ELCARO AI

← Terminal

Order Flow Desk — what we watch

Four perpetuals — BTC, ETH, SOL, XRP — recorded tick by tick on Bybit since 26 Sep 2026: every trade with its aggressor side, the order book once a second, open interest and funding every 5 s, every liquidation on Bybit and Binance, plus Binance's daily archives back to 15 Aug 2026. This page lists each thing we measure, where it comes from, what the test behind it showed, and what it does not mean. Nothing here is a direction call: the gauges describe the market, the register records what we tried to prove and what failed.

Connecting… ·Updated — ·Feeds answering — ·Refresh 30 s

How to read the numbers. A missing measurement shows as “—”, never as 0 — a dead feed must look dead, not neutral. Every frequency carries its sample size. Statements about who is positioned are scenarios inferred from flow and open interest, not facts: no exchange publishes who trades.

Live board

One card per coin, every gauge on this page in its current state. Taker is the 1-hour buy/sell ratio of aggressive volume; OI 24 h is open interest against price; fuel is estimated leveraged positioning within 3 % below (longs) and above (shorts) the price, with where that amount sits in the coin's own history; walls are the largest resting orders near the price.

The gauges — what each one is, and is not

For every gauge: what we measure, the source, what the evidence says, and what it does not mean. The evidence lines quote our own tests on our own recordings; the register below has the full list with the ones that failed.

Taker flowobserved

—

What: Volume traded by aggressive orders, split by side: who crossed the spread. Net taker flow per minute, hour and 4 h; the buy/sell ratio; cumulative delta (CVD).
Source: Every Bybit trade with its taker side (our collector; verified identical to Bybit's public trade archive, 100 % one-to-one match, 0 ms offset). Binance aggTrades for the cross-venue checks.
Evidence: Same-minute flow and return move together (ρ ≈ +0.5): flow explains the move that is happening. The flow state has no measurable value for the next interval (test T8, out-of-sample gain −0.05 %). Flow impact falls when the near-side book is deep (T5, Binance, p 0.001).
Not: Not “pressure that will continue”. Heavy selling tells you what just happened, not what happens next.

Order book depth & wallsobserved

—

What: Resting size within ±10 / 25 / 50 / 100 / 200 bps of the mid once a second, the spread, how much is added and pulled per second, and the largest single resting order within 0.5 % (a wall) with how far price must travel to reach it.
Source: Bybit linear order book, 1000 levels per side. On BTC that feed reaches only ~20–27 bps, so BTC's wider bands are truncated and flagged; ETH ~50 bps; SOL/XRP several per cent.
Evidence: Walls do not hold better than any random level (held-for-5-min rate ≈ the base rate of ~80 % on BTC); 59 % of walls were pulled before price reached them (study 27–30 Sep, 350 walls). Book imbalance → next 5/15 min: ρ ≈ 0. Depth adds nothing beyond volatility clustering for the next-15-min range.
Not: Not support or resistance. A wall is one order that can be cancelled in one message, and usually is.

Open interest & fundingobserved

—

What: Open interest, funding rate, mark and index every 5 s. Read against price: price down + OI up = new positions opened into the fall (fuel for volatility); price down + OI down = positions closed (deleveraging).
Source: Bybit ticker stream (OI verified exact against Bybit REST); Binance 5-min OI archive for the longer history.
Evidence: OI building against the 2-hour move raised subsequent volatility 1.5–2× the base rate on Binance Jan–Sep 2026; the direction that followed flipped between halves of the sample.
Not: Not “longs are trapped, so it will bounce” (or fall). OI says how much is at stake, not which way it resolves.

Long/short account ratioobserved

—

What: The share of accounts net long vs net short (Bybit, every 5 min) and Binance's top-trader position and account ratios.
Source: Bybit account-ratio endpoint (kept ~2 days by Bybit, archived by us since 26 Sep); Binance metrics archive (5 min, since 15 Aug).
Evidence: Descriptive only. It counts accounts, not dollars: 70 % of accounts long can coexist with net short dollar positioning. No predictive test registered yet.
Not: Not “the crowd is wrong, fade it”. We have not shown that, and we do not display it as such.

Liquidations (observed)observed

—

What: Forced closes as they happen: time, side (long or short liquidated), size, and on Bybit the bankruptcy price — the level where the position's margin would reach zero, ~0.7 % beyond where it was actually closed.
Source: Bybit allLiquidation stream (throttled by Bybit — a lower bound; all symbols ≥ $1k since 15 Aug) and Binance forceOrder (one snapshot per second per symbol; BTC/ETH/SOL/XRP). Together roughly half of the multi-exchange totals that aggregator sites print; the long/short split matches them.
Evidence: 100 % of Bybit liquidation items have a same-side trade in the archive within ±1 s (pilot QC, 24 coin-days). The bankruptcy price is not a magnet: in 17,883 events across 30 coins, price reached it first in 48 % of cases vs 50 % for a mirror level — nothing is left at that level once the position is gone.
Not: Not the cause of a move. In the sharpest minutes liquidations were 40–100 % of net selling — an amplifier that fires after the first aggressive sell, not before it.

