AI lab — what survives verification, and what we threw out
Every dashboard shows you the same data. We do the opposite: we test a trigger on the full 2017–2026-09 funding history and publish the verdict, including the ones that failed. Each indicator below carries a verification passport: the test period, the sample, the correlation split by market regime, the conditional forward distribution, and an honest caveat. We report descriptive facts: correlation, lead, trigger frequency, and never a return promise.
Everything that survived this page is used the same way every morning: the indicators fuse into one dated verdict, and the playbook says what each state means for a position.
Laboratory reports
Full studies: the whole rule, its controls, both halves of the window and a passport a bot could trade from. Each has its own page because the conclusion there is longer than one row.
Every algorithm on one row
Full period, one engine, one archive, one cost model. Green is the bull side, long or in cash. Red is the bear side: short, or a filter that only acts while the gate is shut. Controls are marked: they exist to show what a result looks like when the selection is removed, and on this page they have repeatedly matched it.
| Algorithm | Side | Net | CAGR | maxDD | Sharpe | What it is |
|---|---|---|---|---|---|---|
| [TT] HyperGrass v8LONG — long, cash in risk-off | bull | +1607.1% | +53.5% | -41.0% | 1.30 | Regime gate, not stock picking. A random board of the same size scores inside one standard deviation of it. |
| SMA regime gate (BTC+ETH) | bull | +1523.0% | +36.3% | -55.7% | 0.90 | Cuts drawdown, does not add return. Loses to holding in every rising year and wins in every falling one. |
| Hold the basket | bullcontrol | +562.8% | +26.0% | -73.9% | 0.69 | The thing every row above has to beat. |
| [TT] HyperGrass v8LONG — with the bear leg | bear | +1790.8% | +55.9% | -41.0% | 1.33 | The same, plus a small short book while the gate is shut. Improves every number; the funding threshold controls size, not selection. |
| Funding-Heat Short Fade, gated | bear | +5.7% | +0.6% | -34.5% | 0.11 | Flat over the full archive. Its entire positive contribution is 2025-2026. |
| Short everything while risk-off | bearcontrol | -55.9% | -8.7% | -74.5% | -0.11 | No selection at all — short whatever is liquid while risk-off. In 2026 it beat the funding selection. |
| The same, ungated | bearcontrol | -90.6% | -23.1% | -94.9% | -0.51 | The same construction without the regime gate. This is what the gate is for. |
Bull side and bear side, on one measuring stick
Everything on this page that has a tradeable rule, run through the same engine on the same archive with the same costs, and split by which market it is meant for. Net %, the worst drawdown inside that year, and Sharpe from that year's daily returns. Controls are marked: a bear-side result without its “short everything” control has repeatedly turned out to be beta rather than selection.
Bull side — long, or in cash
| Year | [TT] HyperGrass v8LONG — long, cash in risk-off | SMA regime gate (BTC+ETH) | Hold the basket control | ||||||
|---|---|---|---|---|---|---|---|---|---|
| Net | maxDD | Sharpe | Net | maxDD | Sharpe | Net | maxDD | Sharpe | |
| 2017 | — | — | — | +0.0% | 0.0% | 0.00 | — | — | — |
| 2018 | — | — | — | -43.2% | -52.3% | -1.05 | -60.8% | -71.3% | -1.50 |
| 2019 | +0.0% | 0.0% | 0.00 | +72.7% | -36.3% | 1.31 | +9.6% | -62.9% | 0.48 |
| 2020 | -8.9% | -41.0% | 0.04 | +177.1% | -47.5% | 1.96 | +201.5% | -62.9% | 1.76 |
| 2021 | +268.6% | -16.8% | 2.56 | +160.6% | -40.9% | 1.64 | +399.1% | -58.4% | 2.06 |
| 2022 | +0.0% | 0.0% | 0.00 | -20.2% | -20.2% | -1.64 | -58.7% | -63.4% | -0.88 |
| 2023 | +141.1% | -21.0% | 2.15 | +77.6% | -18.7% | 1.66 | +56.3% | -33.0% | 1.29 |
| 2024 | +80.6% | -29.8% | 1.63 | +64.7% | -28.7% | 1.33 | +107.0% | -31.4% | 1.67 |
| 2025 | +16.9% | -12.1% | 0.78 | +2.9% | -24.0% | 0.26 | +4.9% | -35.7% | 0.35 |
| 2026 | +0.0% | 0.0% | 0.00 | -4.5% | -5.8% | -1.67 | -26.8% | -40.2% | -1.01 |
| Full period | +1607.1% | -41.0% | 1.30 | +1523.0% | -55.7% | 0.90 | +562.8% | -73.9% | 0.69 |
Bear side — short, or a filter
| Year | [TT] HyperGrass v8LONG — with the bear leg | Funding-Heat Short Fade, gated | The same, ungated control | Short everything while risk-off control | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Net | maxDD | Sharpe | Net | maxDD | Sharpe | Net | maxDD | Sharpe | Net | maxDD | Sharpe | |
| 2017 | — | — | — | +0.0% | 0.0% | 0.00 | +0.0% | 0.0% | 0.00 | +0.0% | 0.0% | 0.00 |
| 2018 | — | — | — | +0.0% | 0.0% | 0.00 | +0.0% | 0.0% | 0.00 | +0.0% | 0.0% | 0.00 |
| 2019 | +0.0% | 0.0% | 0.00 | +0.0% | 0.0% | 0.00 | +0.0% | 0.0% | 0.00 | +0.0% | 0.0% | 0.00 |
| 2020 | -8.9% | -41.0% | 0.04 | -2.1% | -5.2% | -0.29 | -22.3% | -40.2% | -0.40 | -30.1% | -46.0% | -0.64 |
| 2021 | +263.8% | -16.8% | 2.54 | +0.1% | -2.1% | 0.07 | -69.4% | -76.5% | -1.52 | +2.4% | -34.6% | 0.29 |
| 2022 | -2.0% | -2.7% | -0.95 | +2.4% | -10.2% | 0.25 | +33.2% | -26.7% | 1.06 | -14.5% | -38.2% | -0.13 |
| 2023 | +131.7% | -21.0% | 2.06 | -5.1% | -8.3% | -0.44 | -62.7% | -64.6% | -2.86 | -19.2% | -20.9% | -0.89 |
| 2024 | +77.2% | -31.1% | 1.58 | -20.7% | -20.7% | -1.56 | -44.3% | -57.7% | -0.98 | -32.5% | -34.4% | -1.16 |
| 2025 | +34.3% | -12.1% | 1.32 | +26.7% | -12.9% | 0.99 | +46.9% | -30.6% | 1.15 | +21.6% | -26.3% | 0.67 |
| 2026 | +5.6% | -4.6% | 1.11 | +10.4% | -8.0% | 1.00 | -2.4% | -25.8% | 0.06 | +8.7% | -12.7% | 0.62 |
| Full period | +1790.8% | -41.0% | 1.33 | +5.7% | -34.5% | 0.11 | -90.6% | -94.9% | -0.51 | -55.9% | -74.5% | -0.11 |
Verified — confirmed edges
These carry information about forward price, but only inside specific regimes. Real, useful, and not all-weather. The passport says exactly when they work and when they don't.
