Pick a level above the price, a level below it and a deadline. The table says how often the upper one was reached first, how often the lower one was, and how often neither happened in time — counted from every hour of the tape.
42.6% as of the last day in this build, snapshot taken Oct-23 2026, measured and published by askthetape.com.
Pick a level above the price and a level below it, and a deadline. This page answers with THREE numbers, because the question has three answers: the upper level came first, the lower one came first, or neither happened in time — and the three shares add up to 100 % before rounding. BTC sits at $81,850 (last closed hour, 2026-10-23T05:00Z; the page recomputes every cell from the live price and prints the price and hour it used). Ask it for $90,000 (+10.0%) before $70,000 (−14.5%) within 30 days and, counting from every hour of the last 2 years, the upper level came first on 42.6% of the 16,801 starts, the lower one on 20.1%, and on 37.2% neither level was reached at all. The 95% band on the first number is 27.0–59.2 — and that band describes THIS window's sample, not other epochs: the same cell read 56 % / 55 % / 59 % in the three previous two-year windows. That third share is the one usually left out, and inside a short deadline it is often the biggest: on the same pair over 24 hours it is 99.5%. Shorten the deadline and «nothing happened» eats both winners; stretch it to a year and almost everything prints eventually, which is a different question — the daily long view since 2011 under the window toggle answers that one, and its numbers are not to be mixed with a 30-day answer. Read the band, not the point: starts an hour apart share almost all of their future, so the honest answer is a range from a block bootstrap built for that overlap. And the price is not a detail: these are dollar levels measured from a price, and a 2% move in it shifts a cell by 4 to 25 points.
| definitions | a start is every hour's close inside the window; a start counts only once its full deadline has already elapsed (the youngest 720 starts of this window are held back for exactly that reason). «Upper first» = the first later hour whose HIGH reaches the upper level comes before the first hour whose LOW reaches the lower one; both levels inside ONE hour count as upper first — the watch's own convention — and that count is printed per cell (0 here, 0.0 % of starts). The three shares are of ALL counted starts and add to 100 % before rounding. |
|---|---|
| provenance | Binance BTC/USDT, hourly, 2017-08-17T04:00Z → 2026-10-23T05:00Z — 80,370 hourly bars; engine which_first_hourly.py, recomputed every day at 06:00 UTC, read by this build at 2026-10-23T06:08:51Z. Binance's hourly history is missing 128 hours in 28 outages on BTC and ETH (SOL 20 hours in 10) — none inside the default two-year window; in the since-listing window a 30-day horizon can stretch to 31.4 days across an outage. |
| 🩸 the deadline is the question | the same pair over 24 hours, 72 hours, 7, 14 or 30 days is five different questions: the shorter the deadline, the larger «neither» (here 99.5% at 24 hours against 37.2% at 30d). Quoting a base rate without its deadline is the most common way to be wrong with true numbers. The daily long view (365 days, Bitstamp since 2011) answers the multi-year version of the question and reads up to 20 points higher on the same pair — different horizons, not a contradiction; the page prints both numbers side by side when you switch to it. |
| 🩸 the band, and what it does not cover | starts an hour apart overlap almost entirely, so the count is not the sample: the 95 % band beside every number is a block bootstrap with blocks of three horizons, 4,000 replicas (27.0–59.2 on the headline cell). Since 13.09.2026 the engine bootstraps all THREE shares from the same replicas, so the band printed under a cell is the band of whichever share led it; the daily long view below still carries a band on the upper share only, and says so. It describes the sample of THIS window and nothing else: the same cell read 56 % / 55 % / 59 % in the three previous two-year windows, which is a spread the band does not contain. We do NOT print an «effective sample size»: the usual starts÷median-time formula was re-measured against this very band on 12.09 and disagreed by more than 1.5× on 69 % of cells, so it would have been a number arguing with the band printed next to it. |
| 🩸 the live price, not the anchor | a dollar level is a distance from the price, and that distance moves all day: a 2 % move shifts a cell by 4 to 25 points (measured 12.09: 4.4 pp on 95k/75.5k, 10.7 on 85k/65k, 24.4 on 79k/73k). The engine runs once at 06:00 UTC and anchors at $81,850; the page therefore recomputes every cell from the LIVE price through the percentage grid and prints the price and hour it used. Without JavaScript you get the anchor table instead, stamped with its own hour. |
| 🩸 ties inside one hour | when both levels print inside the same hour the order is unknown, and the watch's convention counts it as upper first. Where that class is large (3 %% of starts or more) the page prints a RANGE instead of a point — the share without ties on one end, with them on the other. On today's round levels ties appear in 100 of the 735 published rows, up to 141 starts on near levels and short deadlines. |
| 🩸 the epoch matters | symmetric ±10 % inside the same deadline: 2 years 41.4% (band 26–59) · since 2017 48.5% (band 43–55). The default window is the one the watch uses (2 years); the longer ones carry older, different markets, and the three earlier two-year readings printed beside every band say how far apart those epochs sit. |