Liquidation fuel (estimated)estimated

—

What: How much recently opened leveraged positioning sits within reach above and below the price — the same construction as the “liquidation heatmaps” on aggregator sites, shown as amounts within ±1 % / ±3 %, not as lines.
Source: Binance 5-min open interest and taker share; new positions split long/short and spread over 10/25/50/100× liquidation levels; a level is consumed when price crosses it; 7-day warm-up.
Evidence: The amount of nearby fuel ranks where real liquidations then happen (ρ ≈ 0.3, p < 0.001, checked on Bybit's real liquidations) — but a placebo that spreads the same amount evenly over 0.5–10 % scores the same, so the exact levels carry no information. “Price goes to the bigger cluster” was +20 pp in one three-week period and 0 in the one before: no direction. A forward test runs 29 Sep – 27 Oct.
Not: Not a map of real open positions — nobody has one; only Hyperliquid publishes positions. Not a target.

Long-liquidation flush ruleregistered rule

—

What: A minute whose long liquidations exceed the coin's trailing-7-day 99.5th percentile (≥ $100k). We then record whether price is higher 60 minutes later.
Source: Bybit liquidation stream; rule H_LIQ_FLUSH_REBOUND_v1, criteria frozen before the data was read.
Evidence: 22 Aug – 28 Sep: price was higher after 60 min in 67 % of 100 flushes vs 45 % in matched control minutes; both halves of the sample agree; the result survives clustering of flushes into 46 episodes (p 0.002). The live record since launch is shown above and keeps counting whichever way it goes.
Not: Not a trade: a causal replay with realistic entry after the flush, fees and spread lost money (−2 to −10 %). A higher hit rate is not a positive expectancy.

Squeeze watch (fuel building)tested flag

—

What: Over the last 2 h: price up while perp takers net sell and OI rises = shorts piling up against the move; the mirror for longs. Shown with the measured frequency of a ≥ 1.5 % move either way in the next hours.
Source: Binance perp and spot 5-min data, Jan – Sep 2026 for the frequencies; live from the same feeds.
Evidence: A flagged state is followed by a ≥ 1.5 % move about 2× as often as an unflagged one; the direction is unstable across halves of the sample. It is a volatility flag.
Not: Not “a short squeeze is coming”. Up and down are shown side by side because the data does not separate them.

Basis & cross-venueobserved

—

What: Perp price vs spot (basis, annualised), and how Binance and Bybit, perp and spot, move relative to each other.
Source: Binance spot and perp trades, Bybit perp trades; Bybit spot from the public trade archive.
Evidence: Venues co-move almost one-to-one at 1 s (max |ρ| 0.66–0.89 at lag 0). Which venue leads is not established: exchange clocks are independent and the lead-lag ratio flips by coin. Sub-second leads are never claimed.
Not: Not “Binance moves first, so trade Bybit”. That would require exchange-timestamped book data we are still building the collector for.

Composite signalrule-based

Shown in the terminal and on the signals feed; its track record is measured, not assumed.

What: The terminal's LONG / SHORT / NEUTRAL reading per coin: eight scored dimensions (technical, order flow, on-chain, macro, correlation, capital flow, valuation, news) combined into one score with a confluence count.
Source: Rules over the feeds above plus the terminal's own data; raw dimension scores are stored and re-centred against their history when combined.
Evidence: Measured continuously: hit rate by coin, by confidence band and by direction at fixed horizons, with the losses shown. Weights are retuned only against those numbers. In ranging markets it flips often; a single caught move is not evidence.
Not: Not AI, not a forecast model. Rules with a published track record, nothing more.

Anatomy of a cascade — what the gauges show during a fast drop

A worked example from our own recordings: SOL on 7–8 Oct 2026, three legs of a −5.7 % day (low −9 %) on which all four coins fell and long liquidations across Bybit reached $120M, the largest day since recording began (median day ≈ $12M). Minute by minute, each leg looked the same.

1An aggressive sell starts it. One minute of net taker selling of $10–16M with zero liquidations in that minute. Price −0.1 … −0.6 %.

2Liquidations amplify it. In the next one or two minutes forced closes arrive; in the sharpest minute they equalled the entire net selling ($7.4M liquidated vs $6.9M net taker flow).

3The book is pulled at the same time. Bids within 50 bps halve (from $16–18M to $8–10M) and the pull rate jumps 5–7×, so the same selling moves price three to five times further. Depth was not thinner before the leg — it thinned with it.

4Why it did not stop. Between legs the book refilled and price bounced +0.5–0.9 %, but open interest kept rising (+9 % on the day while price fell) and the share of long accounts rose: each step down was bought with new leverage, which became the fuel for the next leg. The first real deleveraging came only on the third leg.

