Method: Zero-Skill Baseline (ZSB-1)
Key points
- We compare you with traders with no edge, same rules, same risk, same costs.
- Rank 95 does not mean 95 % skill.
- Every number has source, version, and date.
- Our own strategies: 24 checked, 0 passed – that does not mean no edge exists.
What this page can do: explain how we compute, and back every number we publish anywhere with source, version and date. What it cannot do: tell you whether you have skill. ZSB-1 tells you how a trader with no edge, using your risk, your rules, your trade shape and your costs, would have fared, and where your record sits among 10,000 such simulated no-edge traders. A good rank is consistent with skill and with luck. A bad rank is consistent with no skill and with costs.
Methods paper: ZSB-1 v1.0: Zero-Skill Baseline (English, German one-page summary in the appendix). DOI and timestamp follow with publication. Each method version gets its own DOI; results always cite the version that produced them. Full text: methods paper ZSB-1; German summary: Kurzfassung.
How we test our own strategies: How we test (6 steps). What failed along the way: register, verdicts section.
1. The 15 numbers we use
Each number holds for one rule set, one trade shape and one method version. None is a statement about any firm, about real traders or about you. Historical scenario numbers are not forecasts.
pass with no edge, $1,000 risk per trade, 1:2 (variant)
smallest edge at 87 trades (80 %, σ 1.4 R, α 0.005)
ranked ≥ 95 have skill (50 trades, 5 % with 0.10 R, 1:2)
passed · 19 failed, rest open
For the technically minded: all 15 numbers with source
| No | Number | Holds for | Source | As of |
|---|---|---|---|---|
| 1 | Head math without costs: static floor 40 %, trailing 22 %, trailing with lock 24.5 % | target 12 R, max loss 8 R, continuous limit; formulas D/(D+A), e^(−A/D), e^(−1)·D/A | Methods paper §3 (Taylor 1975, Lehoczky 1977); test vectors T1–T5 | 17 Oct 2026 |
| 2 | A soft daily loss limit barely changes the chance-pass rate (27.5 → 27.4 %); it mostly costs time | base case of the methods paper, static floor | Methods paper §3; run L01, seed 7, 20,000 paths | 17 Oct 2026 |
| 3 | Larger risk per trade raises the chance-pass rate: 5.6 % at $125 per trade, 29.3 % at $1,000 (fewer trades, fewer costs). A warning, not advice | base case of the methods paper | Methods paper §3; run L01, seed 7, 20,000 paths | 17 Oct 2026 |
| 4 | Detection limit: with 87 trades, an edge is detected with 80 % probability from 0.55 R per trade upward (two-sided α 0.005); at 150 trades 0.42 R | spread 1.4 R; formula 3.65·σ/√n (α 0.005 two-sided, 80 % power) | Methods paper §1 (card A2) | 19 Oct 2026 |
| 5 | Countdown (preliminary): an edge of 0.10 R is detected with 80 % probability from 1,332 independent trades; at σ = 1.5 R: 2,996. Never „skill shown“ | σ = 1 R, edge +0.10 R, α 0.005 two-sided, 80 % power | content/zahlen/tape-core-0.2.0_sichtgrenze.json, n_sigma1.5_vorteil0.10 | 19 Oct 2026 |
| 6 | Whoever looks at the 95 % band after every tenth trade gets a false alarm in 33 of 100 cases by n = 600 instead of 5 | 60 looks (from n = 10), no edge | Methods paper §5.3 | 17 Oct 2026 |
| 7 | A naive 80 % band from one's own history hits 58 of 100 at n = 30; the zero-skill forecast beats one's own history there (simulation under an assumed spread of edges, N(0.03, 0.08²) R; not a measurement of real traders) | bootstrap of the history, next 50 trades | Methods paper §9; Masterplan key finding 4 | 17 Oct 2026 |
| 8 | Best of 20: a result that beats 62 of 100 simulated no-edge traders beats the best of 20 simulated no-edge traders 0 times in 100 | Šidák correction | Methods paper §6, test vector TV-8 | 17 Oct 2026 |
| 9 | Rank 95 does not mean "95 % skill": at n = 50 the test has 11 % power (edge 0.10 R, 1:2 bracket, threshold rank 95); if 5 in 100 traders really have skill, about 11 of 100 of those with rank ≥ 95 do | Bayes computation with 5 % prior | Methods paper §4, misreading box | 17 Oct 2026 |
| 10 | 24 preregistered families checked (19 failed, 1 undecided, 2 not testable, 2 still running), 0 passed (our own tests); 43 cards lie in our graveyard, one of them sums up 21 older price-pattern families; one stream is on probation (holdout n 348, +0.14 R per trade, t 1.65, not significant; last 12 months −0.05 R at n 72). "0 passed" does not mean that no edge exists | own strategy research 2026, tests preregistered with hash | Register, verdicts, score | 19 Oct 2026 |
| 11 | Of 43 cards, 17 showed no effect even before costs (gross re-check, lesson 1); for 7, costs alone were the cause. Card labels count differently: each card shows its main cause alone | as no. 10, review of all 43 cards | Register, verdicts, lesson 1 | 19 Oct 2026 |
| 12 | "An effect absent for ten years and present for three is not an effect. It is 2020–2022." | lesson from card B16 (time-series momentum) | Register, verdicts, card G-TSM | 17 Oct 2026 |
| 13 | The second trade of the day points the same way as the opening trade 81–98 % of the time; it doubles risk, not information | own strategy, tested 2026 | Register, verdicts, card G-36 | 17 Oct 2026 |
| 14 | Without MAE/MFE, "at-close" evaluation overstates the chance-pass rate under a trailing floor by about a quarter (18.5 instead of 14.6 %) | base case of the methods paper, intraday trailing floor | Methods paper §7 | 17 Oct 2026 |
| 15 | More trades without an edge do not make an evaluation account positive: the median is negative in every no-edge cell | simulation, 50K-type rules, frequency × risk | own simulation, register, verdicts section, card H-FREQ | 17 Oct 2026 |
How to read this. "Base case of the methods paper" means the rule and cost parameters from methods paper §3. Numbers 1 to 9 and 14 follow from closed formulas or from the reference simulator with a seed; numbers 10 to 13 and 15 come from our own strategy research and are listed with source and hash in the register, verdicts section.
