28 rules tested on history and live. We list the ones that failed as carefully as the ones that did not.
No single rule passed (t ≥ 2.58 on the test). Trend families as a whole were positive in all three periods; mean reversion, funding and seasonality were consistently negative.
Alpha to the basket +54%/yr (t=2.4), +33% (t=1.8), +18% (t=1.3): positive but shrinking. Drawdowns 25–55%.
Sharpe 0.74 / 0.91 / 1.00 versus 0.66 / 0.64 / 0.73 for buy & hold; positive in both halves; drawdowns −68 / −60 / −67% versus −79 / −80 / −95%. On a 12-coin portfolio the Sharpe gain has a 90% interval that includes 0 — the benefit is mainly a smaller drawdown.
+9.8 $/day per 1 MNQ, t=4.7; 8 of 11 years positive; two thirds of the profit came from 2018, 2022, 2023. Drawdown at 1% risk up to about −46%. Rules published by other authors; our check is independent.
+0.48R, t=5.2 on history (inflated by the search over 74 conditions), 28% winners. Same rule on GER40 at 1:8 is negative. A verdict needs about 50 forward trades.
+0.07R ±0.07 overall (longs +0.15R, shorts −0.01R); BTC + ETH alone −0.02R. Likely survivors and a rising market.
+0.12R ±0.18, but the whole gain is in 2024–2025; 2021–2023 about −3R, 2026 −4.8R.
+0.142R, t=2.2; train +0.17R, test +0.11R. A subgroup picked from about ten, so not a proof.
−0.11R per trade after the 0.08% fee (sum −840R); +0.05R before fees. No stop/target pair was positive.
None held on the test part.
−0.07R ±0.06; 3 of 13 months positive.
−0.06R ±0.08. A +0.06R result on one year and 10 coins did not repeat on the long history.
−0.09…−0.10R; both halves of the year negative.
−0.35…−0.55R; 0 of 13 months positive.
−0.03R ±0.01 before fees, −0.26R after the 0.08% fee; 0 of 13 months positive.
Positive in selection: 28 of 831 (15m), 12 of 783 (1h). On the test: none with a lower bound above 0; the best 12 gave −0.14R and −0.17R.
−0.38R per trade after fees (−0.02R before fees); 5m −0.46R, 15m −0.22R, 1h −0.19R. Snapshot of 2026-11-06; the live figures are on /crypto/levels.
Mean ≈ 0 (±0.06R), no better than a random entry with the same risk geometry; the structure score does not help.
Classic −0.085R ±0.038; pivot variant −0.119R; with Storyline −0.215R; with a strict higher-timeframe filter −0.259R.
−0.26…−0.53R after fees; about 0…−0.13R before fees.
No edge in any of 6 groups. Without fees ≈ 0 (−0.06…+0.04R); the 2R target is hit ≈33% of the time, as in a random walk.
15 candidates on train, 0 positive on test, 0 survived.
0 of 21 confirmed. The only factor is the stop width versus the fee.
The dependence is non-monotonic and changes sign from train to test; “skip when a wall is against” does not help.
The 15-minute move after an alert is 1.02× the coin’s baseline: not better than a random moment. Order-book imbalance and large-trade bubbles: no edge after fees (short samples).
The best on BTC (Clenow, Faber) did not carry over to ETH and SOL (Clenow −77% / −88% in the second period). Turtles and Carver are below the benchmark on the portfolio; fast rules showed no stable edge.
None confirmed (test t ≥ 2.8 required). On gold the dependence exists but costs exceed it.
Momentum: Sharpe around 0. London noise bands −0.42R/day, Asian breakout −0.05…−0.10R/day; 0 of 9 pairs positive in both periods.
Most tests use the current top coins, so surviving coins flatter long-only rules. Several rules were chosen after searching many variants; the search size is stated where known. The registry is updated when a new study or a forward journal reaches its pre-set sample. Methodology: Methodology.
This is a single list of the trading rules we tested on history and forward: the Quasimodo pattern and its filters, level retest, Elliott waves, MSNR levels, fair value gap entries, footprint and order book features, radar alerts, trend rules, noise bands on Nasdaq futures and forex ideas. Every entry states the market, the period, the sample size, the result in R after fees and a verdict.
R is profit as a multiple of the initial risk: +1R means the trade earned as much as the stop could have lost. We look at the mean R per trade rather than at single good streaks, and always next to a confidence interval: a result inside the “±” is indistinguishable from luck.
If only good results are shown, the reader never learns how many variants were tried. Out of 132 rules in the broad search none passed the threshold on the test part, and out of more than two thousand cells in the gap search none survived. Those numbers matter more than any single chart of winning trades, so they sit in the registry next to everything else.
“Refuted” means no edge after fees, or the edge did not hold on the test part. “Unproven” means a positive or mixed result that is inside the noise, on surviving coins or on few trades. “Candidate” means positive in more than one period and worth a forward test on data that does not exist yet. A candidate does not mean the rule works: many such results are explained by a rising market and by a sample made only of coins that survived until today.
Rules are written down before the results are seen; selection uses the first part of history and the check uses the last part. Live forward journals collect a pre-set sample, and only then do we revisit a verdict. Nothing in the registry is a promise of profit or advice to trade.
There is no confirmed profitable strategy in the registry. There are candidates: slow trend rules on daily and four-hour crypto candles and noise bands on Nasdaq futures. They are positive on history but were tested on surviving instruments and await a forward test.
Because the number of variants tried is needed to judge any positive result. Showing only the successes is exactly what overfitting looks like.
R is profit as a multiple of the trade’s initial risk. The fee is subtracted as a share of the stop size, so it weighs most on short stops; it is what turns many near-zero results negative.
Every entry names a report file in the project repository. The scripts and study data live there, and the Backtest page lets you test a simple rule of your own against a random-entry control.