Cascade clustering ≥ 0.6
feature.cascade_clustering_index
Read at 1 minute before the move.
This exact combination appears nowhere else in the 30 days searched. The minute it was read from is the only one on record.
q#01dbb12b @ dataset 461da594
PINNED INVESTIGATIONSee this market’s latest move
CYSUSDT fell 29% into Jul 29. The record covers the whole lead-up.
Move window: 2026-07-28 15:00 to 2026-07-29 08:00 UTC. The record for CYSUSDT currently ends 2026-08-04 00:00 UTC.
Counted on 2026-08-05, against a record covering to 2026-08-04 00:00 UTC. These numbers are frozen: this page will always say what it said the day it was written.
Book imbalance · top levels ≥ 0.65 and Cascade clustering ≥ 0.6 and Taker buy share · 15m ≥ 0.65
feature.book_imbalance + feature.cascade_clustering_index + feature.taker_buy_ratio_15m
Read at 1 hour before the move.
Minutes like this are common here: 355 of the 43,201 minutes searched (0.8%), about 11 a day. What followed was close to a coin flip: 164 rose and 166 fell over the next 30 minutes (330 of the 355 matches have a complete 30 minutes forward).
Of the 355 fires, 287 came before the move began and 6 happened while it was already under way. A fire during the move sees it; it does not telegraph it.
About 11 a day, on 30 of 31 days: spread through the record rather than one event.
q#ccd7bf88 @ dataset 461da594
Best supported: Book imbalance · top levels ≥ 0.65 and Cascade clustering ≥ 0.6 and Taker buy share · 15m ≥ 0.65, the combination with the most fires before the move began.
Opens this exact search in the research tool. Nothing runs until you confirm it there, and arming it as an alert sends a message when it next becomes true.
Picked in hindsight: the counts include every other time these readings appeared, not just this move.
What the minutes before it looked like
Hourly candles. The marked lines are the minutes the readings came from: 1 minute, 15 minutes, 1 hour, 4 hours and 24 hours before the move began. Every one of them sits before the move, never inside it.
A reading stands out when it sits outside the range it usually holds. The rules are fixed: the same ones the moment reader on the research page uses.
At 1 minute before the move, 1 of the 39 readings stood out. 2026-07-28 14:59 UTC
| READING | BEFORE THE MOVE | SEARCHES FOR |
|---|---|---|
| Cascade clustering feature.cascade_clustering_index | 0.6089 | ≥ 0.6 |
At 15 minutes before the move, 1 of the 39 readings stood out. 2026-07-28 14:45 UTC
| READING | BEFORE THE MOVE | SEARCHES FOR |
|---|---|---|
| Cascade clustering feature.cascade_clustering_index | 0.6089 | ≥ 0.6 |
At 1 hour before the move, 3 of the 39 readings stood out. 2026-07-28 14:00 UTC
| READING | BEFORE THE MOVE | SEARCHES FOR |
|---|---|---|
| Book imbalance · top levels feature.book_imbalance | 0.6571 | ≥ 0.65 |
| Cascade clustering feature.cascade_clustering_index | 0.6089 | ≥ 0.6 |
| Taker buy share · 15m feature.taker_buy_ratio_15m | 0.6553 | ≥ 0.65 |
At 4 hours before the move, 5 of the 39 readings stood out. 2026-07-28 11:00 UTC
| READING | BEFORE THE MOVE | SEARCHES FOR |
|---|---|---|
| Book imbalance · top levels feature.book_imbalance | 0.8229 | ≥ 0.82 |
| Compression · 15m feature.candle_compression_15m | 0.4244 | ≤ 0.43 |
| Upper wick · 15m feature.candle_upper_wick_15m | 0.7273 | ≥ 0.72 |
| Cascade clustering feature.cascade_clustering_index | 0.6089 | ≥ 0.6 |
| Realized vol percentile feature.realized_vol_pctrank | 0.909 | ≥ 0.9 |
At 24 hours before the move, 4 of the 39 readings stood out. 2026-07-27 15:00 UTC
| READING | BEFORE THE MOVE | SEARCHES FOR |
|---|---|---|
| Book imbalance · top levels feature.book_imbalance | -0.6615 | ≤ -0.66 |
| Containment · 15m feature.candle_containment_15m | inside | inside |
| Lower wick · 15m feature.candle_lower_wick_15m | 0.6667 | ≥ 0.66 |
| Cascade clustering feature.cascade_clustering_index | 0.608 | ≥ 0.6 |
Also counted
This is not a similarity search. Each count matches the exact readings above and nothing else.
Book imbalance · top levels ≥ 0.82 and Compression · 15m ≤ 0.43 and Realized vol percentile ≥ 0.9
feature.book_imbalance + feature.candle_compression_15m + feature.realized_vol_pctrank
Read at 4 hours before the move.
Minutes like this appeared only 8 times in the 43,201 minutes searched. What followed leaned up: 5 of the 8 matches with a complete 30 minutes forward rose (63%).
Of the 8 fires, 7 came before the move began; none were during it.
About 0.3 a day, on 8 of 31 days: spread through the record rather than one event.
q#a8567d14 @ dataset 461da594
Fires often arrive as clusters of consecutive minutes, and forward windows overlap. A cluster is one episode, not many independent trials.
If these counts read weaker than the move suggested, that is the tool working: the lead-up of a picked move always looks special until it is counted everywhere else.