Market Regime (breadth+momentum) TT 70
VERIFIED
A breadth+momentum regime index leads the basket's 30d-forward return with a modest +0.18 corr. Its real value is on the downside: low-regime days ran a median −6.1% over 30 days, up only 38% of the time.
- Full-sample corr of the regime index with the 30-day-forward equal-weight basket return is +0.18 across n=2,142 daily obs: real but modest, in the momentum direction.
- Regime split: corr +0.41 in the 2020-21 bull and +0.12 in the 2023-26 recovery, but −0.07 in the 2022 bear. The trigger follows trends and inverts at turning points.
- Conditional distribution: bottom-regime-tercile days ran a median −6.1% over the next 30 days (up only 38%), vs the top tercile at a median +0.2% (up 51%). The edge is avoiding the downside, not calling tops.
- Built from Binance spot data only; funding was tested and added nothing over breadth + momentum.
Single-coin funding (BTC) TT 58
VERIFIED
BTC funding tracks 30d-forward BTC returns negatively only when leverage runs hot. The link fades to zero in the 2023-26 recovery.
- Full-sample corr_30d = -0.051 on n=2343 daily obs; because 30d-forward windows overlap, the effective sample is ~75, so the full-sample corr sits inside the noise band.
- Regime split: corr = -0.219 in bull 2020-21, -0.099 in bear 2022, +0.002 in recovery 2023-26: negative where funding is elevated, flat once it normalises.
- Conditional distribution: top-funding tercile posts 49% up-days (median 30d-fwd -0.5%) vs 61% up-days for the bottom tercile (median +3%). A 12pp share-of-up gap in the expected direction.
- Sign never meaningfully flips positive. It decays from clearly negative in the leverage era to ~zero, so the trigger is regime-dependent rather than reversing.
Funding Heat (level) TT 52
VERIFIED
Extreme median funding lines up with weaker 30d-forward BTC — but the link only shows up in leverage-heavy bull/bear regimes, not in the 2023-26 recovery.
- Full-sample corr of funding level with 30d-forward BTC return is -0.03 across n=2329 daily obs, sits right at the noise floor once overlap is accounted for.
- Regime split: corr -0.21 in the 2020-21 bull and -0.133 in the 2022 bear, but only +0.025 in the 2023-26 recovery. The effect concentrates in high-leverage phases and fades afterward.
- Conditional distribution: top-tercile funding days are up 48% of the time (median fwd -0.6%) vs bottom-tercile 56% (median fwd +1.7%). An ~8pp share-of-up-days gap in the hypothesised direction.
- Triggers on the highest tercile of market-median funding. A persistent, roughly one-day-in-three condition rather than a rare event.
Single-coin funding (DOGE) TT 48
VERIFIED
DOGE funding runs inverse to 30-day-forward BTC in the 2020-21 bull and 2022 bear, but the relationship decays to noise in the 2023-26 recovery.
- Pooled corr with 30d-forward BTC return is -0.032 over n=2152 daily obs: with heavily overlapping 30d windows the effective sample is ~75, so the pooled number is indistinguishable from noise.
- Regime split shows a consistent inverse tilt in the older leverage-driven regimes: -0.19 in bull 2020-21 and -0.141 in bear 2022, versus a flat +0.057 in recovery 2023-26.
- Conditionally, high-funding (top-tercile) days show a 52% share of up-days vs 57% for low-funding (bottom-tercile) days; top-tercile median 30d-forward move is +0.8% vs +1.9% for the bottom tercile.
- The two negative regimes and the near-zero recent regime are consistent with a real inverse effect in stressed/leveraged periods that fades as DOGE funding decouples from BTC in the recovery.
Momentum acceleration (vs prior) TT 42
VERIFIED
Accelerating momentum is the most robust of the batch: it's the only trigger positive in BOTH the 2020-21 bull (+0.15) and the 2022 bear (+0.11), top-tercile coin-days ran up 43% of the time vs 39% for the decelerating bottom.
- Full-sample corr with 30d-forward return +0.067 across n=157,729 coin-days, modest but real.
- Sign-consistent through opposite regimes: +0.146 in the 2020-21 bull and +0.108 in the 2022 bear: acceleration held up even as the market fell, unlike raw momentum.
- Top acceleration tercile up 43% over 30d vs 39% for the bottom; the edge is a second-derivative confirmation, not a standalone entry.
Vol-scaled momentum TT 40
VERIFIED
Scaling momentum by volatility gives the highest correlation of the batch (+0.085): a coin trending up smoothly beats one that got there on wild swings, but the edge still inverts in the 2022 bear.
- Full-sample corr +0.085 across n=160,459 coin-days. The strongest of the five new tests.
- Regime-dependent: +0.114 in the 2020-21 bull but −0.116 in the 2022 bear. A risk-on trigger, not an all-weather one.
- Top tercile up 44% over 30d vs 41% bottom; the value is filtering out choppy fake trends in favour of clean ones.
Tested & rejected — the spurious pile
Intuitive triggers that look like edges and aren't. We publish these on purpose: a desk that only shows its winners is hiding the survivorship bias.
Price drawdown from high TT 46
REJECTED · NOISE
Correlation with 30d-forward BTC reads 0.144 overall but still flips sign across regimes, and the tercile split is now exactly flat.
- Overall corr with 30-day-forward return is 0.144 on n=3219 daily obs; with heavily overlapping 30d windows the effective sample is about 107, so a reading at this level is not distinguishable from noise.
- Regime corr still flips sign: bull 2020-21 = +0.171, bear 2022 = -0.133, recovery 2023-26 = +0.027. Positive, negative, near-zero, no persistent direction.
- Conditional split is now exactly flat: the deepest-drawdown tercile and the shallowest both show 52% up-days. Not a thin edge, no edge.
- Median 30d-forward moves by tercile are 1.0% deep against 0.7% shallow, a 0.3pp gap that sits inside the daily noise of either bucket.
- What changed: the window. On the desk's own store this battery reaches 2017-08-17 instead of 2020-01-01, and the earlier reading (0.094 on n=2331, bull at -0.124) came from the shorter one. The bull-regime sign inverted outright, which is itself the argument against the trigger: a direction that depends on where the sample starts is not a direction.
Price momentum TT 38
REJECTED · NOISE
Full-sample corr 0.141, but the sign flips across regimes (+0.09 bull, -0.062 bear, +0.036 recovery). The headline is a regime-mix artifact, not a stable trigger.
- Full-sample corr with 30d-forward BTC return is 0.141 over n=2331 daily obs; but 30d windows overlap, so the effective sample is ~75 and a |r| this size sits right at the noise threshold.
- Regime-split corr flips sign: +0.09 in bull 2020-21, -0.062 in bear 2022, +0.036 in recovery 2023-26: momentum that carries in a trend reverses in a bear, so no single-sign effect survives.
- Conditional share-of-up-days is near-flat: top-momentum tercile 56% up vs bottom tercile 54% up. A 2pp gap, well inside sampling error at ~75 effective obs.
Funding sign TT 20
REJECTED · NOISE
Funding sign shows near-zero full-sample corr with 30d-forward BTC and flips direction across regimes, no persistent link.
