Bubbles & sweeps

It is 10:07 ET (22:07 ICT, winter). NQ trades at 20,400.25 and your chart paints a large buy bubble: 46 contracts in one stroke, lifting five prices from 20,400.00 to 20,401.00 in the same millisecond. Many traders read that as “size is buying, follow it”. This lesson is about what that bubble actually records, why its size depends on a setting you chose, and what the data says happens after prints like it.

Key terms

Large-trade marker (bong bóng khối lượng, “bubble”): a circle drawn on the chart wherever an aggressive trade is larger than a size threshold you set Definition. One bubble means one participant, or one burst of orders, crossed the spread with size at that moment. It marks participation, not prediction.

Sweep (quét thanh khoản): one aggressive order that trades through three or more price levels of the book in a single match event Definition. The resting size at the first price was not enough, so the order kept filling at worse prices until it was done.

Threshold (ngưỡng): the size above which a print counts as “large” Definition. It is a display setting about your market, not a constant. Fixed tiers you will see quoted, such as “large on NQ is 20 to 49 contracts, very large 50 and up”, are rules of thumb nobody derived from a test Lore.

Chart first: one order, one bubble, five prints

How it works

Size tells you someone acted, not that they were right

A 500-lot buyer on ES might be informed. They might also be hedging an options book, rolling a position, rebalancing a portfolio, or simply wrong. Large participants lose money too; size does not make anyone clairvoyant. The bubble tells you that someone with size crossed the spread now. It does not say why.

The obvious test is to check whether large single prints are followed by a move their way. A public practitioner study ran that test on NQ and found the opposite of the claim, by a small margin: after a large single print, price reached a bracket in the print’s direction slightly less often than at comparable moments without one Lore (box below).

Thresholds: use percentiles, not folklore

A fixed number goes stale: what counts as large in a quiet week is routine in a fast one, overnight prints run smaller than RTH prints, and the aggregation setting in the figure above can multiply print sizes several times over (M3.13 covers the data settings). A percentile threshold follows the market instead: “mark prints above the 99.5th percentile of the last 20 RTH sessions’ print sizes” stays meaningful when volume changes. Too low and every print is a bubble; too high and you miss the participants you wanted to see. The drill below walks through setting one.

Sweeps: what the data says

The popular reading is that a sweep shows urgency, so price should keep going the sweep’s way Lore. Two independent looks at NQ say the opposite, by a small amount: over the next minute, sweeps continued slightly less often than comparable single-price bursts of the same side and size PlausibleNQ. The wider the sweep, the larger the reversal: about 1.4 points of continuation rate for three or more levels, 2.3 for four or more, 3.1 for six or more. After fifteen minutes the difference is gone.

That pattern looks like temporary price impact: a large order pushes through thin book, overshoots, and part of the push comes back as the book refills. It is a description of what usually happens, not a trading signal, for three reasons that travel with this result everywhere on the site:

• Forward test pending. Our lab’s first screen passed its pre-registered bar; the out-of-sample forward test on new sessions has not finished.
• Trade prices, not quotes. Outcomes were measured on trade prices. The bid–ask bounce was checked on three days and came out near zero, but a quote-based replication has not been run.
• Smaller than a round trip. The average extra move against the sweep after one minute is about 2 ticks, roughly what a round trip costs in spread and fees. A fade trade built on it has not been tested with costs.

Markout: why a fill inside a sweep starts at a loss

Markout (markout) is the mid price a short time after your fill (1 second, 10 seconds, 60 seconds), signed by your side Definition. For a resting limit order it measures adverse selection: did you get filled because someone who knew more traded against you?

A resting buy limit at 20,400.50 that fills during a 46-lot sell sweep is filled because the sweep went through it. By the time your fill reports, the book has traded down to the bottom of the sweep, so the mid already sits below your price: the markout is negative from the first second. That is the classic adverse-selection mechanism from market-microstructure theory (Glosten and Milgrom): passive orders get filled most when the other side has the bigger reason to trade PlausibleOther market. On NQ the size of this effect is not yet measured by our lab.

The slight reversal in the box does not cancel this. A minute later the sweep gives back a little more often than a single-price burst would, but continuation still runs near 48%, and your fill started several ticks under water. Do not treat “sweeps reverse slightly” as “it is safe to rest orders in the path of one”.

Clusters and “campaigns”

Myth / hypothesis: “Clusters beat singles: three 150-lots at one level over an hour are a campaign, and campaigns move markets more reliably than one big print.” Lore The story is appealing, and patient size-building does exist. As a forecast it is untested: neither our lab nor the public study has measured clustered large prints against a baseline, so it is not yet measured. Single large prints, the one version that was tested, showed no edge. Log clusters if you like; do not size up because of them.

Why bubbles are still worth knowing

Bubbles and sweeps show what the candle hides: where size crossed the spread, how thin the book was, how fast a price was taken. They warn you that volatility just arrived and explain why your limit order filled at the worst moment. And they are the kind of claim you can test yourself with a written rule and a baseline (EV1.04).

Worked example: the sell sweep at 20,400, long side

Synthetic session, built to the shape above, not a real date. NQ, 1-minute chart, tick 0.25 points, $5 per tick per contract. Your bubble threshold is 40 contracts, your own 99.5th percentile.

