Accuracy, in plain English

Every number on this page was measured on data the model had never seen, and every number carries three things: the rate, how many times we checked, and the cautious bound — the level the true number is very unlikely to be below. Where we don’t have enough evidence, the page says so. That honesty is the product.

Measured · walk-forward Liquidation-magnet levels — majors (BTC, ETH, SOL, XRP)

In plain English: Our majors levels hit 86% of the time - walk-forward, out-of-sample, checked over 91,053 real windows. The cautious number (95% floor): 82.5%. That is 5.4x better than guessing (the no-model mass-weighted baseline: 15.8%).

86.1% hit rate — predicted levels that were touched

82.5% the cautious number — 95% confidence lower bound

91,053 real windows checked

Scored out-of-sample on data the model had never seen, walk-forward (only the calibration live at prediction time counts). We publish the conservative lower bound, never the flattering mean. Full derivation: /proof/stats · per-symbol table below · how we measure.

Across the whole governed pool (every served symbol, walk-forward out-of-sample): 90.0% of predicted levels were touched, honest 95% lower bound 67.7%, measured over 17 independent days. Descriptive history, not a forecast.

Every coin we track — the governed table

23 of 37 tracked symbols currently publish a rate. A rate appears only with its count and its cautious bound; symbols benched by the served floor (Wilson floor below 25.0%) show the state and the number that benched them — we do not hide failing symbols, and we do not serve you a map whose grade slipped. 14 auto-disabled symbol(s) sit below the publish floor and are grouped in the collapsed section under the table — benched, never hidden, never blended into the graded rows. Descriptive history, not a forecast.

Coin	State	Hit rate	Cautious number (95% floor)	Windows	Independent days	Edge vs no-model baseline
1000PEPE 1000PEPE → 1000PEPE levels hit 50% of the time - checked over 311 windows across 15 independent days. The cautious number (95% floor): 27.5%. That is 3.4x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving	50.1%	27.5%	311	15	+35.6pp (floor +13.0pp)
AAVE AAVE → AAVE levels hit 81% of the time - checked over 269 windows across 15 independent days. The cautious number (95% floor): 56.6%. That is 5.6x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving	81.7%	56.7%	269	15	+67.2pp (floor +42.1pp)
BNB BNB → BNB levels hit 70% of the time - checked over 283 windows across 15 independent days. The cautious number (95% floor): 45.6%. That is 4.8x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving	70.9%	45.7%	283	15	+56.4pp (floor +31.2pp)
BTC BTC → BTC levels hit 91% of the time - checked over 355 windows across 15 independent days. The cautious number (95% floor): 67.8%. That is 6.2x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving	91.5%	67.8%	355	15	+76.9pp (floor +53.3pp)
DOGE DOGE → DOGE levels hit 75% of the time - checked over 305 windows across 15 independent days. The cautious number (95% floor): 50.2%. That is 5.2x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving	75.5%	50.2%	305	15	+61.0pp (floor +35.7pp)
DOT DOT → DOT levels hit 73% of the time - checked over 228 windows across 15 independent days. The cautious number (95% floor): 48.4%. That is 5.0x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving	73.7%	48.4%	228	15	+59.2pp (floor +33.9pp)
ENA ENA → ENA levels hit 49% of the time - checked over 308 windows across 15 independent days. The cautious number (95% floor): 26.7%. That is 3.3x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving	49.2%	26.8%	308	15	+34.7pp (floor +12.2pp)
ETH ETH → ETH levels hit 100% of the time - checked over 355 windows across 15 independent days. The cautious number (95% floor): 79.6%. That is 6.8x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving	100.0%	79.6%	355	15	+85.5pp (floor +65.1pp)