- Full-sample correlation with 30-day-forward BTC return is 0.045 on n=2368 daily obs; with 30d-overlapping windows the effective sample is ~75, so a |r| this small is indistinguishable from noise.
- The sign is not stable: corr is -0.086 in the 2020-21 bull, -0.139 in the 2022 bear, and +0.029 in the 2023-26 recovery. It flips from negative to positive across regimes.
- The conditional split runs opposite to the hypothesis: the bottom funding tercile carried a higher share of up-days (64%) than the top tercile (53%).
Funding percentile TT 18
REJECTED · NOISE
Full-sample corr looks positive, but the sign flips across regimes, no persistent link between funding rank and 30d-forward BTC.
- Full-sample correlation with 30-day-forward BTC return is +0.124 on n=2210 daily obs: but 30d-forward windows overlap heavily, so the effective sample is only ~2210/30 ≈ 75, at which |r|=+0.124 is not cleanly distinguishable from noise.
- The correlation flips sign across regimes: bull 2020-21 -0.037, bear 2022 -0.04, recovery 2023-26 +0.09. Negative in two regimes, positive in one, not a persistent relationship.
- Conditional split is weak and points opposite to the aggregate: the top funding tercile shows 52% up-days (median fwd +0.6%) vs the bottom tercile 56% up-days (median fwd +1.7%), a ~4pp gap that contradicts the positive full-sample corr.
- One of 13 indicators screened, so a marginal full-sample reading like this is exactly what multiple testing throws up by chance.
Single-coin funding (ETH) TT 18
REJECTED · NOISE
ETH's own funding shows near-zero 30d-forward corr with the market and flips sign across regimes, no reliable lead.
- Full-sample corr with 30d-forward BTC return is just 0.021 on n=2343 daily obs; with heavily overlapping 30d windows the effective sample is ~75, so any |r| this small is indistinguishable from noise.
- The correlation flips sign across regimes: -0.149 in the 2020-21 bull, -0.054 in the 2022 bear, +0.021 in the 2023-26 recovery: not persistent, at best a regime artifact.
- Conditional split is inverted: the top funding tercile precedes up-days only 51% of the time vs 57% for the bottom tercile, so high ETH funding carries no up-side edge over low funding.
- tt score 15 out of 100.
Distance from long SMA TT 18
REJECTED · NOISE
A tempting trend trigger that fails where it matters: the coins stretched FURTHEST above their 200d average. Those coins actually had a WORSE 30-day median (−5.6%) than less-extended ones (−3.9%). The headline +0.07 correlation is driven by below-average coins bleeding, not by overextended ones running.
- Full-sample corr +0.072 (n=145,080) looks positive, but the tercile structure is the tell.
- Top tercile (most extended above 200d SMA) median 30d-forward −5.6%, up only 42%, overextension mean-reverts.
- Both terciles up ~42%: no directional separation; the linear correlation is an artifact of the downside, not a usable long edge.
Cross-sectional momentum TT 12
REJECTED · NOISE
Stretching the momentum lookback further washes the trigger out: the top and bottom terciles were up 44% vs 43%, statistically indistinguishable, and it ran −0.23 in the 2022 bear. Too slow to be a trigger, too noisy to trust.
- Full-sample corr +0.055 (n=154,999). The weakest positive of the batch.
- Top vs bottom tercile up 44% vs 43%, no usable separation over 30 days.
- −0.232 in the 2022 bear: 3-month winners were the biggest losers into the downturn, 90d momentum is too slow to catch turns.
Funding change TT 8
REJECTED · NOISE
7-day funding change shows no stable link to 30-day-forward BTC: corr +0.012 overall, sign flips across regimes.
- Full-sample correlation with 30-day-forward BTC return is +0.012 on n=2322 daily obs, effectively zero.
- Regime corr flips sign: bull 2020-21 +0.009, bear 2022 -0.035, recovery 2023-26 +0.054, no consistent direction.
- Top tercile of 7d funding change: 55% share of up-days vs 50% for bottom tercile. A 5pp gap, median forward move 1.7% vs 0%.
Relative strength vs BTC TT 8
REJECTED · NOISE
Chasing relative strength vs BTC is backwards: the coins that most OUTPERFORMED BTC went up only 38% of the time over the next 30 days (median −7.5%), while the BTC-laggards went up 47% (−1.8%). Hot-vs-BTC coins cool off. This is a mean-reversion tell, the opposite of the hypothesis.
- Full-sample corr with 30d-forward return +0.014, indistinguishable from zero.
- Backwards conditional: top RS-vs-BTC tercile up 38% (median −7.5%), bottom tercile up 47% (median −1.8%).
- The relative-strength leaders were the forward losers across the sample: chasing what already beat BTC is a fade setup, not a follow.
Realized volatility TT 5
REJECTED · NOISE
Realized vol 30d holds no reliable link to 30-day-forward BTC return: +0.073 on n=3218, and the bull-regime concentration that once carried it does not survive the longer history.
- Full-sample corr with 30d-forward BTC return = +0.073 across n=3218 daily obs; the sign stays positive in every regime (bull 2020-21 0.026, bear 2022 0.047, recovery 2023-26 0.058).
- The 2020-21 bull, which used to carry the whole effect at 0.115, now reads 0.026: no regime stands out, and none reaches a level worth acting on.
- Conditional split is thin: the high-vol tercile shows 56% share-up days (median fwd +2.0%) against 54% for the low-vol tercile (median fwd +1.4%). A 2pp gap.
- Direction note: the positive corr means high realized vol preceded firmer, not weaker, forward returns. The opposite of the original hypothesis, and too small to trade either way.
- What changed: the window. On the desk's own store the battery reaches 2017-08-17 instead of 2020-01-01, and the earlier reading of +0.126 on n=2330 was measured on the shorter one. A trigger that only exists on the sample it was found in is the definition of the thing this page is here to catch.
Funding dispersion TT 2
REJECTED · NOISE
Wide cross-coin funding spread barely moves with 30d-forward BTC returns: +0.054 over the full sample, and no regime lifts it out of the noise.
- Full-sample correlation with the 30d-forward BTC return is +0.054 (n=2377 daily observations). With 30-day overlapping windows the effective sample is about 79, so this is indistinguishable from noise.
- The sign is positive in all three regimes and small in each: bull 2020-21 +0.061, bear 2022 +0.058, recovery 2023-26 +0.012. There is no regime where the rule pays.
- Days in the top dispersion tercile were followed by an up-day 55% of the time against 53% for the bottom tercile. A 2pp gap, inside the sampling error of this sample.
- The conditional median 30d-forward move was +2.2% after high-dispersion days against +0.9% after low-dispersion ones.
Volume spike TT 1
REJECTED · NOISE
Full-sample corr with 30d-forward BTC return is +0.021: flat, and the sign flips across regimes.
- Over n=2341 daily obs the correlation between volume spike and 30-day-forward BTC return is +0.021, effectively zero.
- Regime corr flips sign: bull 2020-21 +0.011, bear 2022 +0.084, recovery 2023-26 +0.022, no stable direction.
- Conditional split runs backwards: top-tercile days are up 53% of the time vs 57% for bottom-tercile. The high-volume group has fewer up-days, not more.