At 10:12 ET NQ trades down into 20,400.00, yesterday’s value-area low. Three sell bubbles print within four minutes: a 52-lot sweep from 20,401.25 to 20,400.00 (six levels), a 44-lot at 20,400.00, and a 61-lot sweep from 20,400.75 to 20,399.75. Price touches 20,399.75 once and does not trade lower; the next bar closes at 20,400.75.

What you know for certain: aggressive sellers spent 157 contracts here and moved price one tick through the level. That is the description. What you do not know is the next move. Wide sweeps have slightly reversed on average, but by about a tick or two, and continuation still happens close to half the time. The bubbles do not choose the ending, so you write the plan instead of reacting to the size:

• If a 1-minute bar closes back above 20,401.25 (the top of the first sweep) with positive delta, the sellers’ push has been given back. Long 1 contract at 20,401.50 on a stop order.
• Stop 20,399.00, three ticks under the low: 10 ticks of risk, $50 per contract before costs.
• Target 2R, 20 ticks, at 20,406.50.
• Invalid if a bar closes below 20,399.75 with negative delta. You stay flat.

Notice that the plan enters with a stop order after the sweep, not with a resting limit inside it. A buy limit at 20,400.25 would most likely have filled during the 61-lot sweep with an immediately negative markout.

The mirror. A buy sweep into a level above value prints the same way upside down: plan a short only if price closes back below the bottom of the sweep with negative delta, stop a few ticks above the sweep high.

The failure. Same morning shape, another synthetic day: the third sell sweep is followed by a 75-lot sweep down to 20,398.50, and the next bar closes at 20,398.75 on delta −1,400. The trigger never comes and the plan stays flat. A trader who bought the “absorbed” bubbles at 20,400.25 is down 7 ticks at the low, $35 per contract plus costs, with the market still moving away. Sweeps reverse slightly on average; this one did not, and the average does not protect any single trade.

Prop-firm note. In an evaluation or practice account your limit orders fill by the simulator’s rules, not by a real queue, so a sim fill inside a sweep can look kinder than a live one. Firms forbid exploiting simulated fills (PROP1.04, in the Prop Firm course). MKT2.03 covers how sim fills differ from live fills; until it is published, read your firm’s help-center article on prohibited conduct before you trade a firm account.

Do it on the platform

Common mistakes

Common mistake: Treating size as a signal to follow. Large traders are often hedging, rolling or rebalancing, with no directional view at all, and their directional bets lose too. Following bubbles means buying other people’s hedges. On NQ, large single prints showed no edge for direction.

Common mistake: Treating a base rate as an edge. “Sweeps reverse” sounds like a fade signal. The measured difference is about 1.4 points of continuation rate and about 2 ticks of price, roughly a round trip in costs, and the forward test is still pending. It tells you not to chase sweeps; it does not yet tell you to fade them.

Common mistake: Resting a limit order in the path of a sweep to “catch the reversal”. If it fills, it fills because the sweep went through it, and the markout is negative from the first second.

Drill: set your threshold from percentiles

Export or record the aggressive trade sizes for the last 10 RTH sessions of NQ (PLAT2.09 shows an ATAS X export), with fills of one match event grouped into one trade.

1. Sort the sizes and find the 99th, 99.5th and 99.9th percentiles.
2. Count how many trades per session each percentile would mark.
3. Pick the percentile that gives you a number of bubbles per session you can actually read and log (for example 20 to 60), and write that size into your platform.
4. Repeat steps 1 to 3 with the fills left ungrouped and compare the thresholds.

What we don’t know yet

• Whether the slight sweep reversal survives the pre-registered forward test on new sessions: pending in our lab.
• Whether a fade trade on wide sweeps makes money after costs: not yet measured.
• The size of markouts for passive fills inside NQ sweeps, and a quote-based (not trade-price) version of the sweep result: not yet measured; both need a top-of-book recorder.
• Whether clusters of large prints at one level carry any information: not yet measured.

Recap

• A bubble marks size that crossed the spread; its size depends on your threshold and on whether the feed groups the fills of one order.
• Set the threshold from your own print-size percentiles and recompute it when volatility changes.
• Large single prints showed no edge for direction on NQ (Lore). Sweeps through three or more levels reverse slightly over the next minute (Plausible, NQ, forward test pending), by about a round trip in costs.
• A resting limit order filled inside a sweep tends to start at a loss (negative markout); plan entries after the sweep instead.

Further reading

• Order Flow, Measured: All claims. A public, pre-registered practitioner study on NQ, 2018–2026. Read the sweep row (OB-03) and the large-print row (FP-09).
• Glosten and Milgrom (1985), “Bid, ask and transaction prices in a specialist market with heterogeneously informed traders”, Journal of Financial Economics 14(1). The model behind adverse selection of passive orders.
• Our lab: NQ sweeps of three or more levels continued about 1.4 points less often than matched single-price bursts over the next minute, with the effect growing with sweep width and gone after 15 minutes; forward test pending.

Next: M3.12 looks at the other side of the book: orders that are shown and pulled before they trade, and what the law and the data say about them.

Educational content only. Not financial advice. Futures trading involves substantial risk of loss.