HYPE HYPE → HYPE levels hit 70% of the time - checked over 343 windows across 15 independent days. The cautious number (95% floor): 44.8%. That is 4.8x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving	70.1%	44.9%	343	15	+55.5pp (floor +30.4pp)
LINK LINK → LINK levels hit 19% of the time - checked over 247 windows across 15 independent days. The cautious number (95% floor): 6.9%. That is 1.3x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving · floor watch	19.8%	7.0%	247	15	+5.3pp (floor -7.6pp)
LTC LTC → LTC levels hit 38% of the time - checked over 196 windows across 15 independent days. The cautious number (95% floor): 18.9%. That is 2.6x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving · floor watch	38.8%	19.0%	196	15	+24.3pp (floor +4.5pp)
NEAR NEAR → NEAR levels hit 65% of the time - checked over 284 windows across 15 independent days. The cautious number (95% floor): 40.5%. That is 4.5x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving	65.4%	40.6%	284	15	+50.9pp (floor +26.1pp)
NVDA NVDA → NVDA levels hit 66% of the time - checked over 94 windows across 12 independent days. The cautious number (95% floor): 38.5%. That is 4.5x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving	66.1%	38.6%	94	12	+51.6pp (floor +24.0pp)
ONDO ONDO → ONDO levels hit 83% of the time - checked over 280 windows across 15 independent days. The cautious number (95% floor): 58.7%. That is 5.7x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving	83.7%	58.8%	280	15	+69.2pp (floor +44.2pp)
PAXG PAXG → PAXG levels hit 44% of the time - checked over 139 windows across 13 independent days. The cautious number (95% floor): 22.2%. That is 3.0x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving · floor watch	44.9%	22.3%	139	13	+30.4pp (floor +7.8pp)
POL POL → POL levels hit 47% of the time - checked over 107 windows across 14 independent days. The cautious number (95% floor): 24.5%. That is 3.2x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving · floor watch	47.1%	24.6%	107	14	+32.6pp (floor +10.0pp)
SOL SOL → SOL levels hit 78% of the time - checked over 342 windows across 15 independent days. The cautious number (95% floor): 53.4%. That is 5.4x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving	78.7%	53.5%	342	15	+64.2pp (floor +39.0pp)
TRX TRX → TRX levels hit 62% of the time - checked over 148 windows across 14 independent days. The cautious number (95% floor): 37.5%. That is 4.3x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving	62.9%	37.6%	148	14	+48.4pp (floor +23.0pp)
WLD WLD → WLD levels hit 75% of the time - checked over 341 windows across 15 independent days. The cautious number (95% floor): 50.1%. That is 5.1x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving	75.4%	50.1%	341	15	+60.9pp (floor +35.6pp)
XAU XAU → XAU levels hit 65% of the time - checked over 308 windows across 15 independent days. The cautious number (95% floor): 40.3%. That is 4.4x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving	65.1%	40.3%	308	15	+50.6pp (floor +25.8pp)
XLM XLM → XLM levels hit 63% of the time - checked over 250 windows across 15 independent days. The cautious number (95% floor): 38.8%. That is 4.3x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving	63.5%	38.8%	250	15	+49.0pp (floor +24.3pp)
XRP XRP → XRP levels hit 100% of the time - checked over 330 windows across 15 independent days. The cautious number (95% floor): 79.6%. That is 6.8x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving	100.0%	79.6%	330	15	+85.5pp (floor +65.1pp)
ZEC ZEC → ZEC levels hit 76% of the time - checked over 347 windows across 15 independent days. The cautious number (95% floor): 51.5%. That is 5.2x better than guessing (the no-model mass-weighted baseline: 14.5%).	Serving	76.9%	51.6%	347	15	+62.4pp (floor +37.1pp)