Funding Heat z-score TT 0
REJECTED · NOISE
Funding Heat z-score shows a near-zero 30d-forward corr (+0.027) that flips sign across regimes, no persistent directional trigger.
- Full-sample correlation with 30-day-forward BTC return is +0.027 on 2342 daily observations: flat, and below the noise floor once overlapping 30d windows cut the effective sample to ~75.
- The correlation flips sign across regimes: -0.045 in the 2020-21 bull, -0.063 in the 2022 bear, +0.067 in the 2023-26 recovery, direction is not stable.
- Days in the top heat tercile closed up over the next 30d 53% of the time vs 51% for the bottom tercile. A 2pp gap, inside the noise band.
How we test
We take a trigger and look at what bitcoin did over the next 30 days. We measure the link (correlation) on the whole history since 2017, then separately in three periods: the 2020–21 rise, the 2022 fall, the 2023–26 recovery. A trigger passes if the link holds in at least one period and keeps its sign. The second condition: a person who re-checks the numbers finds no error. Anything that flips sign from period to period, or cannot be told from noise, is rejected, and stays on the page.
Where the limit is. The 30-day windows overlap, so there are not 3,305 independent observations but about 110. A link weaker than 0.10 on the full sample cannot be told from noise. We screened 19 triggers, so some "finds" will be chance: normal for a search, which is why we print what we threw out. These are tests on the whole history, not on a held-out part: a number says the link was there in 2017–2026-09 and does not promise it will be there tomorrow. Only the cross-sectional ranker went through a real held-out test, and it is scored separately. A lab verdict is a reason to watch a trigger live, not a return.
How this compares — and what makes it different
Dashboards give you live data, on-chain desks give you depth. The AI lab does something else: every trigger is checked in public. A backtest with controls, a passport, the rejected pile and a forward paper account. The table compares honesty of method, not returns.
- You see whether a rule worked on the whole history, not on a lucky slice.
- You see the control — the same rule on shuffled names. Same result means an empty rule.
- You see Sharpe and Sortino — return adjusted for risk, not a bare percentage.
- You see the paper account — the rule trades forward and the live result is checked against the backtest.
| Capability? | top traders · labthis page? | Funding dashboardsCoinGlass, Laevitas? | On-chain desksGlassnode, CryptoQuant? | Signal servicesTelegram “AI signals”? |
|---|---|---|---|---|
| Live funding & derivatives data | ✓ | ✓ | ✓ | — |
| Multi-year historical funding | ✓ | ✓ | ✓ | ✕ |
| Per-indicator backtest vs forward price | ✓ | ✕ | ◐ | ◐ |
| Regime-split validation (bull/bear/recovery) | ✓ | ✕ | ◐ | ✕ |
| Verification passport per indicator | ✓ | ✕ | ✕ | ✕ |
| Published rejected / spurious indicators | ✓ | ✕ | ✕ | ✕ |
| Descriptive framing — no win-rate / yield promise | ✓ | ✓ | ◐ | ✕ |
| Backtest with controls on the full history | ✓ | ✕ | ✕ | ✕ |
| Paper trading — forward accounts, live | ✓ | ✕ | ✕ | ◐ |
| Sharpe and Sortino in every report | ✓ | ✕ | ◐ | ✕ |
| The data behind every number is open (JSON) | ✓ | ◐ | ✕ | ✕ |
| Open and bilingual | ✓ | ◐ | ✕ | ✕ |
How to read a row → Verification passport per indicator: lab ✓, dashboards ✕, on-chain ✕, triggers ✕: only this page ships a per-indicator test passport; the others don't publish one at all. That is the specific gap the AI lab fills.
✓ yes ◐ partial / paywalled ✕ no — n/a
The passport
Every indicator ships its own test period, sample, regime-split correlation and honest caveats. A dashboard shows you the funding number; we show whether it carries information: when, and how much.
Turned out useless
We publish the triggers we tested and threw out. Signal shops quietly delete their losing trades; data dashboards never test at all. The failures are the proof the winners are real.
Descriptive, not promissory
Correlation, lead, trigger frequency, never “this returns X%” or a 90% win-rate. The framing that survives the first miss, not the one that sells a subscription.
Fair play: CoinGlass and Laevitas are excellent live-data tools, and Glassnode has published serious work on point-in-time backtesting. None of them publish a per-indicator verification passport with the rejected pile for a public audience. That is the specific gap the AI lab fills.
The TT desk thoughts
Where our edge is. Not in the data: candles, funding and on-chain are the same for everyone. The edge is three things. First, every rule is written down before the test and run with costs on the whole history, long and short. Second, every number has a control: the same rule on shuffled names, and a random entry. If the control gives the same result, the rule is empty, and we print that. Third, everything that passed trades forward on paper accounts of $100,000 each, and the live result is checked against the backtest every day.
Why it is hard to copy. The pile of rejected rules costs more than a list of winners: it had to be counted honestly and not hidden. Years of data, the engine, the controls and the accounts are one pipeline, and every new trigger goes through all of it. A dashboard shows a number; a trigger channel shows its wins; we show what we tested, what we threw out, and how it trades right now. The flagship lives on the funding-rate gauge; the working notes are on @toptraders0x.
Related
Scores, candidate passports and the working notes
The 0-100 verification score for every trigger, and the full passport for each candidate still on the bench. This is the working evidence behind the tables above, kept closed because a reader who wants a verdict should not have to scroll past the exhibits to reach it.
score — the effectiveness ranking
One number per indicator, ranked. The score (0–100) rates how well a trigger survived multi-regime verification: effect size, how many regimes carry the same-sign effect, and the conditional separation, penalised for instability. It rates verification strength, not profitability. A high score means the relationship is real and consistent, never that it “returns X%”.