Leg (UTC)	Trigger minute	Liquidations next minute	Bids ±50 bps	Move in 5 min
7 Oct 02:00	−$9.7M net sell	$3.2M (Bybit + Binance)	$14M → $7.9M	−1.3 %
8 Oct 04:19	−$11.3M net sell	$1.7M	$17.6M → $9.8M	−0.7 %, bounce +0.9 %
8 Oct 15:11	−$15.8M net sell	$7.4M in one minute (15:23)	$16M → $7.7M	−3.1 % in 25 min

What this does and does not establish: the mechanics are reproducible and explain the size of the move; nothing in the recordings before the first leg distinguished that night from the range of the day before. Explanation after the fact is not a forecast, and we do not present it as one.

Data sources

What is recorded, at what resolution, since when, and the known limit of each feed. The collector's own health log is checked before any window is analysed; its trade table was verified against Bybit's public archive on 24 coin-days (0 defects, losses only at reconnects, ≤ 28 rows per day).

Feed	Venue	Resolution	Since	Known limit
Trades with taker side	Bybit perp	every trade, ms	2026-09-26	No trade id in the live stream; order inside one millisecond unknown. Public archive (0.1 ms, ids) used for QC.
Order book	Bybit perp	1 s summary · full book every 60 s	2026-09-26	Rows carry our receive time, not the exchange's: physical order of book vs trade events is not identifiable. BTC depth reaches only ~25 bps.
Open interest, funding, mark, index	Bybit perp	5 s	2026-09-26	Verified exact against REST.
Liquidations	Bybit (all symbols) · Binance (4 coins)	event · 1 s snapshot	2026-08-15 · 2026-09-26	Bybit throttles its stream (lower bound). Binance publishes no liquidation history. Other exchanges not collected.
Long/short account ratio	Bybit · Binance top traders	5 min	2026-09-26 · 2026-08-15	Counts accounts, not dollars.
Daily archives: aggTrades, book depth ±1–5 %, metrics	Binance perp + spot	trade · 30 s · 5 min	2026-08-15	Book-depth archive 3–10 Sep 2026 is defective (ask side) and excluded.
Public trade archives	Bybit perp + spot	0.1 ms, trade ids	pilot 2026-09-26 → 10-01	Trades only — no book, no sequence numbers. Expansion to Aug 15 pending review.

Research register

Every question we tried to answer on this data, with its criteria written down before the data was read, and what it returned. Failures stay on the page. “Exploratory” means the sample was too short to confirm anything and the result is a candidate for a frozen forward test, not a finding.

Question	Result	Status
After a burst of long liquidations, is price higher 60 min later more often than in matched control minutes?	Yes: 67 % vs 45 %, n 100, both halves, survives episode clustering (p 0.002). A realistic replay still loses after costs.	passed · live
Does an estimated liquidation map say where liquidations happen, and does price go to the bigger cluster?	Amount within reach: yes (ρ 0.3), but a placebo with evenly spread levels scores the same. Magnet: +20 pp in one period, 0 in the other.	amount only · forward test to 27 Oct
Is the bankruptcy price of a liquidation a magnet?	No: reached first in 48 % vs 50 % for a mirror level (17,883 events, 30 coins). Slight repel, below the bar.	no effect
Price up + perp takers selling + OI up: does a squeeze follow?	A ≥ 1.5 % move follows 1.5–2× more often; its direction flips between halves. Volatility flag, not a trade.	volatility only
Do large resting orders (walls) hold when reached?	No better than any level (base rate ~80 % on BTC); 59 % are pulled before price arrives.	no effect
Does book imbalance or depth predict the next 5–15 minutes?	Imbalance → direction: ρ ≈ 0. Depth → range: nothing beyond volatility clustering.	no effect
Mechanism tests T1–T15 (5 days Bybit, 48 days Binance): does flow impact depend on book state, and does state forecast the next interval?	Impact falls with near-side depth (T5); state explains +3.1 % more of the current move out of sample (T7); state has no forecast value for the next interval (T8); the imbalance “lead” over the mid is a timing artifact (T13); venue leads unestablished (T14/T15). Most Bybit tests underpowered.	exploratory
Does price retrace to the origin of an impulse? Does a liquidity profile predict direction? (five directional studies)	No registrable rule found; one study was not even specifiable without look-ahead. Closed.	closed
Frozen forward tests	Order-book study (run ≥ 18 Oct 2026), liquidation-map forward test (28 Oct), flush rebound pre-registration (10 Oct 2026 → Apr 2027), fuel build-up v2 (to Jan 2027). Nothing is read before its date.	pending

Our rules

Pre-registrationExposure, outcome, sample, success threshold and the test are written down and committed before the data is read. Changes are append-only amendments.

Frozen windowsDays reserved for a forward test are never used to explore. A rule that was tuned on a window cannot be confirmed on it.

No silent zerosA dead feed shows “—” or an error, never a number that looks like data.

Explanation ≠ forecastShowing why a move was as large as it was is not the same as predicting the next one. We label which is which.

Positioning is a scenarioFlow and open interest do not reveal who trades, their entries or their stops. Such statements are marked as scenarios.

Success means net expectancyA significant probability of a move is not a result. Only a positive expectancy after latency, spread, fees, slippage and adverse selection counts — simulated under stated costs, with intervals.