| what the watch does | this is the same instrument the watch answers with when an author names two levels and a deadline — hourly bars, the window it uses, three shares. Of five real questions it published on 12.09, three were reproduced exactly at the anchor hour and two within one point; at 18:00 UTC the page reads those pairs 2–3 points lower after a −0.3 % move in price, which is the same reason the live price is the default here. |
| 🩸 where this is thinner | one semantics (touch = the hour's high or low), dependent starts, and a band that covers this window's sample but not the difference between epochs. Round levels are chosen around the anchor price, so the matrix changes shape as the price moves; the percentage grid does not. Level percentages are printed rounded for the eye; the engine counts on the unrounded ones. Coins beyond BTC, ETH and SOL are not covered. |
| 🩸 what a resampled walk reproduces, and what it does not | we tried to beat the grid with a walk resampled from its own bars. Draw the hours independently — no memory at all — and it does not reproduce it: the median miss on «neither» is 2.8 points (up to 13 on far pairs), «upper first» is out by up to 9 points at 24 hours, and only 516 of the 605 cells land inside the printed band; Gaussian hours with the same volatility and a true within-hour path land 457–506 across 5 seeds. Keep one day of memory and the miss shrinks without vanishing: 24-hour blocks bring the median miss on «neither» to 1.2 points, with 571 of the 605 «neither» cells inside the band (worst still 10.6 points out); «upper first» lands 594–605 of the 605, seed depending. Read that next to the band's own width: on a 30-day deadline it is ±8.4 points wide at the median for «upper first» and ±9.0 for «neither», so no smaller edge is visible here. Two limits: on 24-hour deadlines the path is literally a stretch of the same tape — 121 of the 605 cells answer themselves, the other 484 do not — and the 605 are not 605 independent tests: the levels nest and the starts overlap. The test is not blind: bars carrying 1.5× the volatility (ETH) land only 137 cells in the band. Nor is it a law: on the previous two-year window the same resampling lands only 385 of the 605 inside today's band — the band belongs to this window's volatility. Within this window the grid holds nothing beyond the two distances, the deadline and one day of volatility memory. It prices the question; it does not answer it. Measured 2026-09-14 against the grid this engine built at 2026-09-14T05:00Z. |
| 🎯 how far an interpolated cell sits from the exact count | Any pair that is not on the 11×11 grid is interpolated — the page says so, and this row says by how much. We counted 400 random pairs twice over two years of hourly starts: exactly, with the same code that builds the grid nodes, and the page's way, against deadlines of 24 hours, 3 days, 7 days, 14 days and 30 days — 2,000 cells in all. Inside the grid the interpolation is honest: it misses the exact count by 0.35 points at the median, 2.0 at the ninetieth percentile, 3.7 at the worst. The damage is at the edge. The closest level the grid carries is 2 % away, and anything closer is not interpolated at all — it is served the 2 % node, and 770 of the 2,000 cells we counted landed there: on those the miss is 11.8 points at the median, 32.5 at the ninetieth and 62.5 at the worst. Each of those three is the largest of the three outcomes' OWN statistic — upper first, lower first, neither — so the triple is a boundary across statistics, not the worst cell and not one distribution. By deadline the node costs differently: 24 hours 24.1 pp · 3 days 12.0 pp · 7 days 11.9 pp · 14 days 11.6 pp · 30 days 12.0 pp at the median (and 49.4 pp · 33.4 pp · 28.3 pp · 27.3 pp · 26.8 pp at the ninetieth). This is not a corner case: the levels are rebuilt once a day at 05:00 UTC while the price moves all day, and over the last 30 days at least one cell of the matrix was that node in 51.6 % of hours (51.7 % over a year) — a median of 7 cells, and the matrix itself runs 49 to 64 cells as the price moves, so 12.5 % of it. Only a band of price $719 wide — 0.88 % — leaves every cell inside the grid, and a 0.27 % move from the anchor puts a whole row or column on the node. On round-dollar levels, which is what a reader types, 61 % of the pairs we drew near the price landed there. With both legs within 5 % of the price, node or not, it is 4.2 / 25.5 / 58.9; beyond 10 % it is 0.13 / 1.22 / 3.0. A worked example from 2026-10-23: $80,000 against $76,000 within 3 days, priced off an anchor of $78,946 — the grid printed 52.1 / 26.5 / 21.4, the exact count of the very same tape said 66.6 / 20.8 / 12.6, a 14.5-point miss on the upper level. That sits next to the band, and they are not the same doubt: the band is about the sample, this is about the grid. On the same 30-day deadline the band is ±8.4 points at the median — 16.8 points across — and a pair that lands on the node misses the exact count by 26.8 at the ninetieth, more than the band's whole width; beyond 10 % of the price it is 1.2. Measured 2026-10-23 against the grid this engine built at 2026-10-23T05:00Z, and re-counted on 3 control pairs by a second implementation that shares no code with the first. |
| policy | no backfill · no export · read it here · a base rate is what happened, not what will — we never call direction · read on the site · not exported: this page has no daily series to put in a file |
| this build | data to 2026-10-22 |
| policy | no backfill · no export · read it here · a base rate is what happened, not what will · read on the site · not exported: this page has no daily series to put in a file |
The chart itself is drawn in the browser: turn JavaScript on to hover a day, pick a window and draw your line. The numbers above come from the last build, Oct-23 2026 21:24Z.
What the tape says today · what's new · all 111 metrics