| #? | Indicator? | score? | Grade? | What the backtest shows? |
|---|---|---|---|---|
| 1 | Market Regime (breadth+momentum) | 70 | confirmed | r +0.18 pooled · peak +0.41 in Bull 2020-21 · sign flips across regimes · up-day gap 13pp |
| 2 | Single-coin funding (BTC) | 58 | confirmed | r -0.05 pooled · peak -0.22 in Bull 2020-21 · one sign in the meaningful regimes · up-day gap 12pp |
| 3 | Funding Heat (level) | 52 | confirmed | r -0.03 pooled · peak -0.20 in Bull 2020-21 · one sign in the meaningful regimes · up-day gap 7pp |
| 4 | Market screener (RSI breadth) candidate | 50 | confirmed | r +0.02 pooled · peak -0.26 in Bear 2022 · one sign in the meaningful regimes · up-day gap 7pp |
| 5 | Single-coin funding (DOGE) | 48 | confirmed | r -0.03 pooled · peak -0.19 in Bull 2020-21 · sign flips across regimes · up-day gap 6pp |
| 6 | Price drawdown from high | 46 | confirmed | r +0.14 pooled · peak +0.17 in Bull 2020-21 · sign flips across regimes · up-day gap 0pp |
| 7 | Momentum acceleration (vs prior) | 42 | confirmed | r +0.07 pooled · peak +0.15 in Bull 2020-21 · one sign in the meaningful regimes · up-day gap 4pp |
| 8 | Vol-scaled momentum | 40 | confirmed | r +0.09 pooled · peak -0.12 in Bear 2022 · sign flips across regimes · up-day gap 3pp |
| 9 | Price momentum | 38 | confirmed | r +0.14 pooled · peak +0.12 in Bull 2020-21 · sign flips across regimes · up-day gap 3pp |
| 10 | Funding screener (most-crowded long) candidate | 32 | weak | r -0.01 pooled · peak -0.22 in Bull 2020-21 · one sign in the meaningful regimes · up-day gap 2pp |
| 11 | SMA distance (regime gate) candidate | 32 | weak | r +0.12 pooled · peak -0.38 in Bear 2022 · sign flips across regimes · up-day gap 4pp |
| 12 | Funding sign | 20 | weak | r +0.04 pooled · peak -0.14 in Bear 2022 · one sign in the meaningful regimes · up-day gap 10pp |
| 13 | Funding percentile | 18 | noise | r +0.12 pooled · peak +0.09 in Recovery 2023-26 · one sign in the meaningful regimes · up-day gap 2pp |
| 14 | Single-coin funding (ETH) | 18 | noise | r +0.02 pooled · peak -0.15 in Bull 2020-21 · one sign in the meaningful regimes · up-day gap 9pp |
| 15 | Distance from long SMA | 18 | weak | r +0.07 pooled · peak -0.19 in Bear 2022 · one sign in the meaningful regimes · up-day gap 0pp |
| 16 | Cross-sectional momentum | 12 | noise | r +0.06 pooled · peak -0.23 in Bear 2022 · one sign in the meaningful regimes · up-day gap 1pp |
| 17 | Funding screener (crowding spread) candidate | 10 | noise | r +0.05 pooled · peak +0.10 in Bull 2020-21 · one sign in the meaningful regimes · up-day gap 3pp |
| 18 | Funding change | 8 | noise | r +0.01 pooled · peak +0.05 in Recovery 2023-26 · one sign in the meaningful regimes · up-day gap 6pp |
| 19 | Relative strength vs BTC | 8 | noise | r +0.01 pooled · peak -0.08 in Bear 2022 · one sign in the meaningful regimes · up-day gap 9pp |
| 20 | Realized volatility | 5 | noise | r +0.07 pooled · peak +0.05 in Bear 2022 · sign flips across regimes · up-day gap 3pp |
| 21 | Funding dispersion | 2 | noise | r +0.05 pooled · peak +0.06 in Bull 2020-21 · one sign in the meaningful regimes · up-day gap 2pp |
| 22 | Volume spike | 1 | noise | r +0.02 pooled · peak +0.08 in Bear 2022 · one sign in the meaningful regimes · up-day gap 2pp |
| 23 | Funding Heat z-score | 0 | noise | r +0.03 pooled · peak +0.07 in Recovery 2023-26 · sign flips across regimes · up-day gap 2pp |
How to read a row → #1 Market Regime · score 70 · confirmed · r +0.18 pooled, peak +0.41 in Bull 2020-21: it ranks top because the trigger is strong and same-signed in the bull regime, but “confirmed” means it carries information in some regimes, not all: read the score as a ranking of verification strength, never as a return.
How the score is computed
score = 100 × (0.5·effect + 0.2·separation + 0.3·regime-breadth) × instability-penalty, where effect blends the pooled and best-regime |correlation| (0.05→0, 0.20→1), separation is the top-vs-bottom-tercile up-day gap, regime-breadth is how many of the three regimes carry a same-sign |r|≥0.10, and the instability penalty (×0.5) fires when the pooled correlation opposes the strongest regime. The classic noise fingerprint. Grades: persistent ≥60, confirmed 35–59, weak 20–34, noise <20. It is a descriptive summary of the passports below, on 3,305 daily observations with overlapping 30d-forward windows: read magnitudes as ranking, not as tradable point estimates.
Candidates — what is on the bench right now
Everything the desk is currently testing, in plain words: what the rule is, what it actually returned when tested with costs, the verification score, and whether we would act on it. A candidate is research. Nothing here is a promise.
Funding screener (crowding)
score 41candidate — score 41, conditional; the crowding spread carries the trigger, the crowded long does not
The screener ranks perps by annualised funding and prints the most-crowded longs and shorts. That ranking metric had never been tested against what happens next, so both ends of it were scored the same way every AI lab indicator is: against 30-day forward BTC returns across 2,355 days and three regimes. The gap between the most-crowded long and the most-crowded short scores 41 (conditional). The most-crowded long on its own scores 29 (weak) and its pooled correlation is -0.006, effectively nothing.
| What was measured | Correlation | Top third | Bottom third |
|---|---|---|---|
| Спред скученности | +0,114 | 58% растущих дней | 51% |
| Самый переполненный лонг | −0,006 | 55% | 52% |
| Бычий рынок 2020-21 | +0,097 / −0,207 | — | — |
| Сейчас, 2023-26 | +0,004 / +0,003 | — | — |
2,355 days against 30-day forward BTC returns across three regimes.
What we would do with it: Read the crowding spread as a market stress gauge, never as a reason to open a position. Do not use the most-crowded-long reading on its own: outside the 2020-21 bull it has no support. Both triggers have gone flat in the current regime.
Tested: 2026-07 · v1.0 · 20-coin liquid perp basket, funding archive, 2,355 daily observations with overlapping 30d-forward windows across three regimes
Supply Dilution
score 34candidate factor — cross-sectional basket, long slow float growth / short fast
Rank the perp universe by trailing circulating-supply growth, long the slowest-diluting third and short the fastest, equal weight per leg and legs equal to each other. The idea came from reading what the durable Hyperliquid vaults actually hold rather than from a parameter sweep: they say plainly that they short high-emission, high-FDV names and hold the majors. Over 363 days on 33 perps the basket returned +25.6% at a -7.6% maximum drawdown while holding that same basket returned -68.0%, and it stayed positive on a held-out window it had never seen.
| Construction | Total | Drawdown | Sharpe |
|---|---|---|---|
| Держать корзину (контроль) | −68,0% | — | — |
| Фактор, dollar-neutral | +25,6% | −7,6% | +1,82 |
| Он же вне выборки (июн–июл) | +6,5% | −2,1% | +3,86 |
| Контроль: ранг по капитализации | +9,3% | −20,7% | +0,51 |
| Контроль: ранг по моментуму | −3,8% | −20,5% | −0,09 |
Costs are charged in full, real funding included. 33 coins.
What we would do with it: Run it in its neutral form and in small size. It shows that slowly diluting coins beat quickly diluting ones, and that survives both the majors-versus-alts control and the momentum control. It has only been tested on one falling market over twelve months, so size it such that a sign flip in a bull cycle would not hurt.
