up to 3 days).
Data
Base stream development 518 / validation 261 trades, minute data incl. overnight
Result
All four variants: difference −0.030 to +0.025 R, no t above 0.5. Worst months 0.15–0.29 R worse.
Lesson
The time exit is part of the edge; overnight gaps eat what longer holding brings.
Cause
Not better than the base, paired
Record
rule and code hash before the run, 2026-10-12, sha256 fe974a32…e80f
Test no. 30 (12/13 Oct 2026)
G-32 Macro surprise · Failed · 2026-10-13

G-32

Macro surprise

Failed · Rest in variance.

Hypothesis
The standardized surprise (actual − consensus) of 10 US indicators continues (or reverts) in the minutes and hours after.
Data
2,081 events, 9,427 event × market, development 2014–2019
Result
Main line n 371, mean −0.090 R; all 6 variants negative (−0.09 to −0.21 R).
Lesson
The opening minute after the release cannot be traded by a minute-bar system; what follows is already priced in.
Cause
Price pattern without a mechanism
Record
rule and code hash before the run, 2026-10-13, sha256 94d8c248…ff3c
Test no. 32 (13 Oct 2026); calendar hash 1fcaaae4…526d
G-36 Base stream without daily and weekly cap · Failed · 2026-10-13

G-36

Base stream without daily and weekly cap

Failed · Rest in variance.

Hypothesis
The caps per day and week cost opportunities; the extra stream has its own edge.
Data
six markets, development 2013–2019, validation 2020–2022, final test 2023–09.2026 once
Result
Extra stream total: development n 274, +0.202 R; validation n 196, +0.199 R; final test n 259, mean +0.00008 R (t 0.001).
Lesson
The day's second trade points the same way as trade one 81–98 % of the time; it doubles risk, not information.
Cause
Different in development and validation
Record
rule and code hash before the run, 2026-10-13, sha256 fd8b3f51…7838
Test no. 36/36F (13 Oct 2026; v1 hash 4e1f4b83…2373)
G-61-1 24 exit variants for the base stream · Failed · 2026-10-14

G-61-1

24 exit variants for the base stream

Failed · Rest in variance.

Hypothesis
Some combination of target, trailing stop, partial exit or time beats the existing exit profile.
Data
Base stream with real minute prices, development 647 / validation 328 trades; selection rule fixed in advance
Result
No variant meets the development condition. Significantly worse: early target −0.181 R, partial exits −0.12 to −0.14 R.
Lesson
Partial exits and early targets trade the right tail for a higher hit rate and lose the edge in the process.
Cause
Not better than the base, paired
Record
selection rule before the run, 2026-10-14
Test no. 61 part 1 (14 Oct 2026)
G-61-4 Volume shock: continuation or reversal · Failed · 2026-10-14

G-61-4

Volume shock: continuation or reversal

Failed · Rest in variance.

Hypothesis
A minute with ≥ 20× median volume and a large range signals an information shock that continues (W1) or reverts (W2).
Data
25 markets, minute data, development 2013–2019, validation 2020–2022
Result
W1 development n 116,752, mean −2.83 R; W2 n 105,851, mean −4.31 R. Median costs 2.5–4.0 R per trade.
Lesson
Relative triggers fire at night on tiny bars; without an absolute minimum size every volume rule mostly pays costs.
Cause
Costs larger than the gross effect
Record
rule before the run, 2026-10-14
Test no. 61 part 4 (14 Oct 2026); not repeated with other thresholds
G-22 Second- and third-best candidate of the day · Not confirmed in the final test · 2026-10-11

G-22

Second- and third-best candidate of the day

Not confirmed in the final test · Rest in variance.

Hypothesis
The second-best candidate (another market) carries its own edge and could be a second stream.
Data
six markets, development/validation, final test once
Result
Development n 194, +0.233 R (t 1.96); validation n 125, +0.289 R (t 2.04); final test n 182, +0.005 R (t 0.04).
Lesson
A finding that arose after looking at development and validation did not hold in the final test.
Cause
Idea formed after looking at the data
Record
not pre-registered, 2026-10-11
Tests no. 18 B and 22 (11/12 Oct 2026); not pre-registered
G-71-4 Order-flow signals · No finding · 2026-10-16

G-71-4

Order-flow signals

No finding · Rest in variance.

Hypothesis
The aggressor imbalance of the last 5/15 minutes predicts continuation.
Data
trade-and-quote data, 12 months (10.2025–10.2026), 190,280 candidates; exploration without prereg
Result
650 cells, 44 intervals excluding zero (≈ 32 expected by chance), none survives Bonferroni.
Lesson
Minute-level order flow is already in the price returns.
Cause
Price pattern without a mechanism
Record
exploration without prereg, 2026-10-16
Exploration no. 71 part 4 (16 Oct 2026)
G-02 More frequency: lower threshold, 18 markets · Failed · 2026-10-11

G-02 · U

More frequency: lower threshold, 18 markets

Failed · Rest in variance.