On no. 3: The larger the risk per trade, the more traders with no edge (simulated no-edge traders) pass (5.6 % at $125, 29.3 % at $1,000 per trade). A pass achieved that way says even less about skill.
What we do not say. Not "better than X % of traders" (correct: "better than X of 100 simulated no-edge traders with your risk"). Not „skill shown“. No blanket rate such as "17–25 % pass": the chance-pass rate holds for its rule set and differs strongly between rule sets. No firm names. No probability of skill.
2. The zero-skill baseline
Your trade values and your day structure stay. We reweight your trading days so that a trader with no edge makes gross zero on average (exponential tilting of day sums), draw 10,000 simulated no-edge traders as whole days, and subtract costs per trade.
For the technically minded
For stop/target brackets this is exact (at 1:2 the simulated no-edge trader wins exactly 1 in 3); for time and trailing exits it is an approximation. Limit: days are drawn independently; if they depend on each other (calm and wild phases), the rank tends to be too optimistic. The band corrects for this (VIF, block bootstrap), the rank not yet. Planned: the effective number of days goes on the card, and below 20 it will say "rank uncertain"; the beta does not show this yet.
What we do not use as the null: sign permutation, plain shuffling, random entries on price data. Reasons are in methods paper §2.
3. Three formulas you can check by head
Two consequences that surprise: a soft daily loss limit barely changes the rate, mostly the duration. Larger risk per trade raises the rate because fewer trades mean fewer costs. We show the latter as a warning, never as advice.
For the technically minded
Continuous limit, no costs, target A above the start, max loss D below:
| Rule | Pass probability with no edge | Example A = 12 R, D = 8 R |
|---|---|---|
| Static floor | D ÷ (D + A) | 8 ÷ 20 = 0.40 |
| Intraday trailing floor | e^(−A/D) | e^(−1.5) = 0.22 |
| Trailing floor locked at the starting balance | e^(−1) · D/A | 0.368 · 8/12 = 0.245 (24.5 %) |
The static formula does not depend on the strategy at all (the classical ruin problem with fair steps, Feller 1968).
4. Rank and the misreading
Rank 95 means 5 of 100 simulated no-edge traders were as good as you. It does not mean you have skill with 95 % probability.
For the technically minded
"You beat x of 100 simulated no-edge traders with your risk, your rules and your costs." The p-value follows Phipson–Smyth, is never zero, and carries a Monte-Carlo error. At 50 trades and a 5 % share of skilled traders, about 11 of 100 of those with rank ≥ 95 really have skill (no. 9).
5. Band, countdown, checkpoints
The countdown uses your own spread and the smallest edge you care about (default 0.10 R). Preliminary: the countdown in the calculator already uses the strict rule (α 0.005 two-sided, as for the detection limit).
For the technically minded
The edge band is an estimate with a skeptical prior (τ₀ = 0.25 R); the card prints how much of the estimate comes from your trades, next to the bootstrap band over whole days. It appears from 150 trades; below 150 trades the band cannot be evaluated. The fixed checkpoints are still preliminary. Note: the verdict levels still use α 5 % one-sided; they will be moved to the strict rule. Whoever looks continuously gets 33 % false alarms (no. 6); so verdicts are given at ten fixed checkpoints (O'Brien–Fleming), and in between the band is labelled "preview".
Four verdict levels with fixed thresholds: "Not yet decidable", "Looks like chance", "A signal is emerging", "Hard to explain by luck alone" (always with "This is not proof of skill"). Levels can go down. Thresholds per checkpoint are in methods paper §5.4.
6. Variants
Mandatory question: "How many fundamentally different setups have you tried?" We then compare you with the best of N simulated no-edge traders (Šidák).