Tested: 2026-07 · v1.0.0 · 33 liquid perps, 2025-07-28..2026-07-25, with a held-out Jun-Jul 2026 window
Supply Dilution — regime-gated basket
score 34candidate — the dollar-neutral basket held only below the long-term trend; the gain is switched-off beta, not a better factor
Run the Supply Dilution basket: long the slowest-diluting third, short the fastest, dollar neutral: but hold it only while BTC trades below its long-term average, and sit flat otherwise. Over 2025-07-28 to 2026-07-25 that returned +35.8% against +25.6% for running it continuously, at a -5.3% maximum drawdown against -7.6%, while being in the market 57% of the days. The improvement is real and its cause is not flattering to the factor: measured net dollar exposure is exactly zero, but realised beta is -0.046 and correlation to the universe is -0.25, so the book quietly carries short beta. The gate simply stops paying for that beta while the market rises.
| Construction | Total | Drawdown | Sharpe |
|---|---|---|---|
| Корзина без фильтра | +25,6% | −7,6% | +1,82 |
| Корзина под фильтром | +35,8% | −5,3% | +3,11 |
| В рынке дней | 57% | — | — |
| Хедж мажорами вместо трети | +26,8% | −13,8% | +1,17 |
The same factor as above plus the trend filter. Two fitted parameters now sit on a twelve-month sample.
What we would do with it: Use it only while understanding where the extra return comes from. At zero net dollar exposure the basket still leans short: it loses 7.8bp on up-market days and makes 20.1bp on down days. The gate simply switches it off while the market rises. That is a market-timing overlay, not better coin selection, and the cleaner fix is to size the legs to equal beta.
Tested: 2026-07 · v1.0 · 33 binance_futures perps, 2025-07-28..2026-07-25, gate from the BTC daily close against its own long-term average
SMA regime gate
score 30candidate — halves the drawdown, does not generate the return; TT score 30 because the score measures the wrong thing
Hold BTC and ETH equal-weight while the daily close is above its long-term moving average, otherwise sit in cash, and trade only on the crossing. The rule comes from the published CORE arm of an external spot system and is specific enough to rebuild, so it was rebuilt on our own archive rather than taken on trust. Over 2020-2026 it returned 12.9x against 9.5x for holding the same pair, at a 41% maximum drawdown against 76%. The drawdown number is the finding; the return number is mostly one year.
| Window | Gate | Hold | Gate drawdown | Hold drawdown |
|---|---|---|---|---|
| Всё окно 2020-2026 | 12,9x | 9,5x | −41% | −76% |
| Без 2022 года | 10,8x | 9,3x | −41% | −75% |
| С 2023 года | 2,0x | 2,6x | −30% | −60% |
| Бычий 2020-2021 | 5,9x | 10,3x | −41% | −52% |
BTC and ETH on daily candles, costs charged on every crossing. The length was checked against the whole grid of alternatives: 27 of 39 work, so the number is not fitted.
What we would do with it: This is a way to cut drawdown, not to earn more. The rule loses to holding in every rising year and wins in every falling one; over six years it halved the drawdown. The low score describes the yardstick rather than the rule: the AI lab formula rates prediction of returns, while all this rule claims is sitting out losses.
Tested: 2026-07 · v1.0 · BTC and ETH daily closes rebuilt from the local spot archive, 2020-05-01..2026-07-26, 2,278 bars
Market screener (RSI breadth)
score 27candidate — score 27, weak; the tercile separation runs against the board's own direction
The screener ranks coins by RSI and 24-hour change and prints a long board and a short board. Whether that ranking says anything about what comes next had never been measured, so the share of the basket the board would list long was scored the same way every AI lab indicator is: against 30-day forward BTC returns over 2,132 days and three regimes. It scores 27, weak. The pooled correlation is +0.03, the regimes disagree on sign, and the conditional separation points the wrong way.
| What was measured | Value |
|---|---|
| Корреляция с доходностью через 30 дней | +0,03 |
| Бычий 2020-21 / медвежий 2022 / сейчас | −0,119 / −0,252 / +0,035 |
| Верхняя треть: растущих дней | 50% |
| Нижняя треть: растущих дней | 54% |
2,132 days. The reconstruction is approximate: the live board ranks the full TradingView universe while the archive holds 20 liquid coins.
What we would do with it: Do not read the board as a forecast, on this evidence it is not one. The separation even runs against its own logic: when more coins sit above the RSI midline, what follows is slightly worse. As a filter of what currently passes liquidity and momentum screens the board stays useful, and that is its actual job.
Tested: 2026-07 · v1.0 · 20-coin liquid basket reconstructed from the spot archive, 2,132 daily observations with overlapping 30d-forward windows across three regimes
TradFi Breakout
score 0the rule is written, the measurement is not
Breakouts on trade.xyz markets: equities, indices, commodities, FX — with a regime gate built from the universe's own breadth, a relative-strength check, a toxicity veto, and funding treated as a cost filter. There is no backtest and there cannot be one soon: the longest history on the venue is 292 daily bars.
| What was measured | Value |
|---|---|
| Кривая капитала | нет |
| Контроль (удержание списка инструментов) | нет |
| Перестановка имён со совпадением по стороне | нет |
| Максимальная история площадки | 292 дневных бара |
| Медиана истории | 129 баров |
| Рынков с историей от года | 0 из 88 |
None of the six figures the report standard requires can be computed on this much history. The product exists so the rule is written down and falsifiable before a sample exists, not as a substitute for measuring it.
Measured on Yahoo daily history
Verdict: rejected. Failed: по итогу ниже эталона «купить и держать» — при том что эталон собран из сегодняшних выживших и потому завышен; sharpe 0.74 ниже пола деска 1.0. Passed: итог выше p95 перестановки имён (запас 17%), выиграли у правила 4 из 200; итог выше p95 случайного входа при тех же днях входа и том же числе долей счёта на инструмент (запас 35%), выиграли у правила 3 из 200; по риску выше эталона: sharpe 0.74 против 0.64, просадка -22.13% против -73.1%, sortino 1.09 против 0.94.
Controls
| control | median | p95 | rule |
|---|---|---|---|
| name permutation | +1 108,7% | +2 277,6% | +2 665,5% |
| random entry, same gate and days | +1 009,6% | +1 979,0% | +2 665,5% |
| buy and hold the basket | +27 859,5% · Sharpe 0.64 | +2 665,5% · Sharpe 0.74 | |
| S&P 500 | +6 012,7% · Sharpe 0.59 | +2 665,5% · Sharpe 0.74 | |
Halves of the window
| window | total | Sharpe | max drawdown |
|---|---|---|---|
| 1980-09-04..2003-08-20 | +243,0% | 0.65 | −16,3% |
| 2003-08-20..2026-09-02 | +705,5% | 0.83 | −22,1% |
Phases (S&P 500 against its own long-term average)
| in its phase | outside it | |
|---|---|---|
| total | +3 623,8% | −26,1% |
| Sharpe | 0.95 | -0.56 |
| trades | 1747 | 108 |
By year
What differs from the live rule
- Фандинга на Yahoo нет: фильтр стоимости удержания в этом прогоне не работал. На площадке он снимает часть имён, поэтому результат здесь оптимистичнее живого правила, а не осторожнее.
- Спот вместо перпа: отрыва от оракула, открытого интереса и потолка OI площадки в этих данных нет вовсе, и четыре фильтра гигиены исполнения тоже не работали.
- Список инструментов — СЕГОДНЯШНИЙ состав площадки, отыгранный назад. Это отбор по выжившим: имена, которых площадка не листит, в замер не попали ни одного дня. Число — верхняя граница.