U · Numbers from an older list (11 Oct 2026), in no result report – not verifiable.

Hypothesis
More markets and a lower threshold give more trades at the same edge.
Data
6 and 18 markets, development/validation
Result
Failed according to the 11 Oct list; result lines in archives alone (U). Comparison from later tests: higher thresholds better gross per trade, but 3 to 10 times rarer.
Lesson
The edge sits in the top percent; lowering the threshold buys noise at the price of costs.
Cause
Costs larger than the gross effect
Record
rule and code hash before the run, 2026-10-11, sha256 fddb09a0… (Auszug / excerpt)
Test no. 02 (11 Oct 2026)
G-ML Combined model over all features · Failed · 2026-10-19

G-ML · U

Combined model over all features

Failed · Rest in variance.

U · Numbers from an older list (11 Oct 2026), in no result report – not verifiable.

Hypothesis
A learning model over all price-pattern features finds what single rules miss.
Data
U (no result report)
Result
Listed as “failed” on the 11 Oct list (U).
Lesson
More features without a mechanism raise the number of variants, not the edge.
Cause
Price pattern without a mechanism
Record
just a list survives
List of 11 Oct 2026
G-TSM Time-series momentum 1–12 months · Failed · 2026-10-19

G-TSM · U

Time-series momentum 1–12 months

Failed · Rest in variance.

U · Numbers from an older list (11 Oct 2026), in no result report – not verifiable.

Hypothesis
The return of the last 1–12 months continues.
Data
U
Result
Development failed, validation good (numbers U).
Lesson
An effect that is missing for ten years and present for three is not an effect; it is 2020–2022.
Cause
Different in development and validation
Record
just a list survives
List of 11 Oct 2026
G-KAL Carry, overnight long, turn of the month, day before the rate decision · Failed · 2026-10-19

G-KAL · U

Carry, overnight long, turn of the month, day before the rate decision

Failed · Rest in variance.

U · Numbers from an older list (11 Oct 2026), in no result report – not verifiable.

Hypothesis
Known calendar and premium effects from the literature.
Data
U
Result
Overnight long = index drift alone (excess over drift ≈ 0); other numbers U.
Lesson
What the literature shows gross over decades rarely held for us with 1 tick slippage per side over 2010–2019.
Cause
Costs larger than the gross effect
Record
just a list survives
List of 11 Oct 2026

Early period (a list alone survives)

V-21 · 21 patterns · 0 passed
V-21 21 intraday price patterns · Failed · 2026-10-19

V-21 · U

21 intraday price patterns

Failed · Rest in variance.

U · Numbers from an older list (11 Oct 2026), in no result report – not verifiable.

Hypothesis
Common chart patterns (opening-range breakout, pullbacks to averages, gaps, prior-day high/low, structure breaks, last-half-hour momentum and others) hold intraday after costs.
Data
6 markets, minute data, development 2010–2019 (U)
Result
All 21 below the bar (Sharpe ≥ 0.3, t ≥ 2, excess over drift); single numbers U.
Lesson
In our tests (6 markets, 2010–2019, after costs) none of these patterns cleared the bar. That says nothing about how others trade them. Since then we take families with a mechanism: who has to trade, and when.
Cause
Price pattern without a mechanism
Record
just a list survives
List of 11 Oct 2026 (early tests, written down per family in advance)

Levers we tested that are not strategies

Levers · 3 checked · 0 passed
H-FREQ “No edge, but positive through frequency” · Failed · 2026-10-16

H-FREQ

“No edge, but positive through frequency”

Failed · Rest in variance.

Hypothesis
An evaluation account is positive on average even without an edge at 3–4.5 trades a week (+1.4 to +3.7 k$).
Data
real trade sequences, real trading days, 10,000 paths, seed 61; forecast in advance
Result
The no-edge mean falls with frequency (+0.8 → +0.3 k$); the median is negative in every no-edge cell (−0.8 to −1.1 k$).
Lesson
Independent days and instant payouts create money from the model, not from the market.
Cause
Gain from the model, not the market
Record
forecast before the computation, 2026-10-16
Simulation window 2 (17 Oct 2026 00:25), forecast of 16 Oct 2026
H-1S Better entry in the opening second · Failed · 2026-10-16

H-1S

Better entry in the opening second

Failed · Rest in variance.

Hypothesis
A market order one second after the minute starts fills better than the minute's open.
Data
second data, three index futures, 2010–2026, n 581
Result
−1.8 ticks = −0.013 R worse; 28 % of fills better; after 20 seconds ≈ 0.
Lesson
A few unusually good fills are selection: the trades that run away at once are missing from memory.
Cause
Gain from the model, not the market
Record
measurement inside a prereg, 2026-10-16, sha256 6b229f5d…d267
Round 3, measurement M1 (16 Oct 2026)
H-SIZE Smaller size as the buffer shrinks · Failed · 2026-10-19

H-SIZE · U

Smaller size as the buffer shrinks

Failed · Rest in variance.