Auto-disabled (below publish floor) — 14 symbol(s). These coins fell under our auto-disable bar and are benched from serving; their rows keep the exact measured state and counters, grouped here so they never blend into the graded table.

Coin	State	Hit rate	Cautious number (95% floor)	Windows	Independent days	Edge vs no-model baseline
AAPL AAPL → Benched by the served-floor governance. The number that benched it is public: at-exclusion floor 3.1% over 13 independent days. Re-inclusion is earned, not assumed: floor ≥ 30.0% held over consecutive confirmed cycles plus sustained time. Live status: accruing.	Excluded · accruing	—	—	82	13	—
ADA ADA → Benched by the served-floor governance. The number that benched it is public: at-exclusion floor 23.5% over 15 independent days. Re-inclusion is earned, not assumed: floor ≥ 30.0% held over consecutive confirmed cycles plus sustained time. Live status: accruing.	Excluded · accruing	—	—	318	15	—
AVAX AVAX → Benched by the served-floor governance. The number that benched it is public: at-exclusion floor 15.9% over 15 independent days. Re-inclusion is earned, not assumed: floor ≥ 30.0% held over consecutive confirmed cycles plus sustained time. Live status: accruing.	Excluded · accruing	—	—	226	15	—
COIN COIN → Measured continuously; the evidence is still below our publish floors (7 independent days, 21 graded windows so far). The counters are public; the rate is not, yet.	Accruing	—	—	21	7	—
HOOD HOOD → Measured continuously; the evidence is still below our publish floors (7 independent days, 24 graded windows so far). The counters are public; the rate is not, yet.	Accruing	—	—	24	7	—
MRVL MRVL → Benched by the served-floor governance. The number that benched it is public: at-exclusion floor 29.8% over 13 independent days. Re-inclusion is earned, not assumed: floor ≥ 30.0% held over consecutive confirmed cycles plus sustained time. Live status: accruing.	Excluded · accruing	—	—	113	13	—
MSTR MSTR → Benched by the served-floor governance. The number that benched it is public: at-exclusion floor 18.1% over 10 independent days. Re-inclusion is earned, not assumed: floor ≥ 30.0% held over consecutive confirmed cycles plus sustained time. Live status: accruing.	Excluded · accruing	—	—	55	10	—
SNDK SNDK → Benched by the served-floor governance. The number that benched it is public: at-exclusion floor 10.2% over 15 independent days. Re-inclusion is earned, not assumed: floor ≥ 30.0% held over consecutive confirmed cycles plus sustained time. Live status: accruing.	Excluded · accruing	—	—	337	15	—
SOXL SOXL → Benched by the served-floor governance. The number that benched it is public: at-exclusion floor 8.6% over 15 independent days. Re-inclusion is earned, not assumed: floor ≥ 30.0% held over consecutive confirmed cycles plus sustained time. Live status: accruing.	Excluded · accruing	—	—	273	15	—
SUI SUI → Benched by the served-floor governance. The number that benched it is public: at-exclusion floor 43.8% over 15 independent days. Re-inclusion is earned, not assumed: floor ≥ 30.0% held over consecutive confirmed cycles plus sustained time. Live status: accruing.	Excluded · accruing	—	—	281	15	—
TAO TAO → Benched by the served-floor governance. The number that benched it is public: at-exclusion floor 18.2% over 15 independent days. Re-inclusion is earned, not assumed: floor ≥ 30.0% held over consecutive confirmed cycles plus sustained time. Live status: accruing.	Excluded · accruing	—	—	241	15	—
TSLA TSLA → Benched by the served-floor governance. The number that benched it is public: at-exclusion floor 19.0% over 13 independent days. Re-inclusion is earned, not assumed: floor ≥ 30.0% held over consecutive confirmed cycles plus sustained time. Live status: accruing.	Excluded · accruing	—	—	94	13	—
UNI UNI → Benched by the served-floor governance. The number that benched it is public: at-exclusion floor 24.4% over 15 independent days. Re-inclusion is earned, not assumed: floor ≥ 30.0% held over consecutive confirmed cycles plus sustained time. Live status: accruing.	Excluded · accruing	—	—	277	15	—
XAUT XAUT → Benched by the served-floor governance. The number that benched it is public: at-exclusion floor 25.1% over 15 independent days. Re-inclusion is earned, not assumed: floor ≥ 30.0% held over consecutive confirmed cycles plus sustained time. Live status: accruing.	Excluded · accruing	—	—	172	15	—

The independent unit is the UTC DAY, not the correlated intra-day window — the cautious bound is computed on held-out day clusters so the interval never reads tighter than the evidence supports. Edge = hit rate minus the governed pool's mass-weighted no-model baseline, with its own 95% floor; per-coin pages: /coins/<SYMBOL> · full governed JSON: /proof/stats.

How we measure — in plain English

1. We predict, then we wait. The map publishes predicted liquidation levels. Later, the real tape tells us which levels were actually touched. Nothing is scored on data the model had already seen.

2. Walk-forward only. A prediction counts only against the calibration that was live when it was made — never a refitted after-the-fact model. That is what “out-of-sample” means here.

3. Every number carries its count and its worst case. Next to every rate you see how many windows or independent days it rests on, plus the “cautious number”: a 95% confidence lower bound — the level the true rate is very unlikely to be below. We publish the floor, never the flattering mean.

4. Honest zeros stay visible. Things without enough evidence say UNVALIDATED or accruing, with the counters public. Symbols whose grade slips are benched and shown as benched. We would rather show you a gap than a guess.

The frozen metric contract: /methodology · the full scored ledger: /proof · machine-readable: /proof/stats · cryptographic anchor chain: /proof/anchors. Everything on this page is descriptive history, not a forecast, and not financial advice.

Directional calls — the honest state: accruing

A separate experiment tries to call DIRECTION (up / down / sideways), pre-registered so the rules cannot be tuned to the result. It publishes a number only after the pre-registered gate clears (enough resolved calls, enough days, and a cautious bound that beats guessing). Until then the state is UNVALIDATED and the counters are the disclosure — evidence in progress, never a signal.

UNVALIDATED · accruing Engine Read directional stream (rule registered 2026-07-26T17:20:00Z)

0 resolved calls so far (need 60)

0 days so far (need 14)

0 calls emitted, last 30d (of 28,391 evaluated)

Abstentions are disclosed, never graded; a realized sideways outcome counts as a miss. Descriptive history, not a forecast.