- Цены сплит-скорректированные, дивиденды НЕ реинвестируются: перп их тоже не платит.
- Календарь решений — торговые дни S&P 500. Имя без бара в этот день входит последним известным закрытием, но не старше порога свежести; молчание длиннее убирает имя из списка инструментов, а не тянет за собой позавчерашнюю цену.
- R считается как доходность сделки, делённая на трейлинг-стоп правила: стопа по цене входа у правила нет.
- Допуск имени в список инструментов спрашивает объём за ПОСЛЕДНИЙ год ряда, и ответ действует на всю его историю. Имя, у которого источник перестал давать объём сегодня, выпадает целиком, включая годы, где объём был. Это заглядывание вперёд на одну величину — «даёт ли источник объём» — и оно рядом с отбором по выжившим.
- Соответствие тикера площадки символу биржи подтверждено двумя картами деска не у всех имён; подтверждённые одной перечислены в артефакте замера — их сверять было не с чем.
- Эталон «купить и держать» покупает имя один раз в день его допуска, равной долей стартового счёта, платит издержки одной стороны и больше книгу не трогает. Рядом печатается корзина с ежедневным ребалансом БЕЗ издержек — она завышена ребалансом и оставлена только ради сопоставимости с прежними отчётами.
Yahoo Finance, daily bars · 84 names · window 1980-09-04 → 2026-09-02
Tested: 2026-07 · v0.1.0 · не тестировано — на площадке максимум 292 дневных бара (XYZ100), медиана 129, рынков с историей от года ноль
Donchian breakout — the long channel with a trailing stop
backtest +137.1% · Sharpe 0.45 · Sortino 0.61 · maxDD −71.1%candidate algorithm — long-only spot trend following, no leverage
A bar closes above the highest high of the channel, you buy the next open, and a stop trailing a multiple of the daily ATR below the highest close since entry takes you out. No ranking, no profit target, no moving averages. The rule as specified fails its own placebo and loses money after February 2022. One of the thirty variants tested clears both nulls: the daily bar with a breakout-quality filter that requires the breaking bar to clear one bar-level ATR and leave a fair value gap.
| Variant | Total | Sharpe | Drawdown | Second half | Nulls |
|---|---|---|---|---|---|
| 1h база | +137.1% | 0.45 | −71.1% | −31.7% | не проходит |
| 1h vol+regime | +184.0% | 0.50 | −61.1% | −16.5% | не проходит |
| 4h vol+regime+fvg | +177.3% | 0.51 | −56.7% | −7.1% | не считались |
| 1d fvg | +197.1% | 0.53 | −58.4% | +7.7% | проходит оба |
| 1h adx | +86.1% | 0.38 | −70.9% | −38.1% | не считались |
| BTC, купить и держать | +1416% | 0.79 | −83.2% | — | — |
Thirty cells: three decision bars crossed with ten rule variants, one engine and one universe. Nulls were run at 200 seeds on four rows; the rest say so.
What we would do with it: Do not trade it. The base rule fails its placebo and has lost money since 2022; the surviving variant was picked as the best of thirty and has to be confirmed on a fresh window. All thirty variants are stored as specs with Binance spot execution profiles, so the confirmation can be checked with the same file a bot would trade from.
Tested: 2026-08 · v0.1.0 · 594 Binance spot instruments including delisted ones, 2017-08 to 2026-08, decision bar aggregated from intraday bars
Fast-channel breakout with a channel exit and a volatility filter — a bull-phase algorithm
backtest +1,583.8% · Sharpe 1.42 · Sortino 2.37 · maxDD −32.6%the fast channel, exit on the opposite channel, a volatility-expansion filter — long-only, no stop
Entry on a fast channel breakout, taken only when volatility is expanding. The short ATR above the long one. Exit when a bar closes below the low of the opposite channel. The filter lifts sharpe 1.31 to 1.42. The highest total among the finalists (+1584%), paid for with the deepest drawdown of the three (-32.6%).
Tested: 2026-08 · v0.1.0 · 10 Binance perp majors on daily bars, 2020-09 to 2026-08
Slow-channel breakout with a channel exit and a volume filter — a bull-phase algorithm
backtest +1,415.7% · Sharpe 1.42 · Sortino 2.44 · maxDD −23.7%the turtle exit survives where stop lines fail — long-only, exit fills never better than the open
Entry on a slow channel breakout confirmed by volume. The trigger bar trades well above its own median, the Effort filter. Exit when a bar closes below the low of the opposite channel. The channel exit is the only non-time exit that survived every control; the volume filter lifts sharpe 1.36 to 1.42 and cuts the drawdown from -33% to -24%.
Tested: 2026-08 · v0.1.0 · 10 Binance perp majors on daily bars, 2020-09 to 2026-08
Fast-channel breakout — a time exit
backtest +1,382.3% · Sharpe 1.51 · Sortino 2.66 · maxDD −29.9%candidate from the exit laboratory — long-only, no leverage, no stop by design
A bar closes above the highest high of the fast channel, you buy the next open and sell at the close a fixed number of bars later. No stop: the exit laboratory measured 14 exit rules against per-rule random-entry nulls on 560 runs, and every stop-based rule cut the entry's edge before it showed. The top of the replication table is held entirely by time exits. This rule cleared the 95th percentile of its own null on 7 of 10 markets.
Tested: 2026-08 · v0.1.0 · 10 Binance perp majors on daily bars, 2020-09 to 2026-08, common calendar from AVAX's listing, no overlapping trades
Slow-channel breakout, a time exit — a bull-phase algorithm
backtest +1,098.6% · Sharpe 1.59 · Sortino 3.03 · maxDD −21.2%the cross-laboratory's winner over 98 entry-exit pairs — long-only, no leverage, no stop by design
A bar closes above the highest high of the slow channel, you buy the next open and sell at the close a fixed number of bars later. The slower channel beats the published faster variant on both sharpe (1.59 vs 1.51) and drawdown (-21% vs -30%) with half the trades. The best phase goes in the headline per the desk's rule: 88% of the log return accrues in bull phases; in bears it sits near zero without losing.
Tested: 2026-08 · v0.1.0 · 10 Binance perp majors on daily bars, 2020-09 to 2026-08, one position per symbol; floor measured on the point-in-time top-30 basket
Funding-Heat Short Fade
backtest — table belowcandidate indicator — short-side tactical index
Tactical short-side index that targets overheated names. Exploits elevated funding-rate z-score by entering short on names paying abnormally high funding to capture mean-reversion, supplemented by earning the high funding carry. Regime-dependent (only active in risk-off/bear windows; flat in risk-on).
| Condition | Result |
|---|---|
| Бычий рынок 2023-2024 | −54,2% годовых |
| Медвежий режим | сильный результат |
| Доля альфы от фандинга | значительная при высоких ставках |
The regime breakdown was not re-run in the AI lab framework, so there is no comparable score.
What we would do with it: Do not switch it on without a falling-regime gate. In a bull market it is destroyed. That is not a caveat but a measured −54.2% a year over 2023-2024.