For the technically minded
Rank 62 at N = 1 becomes 9 at N = 5 (rank against the best of 5 simulated no-edge traders, not the multi-account figure) and 0 at N = 20 (no. 8; synthetic example stream, 1:2 bracket, net mean −0.02 R, n = 300, seed 80130001; null tilting with 20,000 simulated no-edge traders per step, seeds 80130003–80130005, tape-core 0.2.0; p = 0.382 at N = 1).
7. Prop path model
With MAE and MFE we bound the path. Exact computation for stop/target brackets; for time and trailing exits we state the assumption and show a span.
For the technically minded
Whoever has close prices per trade alone overstates the chance-pass rate under a trailing floor by about a quarter in the base case (no. 14); without MAE/MFE we therefore show a range and mark the point value as optimistic. Not yet in the model: consistency rules (no day above a fixed share of the target) and minimum sizes for winning days. Both lower the chance-pass rate; our number is then closer to an upper bound.
Several accounts (model example, pure arithmetic, not a measurement). Assumption: assumed pass rate per account 28 %. Five identical (copied) accounts pass or fail together: at least one passes with 28 %. Five independent accounts: 1 − 0.72⁵ = 80.65 % ≈ 81 %. Cost per pass is the same for copied accounts; the spread changes: all or nothing. Three copied accounts are not three independent samples.
8. Multipliers
Contract values live in a versioned table with one golden test per contract (for example $2 per point for the smallest Nasdaq contract, $5 for the smallest S&P contract).
For the technically minded
Deviations from exchange definitions fail the test before any number is shown.
9. Calibration: we publish how often we were wrong
Every forecast is stamped (hash, version, seed, τ, band) and later scored against the zero-skill forecast (Kupiec coverage test, PIT histogram, CRPS skill score). Our own stream to start is preregistered.
For the technically minded
User calibration appears from 137 independent windows (what it takes to detect a drop from 80 to 70 % coverage with 80 % power, α 5 % two-sided), fully from 300. Until then this page reads: "So far the operator stream (preregistered). Status: n windows."
Return bias: losers report back less often. We weight by the return probability and print Manski bounds next to it. At a return rate of just under one third (example from calibration run calibration.py, run ID calibration_mar_seed11, seed 11, case "return depends on recorded features alone"; not the multi-account figure) the bounds are so wide that automatic data sources are needed for a statement. That is why every stamp counts in the denominator.
10. Validation
Five golden tests against closed formulas (T1 0.40 · T2 0.814 · T3 0.2252 · T4 0.2468 · T5 96), one deterministic test vector for day tilting, one test of the type I error (target ≤ 1.2 % at level 1 %; measured up to 1.24 %, slightly above), three coverage tests for τ. The suite runs on every change; if a test deviates by more than 3 standard errors, nothing is published. Test vectors: E_testvektoren.json, SHA-256 ef0ff76cf29f0f069e078651766d97865edf7f3a3de609a97e631aad9f552b3f.
Recomputable soon. Calculation method, code version and seed are on every proof receipt. With the receipt and your trade file any number can be recomputed once the engine for the zero-skill baseline is published open source (MIT licence); the multi-account computation and internal data will not.
How to verify the checksums
The shipped HTML pages have a checksum list:
curl -O https://irrationaltape.com/SHA256SUMS.txt
shasum -a 256 -c SHA256SUMS.txt
Windows (PowerShell): certutil -hashfile <file> SHA256 and compare the hash to the line in SHA256SUMS.txt. That tells you whether a downloaded file is unchanged — nothing more.
11. Error changelog
Errors we find are listed here with their size.
For the technically minded
| Version | Date | Change |
|---|---|---|
| 0.1 | 16 Oct 2026 | Initial calibration draft: tilting per trade, τ = 0.10 R. |
| 1.0 | 17 Oct 2026 | Day tilting with whole days and an effective-days flag; τ₀ = 0.25 R with coverage tests; fixed verdict levels at O'Brien–Fleming checkpoints; card texts changed (new term "simulated no-edge traders with your risk"; "detected with 80 % probability" instead of "visible from"); at-close bias per user instead of a constant; "probability of edge > 0" removed from cards. |
One example from development: a wrong boundary condition at the floor shifted the expected number of trades to a decision by two trades; golden tests have been mandatory since.
12. What we do not do
No "probability of skill". No sign permutation, no plain shuffling, no random entries on price data in the product.
For the technically minded
No plug-in bands. No blanket rate. No firm names and no firm presets. No price series or candle charts from licensed market data in public outputs; publicly we show R, rates and distributions. No claim to have coined "zero-skill baseline"; the term stands in the tradition of skill scores in forecast verification.
13. What we do not confuse
- Three copied accounts are not three independent samples.
- Historical scenario numbers are not forecasts.
For the technically minded
- The account size of an evaluation is not your own capital.
- Gross payout is not profit after fees, resets and taxes.
Check with your own trades: beta from 23 Oct → waiting list · input format → data formats