BETA · measured, edge not proven Shadow directional stream — the graduation track

10.1% hit rate on resolved shadow calls

9.0% 95% lower bound

2,415 resolved calls over 17 days

82.0% do-nothing base rate, same calls (floor 80.4%)

BETA — enough resolved evidence to publish the track record with its bound; the beat-the-base-rate gate has NOT cleared (no edge claim). The graduation gate that remains: the cautious bound of the hit rate (9.0%) must beat the cautious bound of doing nothing (80.4%) — not yet cleared.

Watch the validation run live — next rung: VALIDATED

resolved calls

2415 / 120

resolved days

17 / 28

At 120 resolved calls and 28 days the stream is re-measured against its pre-registered bar — pass or fail, we publish the result. Zero is a real state: the counters ARE the disclosure.

A never-served side ledger; the ladder is pre-registered (UNVALIDATED → BETA needs 60 resolved calls over 14 days → VALIDATED needs 120 over 28 and a cautious bound that beats doing nothing — the realized sideways base rate over the same calls, never just beating zero). Verdict computed 2026-08-16T20:19:02Z by the accuracy loop; this page renders that exact block. Descriptive history, not a forecast.

The 3-class number — read this one carefully

Sideways-base-rate number — not directional edge Engine Read 3-class hit rate (1h horizon, trailing 7 days)

Our engine currently calls sideways 100.0% of the time, and the market actually moved sideways 87.5% of the time in this window. So the 87.5% below mostly measures how sideways the market was — it is a sideways-base-rate number, not directional skill. The honest check is the residual (our rate minus the base rate): +0.0pp here. Directional calls emitted in the window: 0 — directional accuracy is UNVALIDATED and we say so rather than dress this number up as proof of foresight.

87.5% 3-class hit rate (= the sideways base rate while all calls are sideways)

86.8% 95% lower bound

8,973 scored windows

+0.0pp residual over doing nothing

The per-regime residual check (does the engine beat always-saying-sideways in any volatility regime?) has not been computed yet — the accuracy loop writes it each round. We don’t estimate it.

Measured 2026-08-16T20:19:02Z · descriptive history, not a forecast.

Discipline advisories — walk-forward record

Measured · descriptive Advisory correctness, trailing 30 days

The product's discipline advisories (e.g. “hold”) are scored after the fact: 45.4% proved correct, 95% lower bound 36.7%, over 119 independent advisory windows in the trailing 30 days. Independent means de-duplicated — the raw repeat count runs about 1.0× higher from re-emissions of the same open position, and we never count those as independent evidence. Descriptive history, not a forecast.

Reference-desk advisories — no open position needed

Hypothetical · measured Reference watchlist (BTC/ETH/SOL/XRP/DOGE/BNB/ADA/AVAX/LTC/NEAR/DOT/TRX/XLM/AAVE/POL), trailing 30 days

The same discipline engine also scores a fixed reference watchlist as hypothetical advice — reference-desk units with no open position behind them, graded by the identical 4-hour rule as the position advisories above. 58.6% proved correct, 95% lower bound 55.9%, over 1,349 graded reference windows in the trailing 30 days. This block is never blended into the position-advisory record — it has its own count and its own basis. Descriptive history, not a forecast.

58.6% reference advice proved correct

55.9% 95% lower bound

1,349 graded reference windows

30 trailing days

The full public track record of this stream: /reference-desk — hypothetical, no capital, measured.

Is the underlying tape real? — independently re-checked

Every few minutes we re-verify our own liquidation tape against independent exchange references. Match rate = how often our tape agreed when a check was possible; coverage = how often a check was possible. Both shown with the check count; neither hides the other.

Venue	Match rate	95% floor	Coverage	Checks (7d)
binance	100.0%	100.0%	99.9%	50,400
bybit	100.0%	100.0%	100.0%	50,400
deribit	99.0%	98.9%	100.0%	30,240
deribit: measured since 2026-07-29 (REST capture era) — before that, the feed structurally couldn’t see liquidations (no data (capture blind 2023-10→2026-07)), and we say so.
kraken	100.0%	100.0%	100.0%	50,400
okx	100.0%	99.9%	100.0%	50,400

See it live

The same measured levels, in real time, inside the Money Map — free tier available.

Open the Money Map Full proof ledger → Start free →

Data freshness — levels calibrated 2026-08-16T20:47:21Z · ledger built 2026-08-16T18:49:02Z · engine 3-class measured 2026-08-16T20:19:02Z · advisories scored 2026-08-16 20:50 UTC · X/Twitter news corpus PAUSED (API credits depleted since 2026-08-12 01:44 UTC) — an honest gap, disclosed