AXIS Desk · NQ microstructure · tick-level event study
Large Orders on NQ
What actually happens after a big aggressive order hits the Nasdaq futures book — the moment it lands, the next few seconds, and the next thirty minutes — measured on every sweep of the last 5.7 years, and split by the environment it landed in.
2020-11-27 → 2026-08-141,472 RTH sessions3.68M sweeps scored · 56,401 of 100+ lotsmeasured on the mid, from the post-sweep bookawareness lens · not an entry study
The five questions, answered
When a big buy comes in, do we move down first to fill it with more sellers, then deliver?
The dip is real; the delivery is not. After a 100–499-lot buy sweep, the mid gives back a point (4 ticks) before it adds a point about 50% vs 31–38% of the time — usually within 1–3 seconds, and it is the book refilling, not new sellers being sourced. From that dip, price is back above the post-sweep level at 15 minutes only 47–48% of the time. Big buyers do not, on average, get "delivered."
Do big sellers get the mirror image?
Yes — the tape is symmetric. Buy and sell sweeps of the same size print the same impact (10.2 vs 11.0 ticks for 100–199 lots), the same give-back and the same flat drift. No side has an edge on NQ over 5.7 years.
Do they come in and immediately trade in their favor?
Only by their own impact. A 100–199-lot sweep moves the mid 10.6 ticks (2.6 pts, ~13 ticks in the thinner 2025–26 book), 200–499 lots ≈ 20 ticks, 500+ ≈ 33 ticks — and that impact is retained (98–104% at 1–15 min). Beyond it, the average forward move is ≈ 0 ± 1 pt at every horizon to 30 minutes, against a 15-minute outcome spread of ±44 pts. The aggressor pays the impact and the market goes on as before.
What differs in the environment when the results differ?
Not the direction — the width and the give-back. Time of day, prior 5-min move, vol regime, clustering, side and year all leave the forward mean inside ±1.5 pts (and small orders show the same values, so size gets no credit). Three things do move the aftermath: how much the sweep moved the book (absorbed → continuation-first; blew-through → pullback-first, stable every year 2021–26), whether it printed a fresh session extreme (−1 pt in 5 min, size-specific, 5 of 6 years), and on the 241 depth days, book balance (thin ahead + thick behind → +4–5 pts over 15–30 min; thin both sides → wide and slightly negative).
What can we do with that?
Treat a big print as a liquidity and location fact, not a direction fact. Read the impact (thick or thin book?), read where it printed (fresh extreme?), and read what regime it landed in (how wide the next 15 minutes will be). Don't chase it, don't fade it. Rules and evidence levels are at the bottom.
What "big" means on this tape
A sweep is one aggressor's fills grouped: consecutive same-side prints within 1 millisecond (Sierra appends a sequence number per millisecond, so a market order that walks several resting orders and levels shows as N records 1 µs apart). Sizes below are per sweep, RTH 09:30–16:00 ET, and — unlike single-print size, which quadrupled since 2020 — the sweep-size distribution is stable across all 23 contracts (RTH p99.9 sat between 41 and 50 lots in every one).
47
lots = top 0.1% of RTH sweeps
p99 = 14 · p99.99 = 121
32
sweeps of 100–199 lots per session
~1 every 12 minutes
6.1
sweeps of 200–499 lots per session
~1 an hour
0.3
sweeps of 500+ lots per session
one every ~3 days · n=446 total
150
sweeps of 50–99 lots per session
"large" on most footprint settings
Two morphologies exist and matter later: the sweeper (many small fills across many levels — a market order eating the book; 81% of 100+ events, mean 15 levels) and the block (one print carrying ≥70% of the sweep at one level — a large resting order was matched; 5% of events, mean impact only 3.4 ticks).
1 · The moment it lands: impact scales, and it sticks
Impact is the change in the mid from the book before the sweep to the book after it, in the aggressor's direction. It grows with size but sub-linearly (per-lot impact falls from 0.16 ticks at 10–24 lots to 0.056 at 500+), and it does not snap back: the level of the mid 1, 5 and 15 minutes later still sits at 98–104% of the impact for every tier from 100 lots up.
Impact of a sweep, by size (RTH, 2020–26)
Mean mid move caused by the sweep itself, in ticks (0.25 pt), with the share of that move still present 5 minutes later.
1–3-lot "control" prints are the baseline sample; tiers 10–24 are subsampled to 60k per contract for the path loop, all larger tiers are complete. Retention = (impact + drift at t) ÷ impact, aggregate means.
The book is thinner in points, the same in percent
Impact of a 100–199-lot sweep by year — ticks vs basis points.
8.5 ticks in 2021/2023 → 13.4 ticks in 2026 while price doubled; in bps it holds ~1.5 (2.2 in the 2022 vol year). Percent-normalise, as everywhere on this desk.