U · Numbers from an older list (11 Oct 2026), in no result report – not verifiable.

Hypothesis
Size is capped once the distance to the loss limit shrinks (40 % of the distance).
Data
U
Result
Expected value per attempt +86 instead of +353 $ (U).
Lesson
Fixed size stays.
Cause
Gain from the model, not the market
Record
just a list survives
List of 11 Oct 2026

Not in the graveyard

So you do not look for them here: what is running, what is open, what was not testable, and the one stream that is alive.

For the technically minded: open families
FamilyStatusReason / next step
R1-F1 · Month-end reversal (index, bond)Running forwardrunning forward; next signal Fri 23 Oct 2026 at the close
R1-F2 · Front-running the index roll (oil)Running forwardwaiting for the roll window 6–12 Nov 2026
R1-F4A · Earnings cluster of the largest tech stocks (Nasdaq future overnight)Undecidedholdout n 5, t 4.41 under the fixed reading; not robust with 5 cases, forward test alone (next clusters Jan/Apr 2027)
R2-F2 · Base stream on bitcoin futuresNot testablethe model does not know the market; any mapping would be fitting
R2-F3 · Overnight Nasdaq after a sell-off, wrapped in optionsNot testableoption data for 88 or 27 of 436 signals; revisit after a data pull (also round 3, F1)
R4-A3 · Round 4, A3Not computed yetfrozen, run after a data pull (by 5 Nov 2026), own holdout
R5-GAM · Round 5, gamma conditionNot computed yetwaiting for a recomputed gamma series (hash into the log before the run)
BASIS · Base stream (on probation, edge not shown)Alive, on probationholdout n 348, +0.138 R per trade, t 1.65, not significant; last 12 months −0.05 R (n 72). Checkpoint 28 Feb 2027.
H-MM · Mini + micro instead of mini alone (costs)Passed as a cost lever+0.023 R per trade from lower costs; paper test before use

What the cards have in common

  1. Effect before costs: in the review of all 43 cards (as of 19 Oct), 17 showed no effect even before costs; for 7, costs were the cause. Every intraday futures card with a gross figure was below zero before costs. Every new family computes before the run what gross advantage it needs and whether its holdout can detect it at all.
  2. Paired against the base: filters on a positive strategy look good in absolute terms; the paired comparison counts. None passed it.
  3. The holdout is untouched exactly once: we open it a single time, with a log line before.
  4. Mechanism over pattern: the open candidates all have a constraint (who has to trade).
  5. Power in advance: with fewer than 60 events in the holdout, “not testable” is more honest than “failed”.

Source: Review of all 43 cards, as of 19 Oct 2026.

Registrations (full hashes)

Each round was fixed before any run was computed. The hash is the fingerprint of the text; anyone who gets the text later can recompute it.

For the technically minded: registrations (full hashes)
RoundFixedsha256
R12026-10-15 12:17708e99ca18515862b927783a6f31c93351f6ae640bf0b1ccafe1ea9abe5e0480
hash over the text without the hash section
R22026-10-16 06:50240d7e98069f390b95fc427ee1c96ea99fffdbbc035e3b280670305e257a16dd
R32026-10-16 19:406b229f5d0917c6d478bce27cef3b9da2d01052dd363dc602c11fad0dd53cd267
R42026-10-17 09:15 · OpenTimestampsc533abac57fbf7c52c165980b09663fdd9fb1f1972f0cdb6e50a2bd640b83133
R52026-10-17 11:50 · OpenTimestamps531a2b3343b92c5e3d960784a365b2e7b559bb577d49b190713fc6a82f68f733
R62026-10-18 11:25 · OpenTimestamps9027fef3758d1deff3103d2833949d6ce609b5a823ef884b674326177add3afb
R72026-10-18 18:13 · OpenTimestamps5ebaa20c7a65b17ded163664ef7bf3cf90070242c9138947b62e69b842b34ea3

The cards as a file (for your own analysis): friedhof.json.

So far the marked pre-registrations carry a timestamp (OpenTimestamps); register entries get theirs at launch. The counter never resets. Earliest entry: 2026-10-15.

Entry #1 · coming

Our own trading system

Our own prop account will be register entry #1 once it is running. Before trading begins we will file a timestamped forecast; we publish results once enough trades exist.

Opt-in. What you choose goes public — never your trades.

Your registrations (this browser)

None yet.

    Corrections log

    The counter never goes back. If an entry turns out wrong, the correction goes here — with date and reason.

    No corrections so far.

    Check with your own trades: beta from 23 Oct → waiting list · input format → data formats

    Back to top