Tested: 2026-07 · v1.0 · full 2020-2026 funding archive, split by regime window
[TT] HyperMomentum v1
backtest — table belowcandidate — long-only momentum on 16 majors; the finding is that a high win rate is the defect, not the edge
Built by measuring what the Growi HF vault actually does rather than what it says. Its description claims mean reversion; joining 79,981 of its Open fills to market hourly bars says the opposite: at the hour of a long entry price had already risen 0.32% over the prior hour and 2.78% over the prior day and sat 1.93% above its own 24h mean, against an unconditional background of -0.006% and a 50/50 split. That is a momentum entry. Rebuilt under our constraints: long only, no leverage, no hedge, costs charged in full. The mechanic returns +19.6% against -44.6% for holding the same basket, at a -39.9% drawdown against -67.8%.
| Variant | Trades | Total | maxDD | Sharpe | win% | Edge |
|---|---|---|---|---|---|---|
| широкая цель, стоп есть, длинное удержание | 2 253 | +19,6% | −39,9% | +0,68 | 65,7 | +8,4 бп |
| цель уже, без стопа, среднее удержание | 2 643 | +4,2% | −55,5% | +0,28 | 77,5 | +3,0 бп |
| узкая цель, без стопа, длинное удержание | 2 840 | −41,2% | −64,6% | −1,55 | 91,3 | −13,6 бп |
| самая узкая цель, без стопа, короткое удержание | 9 607 | −97,8% | −97,8% | −9,29 | 85,3 | −31,0 бп |
| Держать корзину (контроль) | — | −44,6% | −67,8% | — | — | — |
Placebo for the best cell is -39.8 +/- 7.8bp, so the edge is 6.2 sigma clear. Costs are charged in full, real funding included.
What we would do with it: Do not ship as a product. The value here is the negative result: a high win rate measurably hurts, and that applies to any strategy that takes profit early and does not cut losses. Before returning to the positive cell it needs a second regime and a universe with no gaps.
Tested: 2026-07 · v1.0 · 16 liquid majors, hourly bars 2025-06-29..2026-07-25, 9,384 bars, long only, gross <= 1x, held-out window from 2026-06-01
OI-Momentum (DD-filter)
backtest — table belowcandidate indicator — risk-on gate / drawdown filter
A risk-management filter that uses open interest (OI) change to gate entry. While long-side continuation suffers a negative pooled CAGR of -25% in clean bear markets, applying this as a DD-filter successfully cuts maximum drawdown by 2x to 3x across 10 out of 10 backtested assets and beats buy-and-hold on Sharpe ratio.
| What was measured | Value |
|---|---|
| Корреляция с доходностью через 1 день, BTC | +0,310 |
| То же, ETH | +0,186 |
| То же, в пуле | +0,179 |
| Терцильный спред к 1-дневной доходности | +0,96% |
| Просадка при использовании как фильтра | в 2–3 раза меньше на 10 из 10 активов |
No regime breakdown was measured, so a comparable TT score is not computed: the numbers above are per coin, not per regime.
What we would do with it: Use it only as a risk filter: it does not make money, it stops you buying strength while the regime is red. As a standalone long trigger in a clean bear market it returns −25% a year.
Tested: 2026-07 · v1.0 · BTC and ETH open-interest archive, 1-day forward returns, pooled and per-coin
Structure break with a channel exit — a bull-phase algorithm
backtest +717.1% · Sharpe 0.96 · Sortino 1.47 · maxDD −45.0%market structure as an entry rule, both sides, no leverage
A close above the last confirmed swing high opens a long, a close below the last confirmed swing low opens a short; the position leaves on the opposite channel. Swings are published at their confirmation bar, so nothing in the rule knows a price that had not printed yet.
Tested: 2026-07 · v0.1.0 · 10 USD-M perp majors on daily bars, 2020-09 to 2026-07, funding accrued daily, one position per symbol, long and short
Candidates we rejected — and the measurement that did it
A rule that failed its own test stays on this page. Deleting it would leave the bench looking like a list of things that worked, which is the one shape research must never take. Each card carries the number that ended it.
Long — Episodic Pivot (бычий режим)
-42.1%candidate — measured, negative result
Кулламагги — Episodic Pivot on the panel engine. Window 365d: -42.14% at 71.7% drawdown, 151 trades, 6.0% winners.
- Window 365d: -42.14% at 71.7% drawdown, 151 trades, 6.0% winners.
- Entry: impulsePct4h >= 10, volSurge >= 1, rvol >= 3, perf3m <= 30, aboveMa20 = 1, regimeRiskOn = 1.
- Exit: стоп 1 ADR, трейл по MA20. Position 15% of the account.
- Trigger at bar close, fill at the next open; costs charged on both sides.
- Without the best three trades: -83.40%.
- Fingerprint: input 974d7ad1b1174bc3 → result 828fcfedafbb2161; method repo-e2667d2285de6fbd.
2026-08 · v2 · окно 365d, движок панели, издержки с обеих сторон, доля на сделку 15%
Cross-sectional ranker
higher-timeframe research record
Model-card status is derived at refresh time; it is never a standalone trading instruction.
- Method: closed UTC daily candles on the higher timeframe.
- Current gates: regime_edge_concentration.
- Positive OOS edge concentration: 60.95% by regime, 49.54% by episode; limit 60%.
2026-08-09 · vcross-sectional-ranker-v0 · held-out out-of-sample evaluation
Liquidation Cascade (fade)
rejected — the effect is real but an order of magnitude too small to pay for execution
A liquidated position is sold because margin ran out, not because anyone decided the asset was worth less, so a cascade should push price past what information justifies and then come back. Tested on 1.37 million liquidation events across BTC, ETH and SOL over 392 days, that reversal is measurably there and decays with horizon exactly as the hypothesis predicts: and it is worth 0.2 to 5 basis points, a fraction of what a round trip costs. It is rejected on magnitude, not on absence.
- Against a side-matched placebo — the same trades placed at random times — 17 of 216 cells clear two standard deviations, against roughly 5 expected from noise.
- The significant cells cluster at the shortest horizon and vanish by 24 hours: 12 of 54 cells at 15 minutes, 4 at one hour, 1 at four hours, 0 at 24 hours. That decay is the signature of a genuine micro-reversion.
- Median gross edge across all 216 cells is +0.4bp; the best horizons run +0.2bp to +5.4bp. A round trip costs 4 to 10 times that, not a few percent more.
- Ten cells did clear the cost line, all short-side, all at a 24-hour hold — the horizon where the effect has already vanished. Against a short-only placebo every one sits inside a single standard deviation, and ETH runs the other way. Their profit is shorting a falling market, not a trigger.
- Direction is not consistent across assets at the tradeable horizons: the short side on BTC grosses +7.4bp while the identical construction on ETH grosses -2.1bp.
- A look-ahead was found and removed mid-study: defining exhaustion as 'the next bar is smaller' let the trigger read a liquidation total that does not exist yet at the entry it triggers. Before the fix the gross edge appeared to grow with horizon up to +20.9bp; after it, that growth disappears entirely.
2026-07 · v1.0 · CryptoHFTData liquidations, binance_futures BTC/ETH/SOL, 2025-06-29..2026-07-25, 392 days with no gaps, 1.37M events