Where the environment does act: width, not direction
Median favourable and adverse mid excursion in the 15 minutes after a 100+ sweep, by time of day (ticks).
The first half hour travels ~2× the midday session in both directions; the mean drift stays within ±3 ticks throughout.
2 · Then what? On average, nothing
Measured from the post-sweep mid, so the sweep's own impact is not counted as "continuation." The picture: a one-to-two-tick give-back inside the first minute for the biggest sweeps (the book refilling behind them), then a flat line. Only 500+ lots show a lean toward continuation (+8 ticks at 5 min, +16 at 30 min) — with n=446 and standard errors of 6–11 ticks, that is a hint, not a fact.
Mean forward move after the sweep, in the aggressor's direction
Ticks vs seconds since the sweep (log time). Bands are ±1 standard error. Small orders' own tiny positive drift is the informed-flow baseline.
1 pt = 4 ticks. Pre-sweep momentum is large — 100–199-lot sweeps arrive after +8 pts in their direction over the prior 5 minutes, 500+ after +14 pts — which is why the level split in section 4 needs a small-order control.
The individual outcome is noise around that zero
Distribution of the 15-minute forward move after 100+ lot sweeps (ticks, ±190 shown).
Mean +0.5 tick, standard deviation 175 ticks (44 pts); 10th/90th percentile −181 / +183 ticks. Any single big print tells you almost nothing about the next 15 minutes' sign.
3 · Does it come back first? Sequence, then delivery
For each sweep: which came first, a move of k ticks against the aggressor or k ticks in their favour, within 15 minutes — and given the sequence, where price sat at 15 minutes. At 1 pt the answer is about the first seconds (median time to the 1-pt dip after a 200+ sweep is 1 second; to the 1-pt continuation, 2–3 seconds). At 4 pts it is about the next few minutes.
Which came first, and did it deliver?
Share of sweeps by sequence within 15 minutes. Left block: k = 4 ticks (1 pt). Right block: k = 16 ticks (4 pts).
dip first, then delivereddip first, never delivereddelivered first, then dippeddelivered first, never dippedsame second
"Dip" = k ticks against the aggressor from the post-sweep mid; "delivered" = k ticks in favour. Absorbed / blew-through = bottom / top quartile of the sweep's own impact among 100+ lot sweeps.
Read on the user's own question. For a 100–199-lot buy the sequence dip → deliver (1 pt) happens 46% of the time vs 38% for deliver → dip; for 200–499 lots it is 51% vs 31%. So yes: the more common order of events is that price comes back against the big buyer first. But the second half of the story fails — after that dip, price is above the post-sweep level at 15 minutes 47–48% of the time (52–53% after a deliver-first). Whichever way it went first is mildly where it still is; there is no delivery premium for the big order. Small 1–3-lot prints show the same 48/52, so even that is just short-horizon autocorrelation.
4 · When the outcomes differ — and what doesn't move them
Every split below is on 100+ lot sweeps in RTH. Direction columns carry standard errors; the "control" is the same split on 10–49-lot sweeps, so that a location or time effect that also shows up for small orders is not credited to size.
a. What the sweep did to the book L3 · robust mechanism
The single most stable pattern in the study. When 100+ lots barely move the mid (bottom quartile, ~3 ticks: the book was thick or a block absorbed them), the next minutes lean with the aggressor: continuation-first beats pullback-first (50% vs 44%) and the mean drift is +2.4 ticks at 5 minutes. When 100+ lots blow the mid 27 ticks through a thin book (top quartile), pullback-first dominates (56% vs 31%) and the mean drift is −2 ticks at 1–5 minutes. Same signs in every year 2021–2026 for the sequence, and 5 of 6 for the drift.
Absorbed vs blew-through, year by year
Continuation-first minus pullback-first (percentage points, k = 1 pt) for the low-impact and high-impact quartiles of 100+ lot sweeps.
absorbed (impact q1)blew through (impact q4)
Positive = the market more often kept going the aggressor's way before giving back a point. Absorbed sweeps stay positive, blew-through sweeps stay strongly negative, in every year. The mechanism is transient impact: the more a sweep walked, the more the book refills behind it.
b. Fresh session extreme L2–L3 · size-specific, 5 of 6 years
A 100+ sweep that prints through the running session high/low (20% of them) gives back −4.1 ticks in 5 minutes (SE 1.1) and shows pullback-first 54% vs continuation-first 35%; interior sweeps run +1.0 (SE 0.5). Small orders at fresh extremes show −0.5 — so this one is about the size. It washes out by 15 minutes (−2.4, SE 1.7). The mirror — sweeps near but not through the extreme continue +6 ticks at 15 min — is present at every size including 1–3 lots, i.e. it is the location ruler, not the order.
c. Book balance, on the 241 days with depth L1 · descriptive, one year, thin
For 50+ lot sweeps between 2025-07-31 and 2026-07-31 (34,030 joined to the resting book within 20 s), split by resting size in the 10 pts ahead of the sweep and behind it (z-scored within the day). Thin ahead is what makes impact big (r = −0.24) — so this is the depth-side view of pattern (a). Numbers are mean drift at 15 min in ticks; travel is the median ±excursion.
The prior desk test of this idea as a trade (whale_book_filter, 2026-08-14) did not confirm an EV edge, and one year is one regime. Read it as: a big order that clears a thin far side while its own side is stacked tends to hold and extend; a big order into a book that is thin on both sides marks a wide, fragile tape.
d. Everything that does not change the direction
Mean forward drift for 100+ lot sweeps in ticks (± SE), share of pullback-first vs continuation-first at 1 pt, median 15-minute favourable / adverse travel, and the same drift for 10–49-lot sweeps in the same state. Where a big-order cell and its small-order control agree, size is not the cause.
5 · What to do with it
Read a big print as "liquidity was consumed here," not "someone knows something."
Across 56k sweeps of 100+ lots the forward mean is 0 ± 1 pt at every horizon, in every year, on both sides. Chasing it or fading it are both coin flips minus the spread. The one exception with a lean is 500+ lots (+2–4 pts by 5–30 min), and that is 446 events with errors as large as the mean.
Read the impact, not the size.
A 100+ lot sweep that moved the mid a tick or two was absorbed — the book is thick there, the aggressor is not done, the tape leans on continuing. A 100+ lot sweep that moved it 5+ pts found air — expect the immediate 1–2 tick give-back and no follow-through premium; the "dip" you see is the book refilling, not a buyer being filled. On the DOM this is the same read as thin-ahead vs thick-ahead.
A big sweep that makes the new session high/low is not confirmation.
Those lean back ~1 pt over the next 5 minutes, more than small orders in the same spot. It is a small, short-lived lean, not a fade signal — but it means the print itself adds nothing to a breakout thesis.
Use the environment to size the box, not to pick the side.
After a big sweep in the first 30 minutes or in the fast-vol tercile the next 15 minutes travel ~±30 pts either way; midday or quiet, ~±14–17. Time-of-day, prior move, cluster and regime never moved the mean more than 1.5 pts, and small orders moved the same.
Calibrate "big" to the sweep, in lots — that number is stable; single-print size is not.
100+ ≈ 32 a session, 200+ ≈ 6, 500+ ≈ one every three days. A 100-lot moves NQ ~13 ticks today (2021: 8.5) but the same ~1.5 bps; percent-normalise before comparing eras.
Where a real read could still be earned: MBP book balance around a big sweep.
Thin-ahead + thick-behind held and extended +4–5 pts on the depth year (n≈1,500, t≈2.5). One year, no EV confirmation. Worth watching live on the DOM; not worth a rule yet.
Method & caveats
Data. Sierra Chart .scid ticks, 23 front-month NQ contracts by the validated roll table (+ NQU26_FUT_CME rename after 2026-07-28), 2020-11-27 → 2026-08-14. Trade price = c, ask/bid = h/l, aggressor volume = askv/bidv. Prices ×100 scale detected, not assumed.
Reference points. Pre-mid = quotes on the sweep's first record (the book before it). Post-mid = quotes on the first record after the sweep (≤2 s later; else last fill). Drift is mid-to-mid, so bid-ask bounce cannot manufacture a "dip." Forward path sampled on a 1-second grid (last / high / low mid within each traded second); horizons flagged stale if no trade within 300 s.
Controls and traps handled. Small-order control on every environment split; the location effect fails the control (present at all sizes) and is reported as such. Pre-sweep momentum measured and shown. Era comparability by stable sweep-size distribution and bps units. Sequence question answered as first-passage on the mid, not on fills. Not tested: entries, stops, EV — this is an awareness study by design.
Known limits. A 1-ms grouping can merge two different aggressors that arrive in the same match event (a stop cascade) — that is still one burst of aggression. Print size = min(aggressor remaining, resting order), so a "block" tells you a big resting order was matched, not who was bigger. Tiers ≤24 lots are subsampled (uniform 1% of <10 lots; 60k per contract of 10–24). Depth join covers one year only. MBO (per-order queue) is not recordable, so "was it one buyer" is unknowable here.
Evidence levels per the desk standard: impact ladder and width-by-environment = L1 descriptive, robust; impact-quartile aftermath = L3 (stable across years, mechanism, incremental over control); fresh-extreme give-back = L2–L3; book-balance = L1; 500+ continuation = L1 thin; environment → direction = null.