Options Sweep Detection: The 500ms Coalescing Rule and Cross-Venue Routing

If you have ever watched a fast-moving SPY contract get lifted across three exchanges in the same second, you have seen a sweep. To the underlying institutional order, it was one decision: "fill 2,000 contracts of this strike now, at whatever venues will quote them." To the OPRA tape, it looks like four prints landing in 80 milliseconds, each on a different exchange, each showing a fraction of the intended size. To a naive flow tracker, those four prints look like four independent trades.

That mismatch between intent and tape is the central problem in options sweep detection. Get it wrong and you either inflate the unusual-activity rank (counting one $5M order as four $1.25M orders) or lose it entirely (the largest single print being too small to qualify as a block on its own). The FlashAlpha Flow Signals API handles it with a deterministic coalescing rule. This is the article that explains the rule.

What a Sweep Really Is

The textbook definition: a sweep is a multi-leg order that walks the inside markets of multiple exchanges in immediate succession to fill faster than a single-venue order could. In equity options that means the routing layer takes a single parent order, breaks it into child orders sized to the displayed liquidity at each exchange, and fires them off in parallel.

The signal you care about is the parent intent, not the per-venue child print. The child prints are an artefact of how exchange routing works, not the trader's decision. A sweep is the trader's decision; a print is the exchange's record.

Why a Single-Print Scorer Gets It Wrong

Imagine a 2,000-lot SPY call sweep that prints as four 500-lot fills across PHLX, CBOE, AMEX, and MIAX in 60 milliseconds. A scorer that treats each print independently has two failure modes.

Neither outcome is acceptable for an unusual-activity feed. The coalescing rule is the deterministic preprocessing step that fixes both.

The 500ms Coalescing Rule

The rule is a single line of code in the scorer and a single sentence of policy:

Three conditions must hold simultaneously for two adjacent prints to join the same group:

1. Same contract. Different strike, different expiry, different right are different intents by definition.
2. Same side. A buy followed by a sell, even at zero latency, is two opposing decisions. Side is preserved.
3. Within 500ms of the prior print. The window is interval-to-interval, not relative to the group's first print. Two prints 300ms apart join; a third print 400ms after the second also joins; a fourth print 600ms after the third does not.

A coalesced group with two or more prints is classified as a sweep. A singleton block-sized print is a block. The combined group is treated as one execution event: the size is the sum, the price is the size-weighted average, the timestamp is the last print in the group.

Why 500 Milliseconds

The window is wide enough to catch most observed sweeps and narrow enough that unrelated prints rarely join in our calibration set. We arrived at it empirically, not theoretically.

500ms sits comfortably in the gap: well above any plausible sweep latency, well below any normal cadence of independent trades on the same contract. Setting the window to 100ms would risk splitting genuine sweeps on slower routing paths; setting it to 2 seconds would risk joining unrelated prints on the busiest names. 500ms is the deliberate, defensible middle.

Block vs Sweep vs Single: The Three-Way Distinction

Every scored signal in the feed gets one of three structure labels. Each carries different information about how the trade was executed and how strong an urgency signal it represents.

The score weights are not arbitrary. Sweeps get the maximum because urgency is the strongest signal in the structure dimension; blocks get a substantial discount because size without speed is consistent with patient execution; singles exist only for upstream completeness and rarely appear because the notable-trade ring filters to block size at ingestion.

How the API Returns It

Every entry in the signals array of the Flow Signals response carries a structure field and the coalesced group's combined fields. Inspecting a single signal tells you whether it was one print or many.

{
  "ts": "2026-05-22T18:14:32.421Z",
  "expiry": "2026-05-22",
  "strike": 587.0,
  "right": "C",
  "side": "buy",
  "price": 1.43,
  "size": 2000,
  "premium": 286000.0,
  "dte": 0,
  "structure": "sweep",
  "aggressor": "at_ask",
  "open_close_bias": "opening_bias",
  "open_close_confidence": 0.43,
  "intent": "bullish",
  "score": 81,
  "conviction": "high",
  "tags": ["sweep", "opening", "0dte", "golden"],
  "score_breakdown": {
    "premium": 15,
    "size_vs_oi": 14,
    "aggressor": 14,
    "sweep": 18,
    "opening_bias": 9,
    "tenor": 11
  },
  "enrichment": { /* greeks + IV vs ATM + delta notional */ }
}

That single record represents four child prints that landed within 240ms total. The size is the parent sum, the price is the size-weighted average, the timestamp is the last child. The tag list includes sweep, which is true for any structure-coalesced group with at least two prints, and the score_breakdown.sweep bucket carries the maximum-weight contribution from this dimension.

How to Filter for Sweeps Only

Pass structure=sweep to the signals endpoint to drop blocks and singles. Combine with a score floor and an intent filter for a focused alert ruleset.

import requests

resp = requests.get(
    "https://lab.flashalpha.com/v1/flow/signals/SPY",
    headers={"X-Api-Key": "YOUR_KEY"},
    params={
        "structure": "sweep",
        "intent": "bullish",
        "minScore": 70,
        "windowMinutes": 60,
    },
)

for sig in resp.json()["signals"]:
    n_prints_estimated = "multiple" if sig["structure"] == "sweep" else "1"
    print(
        f"{sig['ts']}  {sig['expiry']} {sig['strike']}{sig['right']}"
        f"  size={sig['size']:>5} (sweep group)"
        f"  ${sig['premium']:>10,.0f}"
        f"  score={sig['score']}"
        f"  {','.join(sig['tags'])}"
    )

The opposite filter, structure=block, gives you the patient single-venue side of unusual activity. Both filters together (which is the default - leave structure unset) give you the full feed.

Edge Cases and Honest Limitations

The coalescing rule is a heuristic. It is right almost all of the time and worth knowing the boundary cases for the times it is not.

Genuine Independent Prints Within 500ms

On the very busiest contracts (SPY 0DTE near close, mega-cap earnings nights) two unrelated traders occasionally happen to print same-side block trades within the window. The coalescer joins them. The result is one signal with combined size from two different parents. Detection cost: the size is real (it really did print that aggregated quantity in 500ms) but the "single intent" interpretation is wrong. There is no clean way to separate them on the OPRA tape.

Slow Multi-Venue Routes

A few exchanges have acknowledgement paths slow enough that the final child print of a sweep can land beyond 500ms after the first. The coalescer treats the late child as a separate event. The resulting two signals (the early coalesced group and the late singleton) report a smaller-than-actual parent size. The bias is conservative - splitting a sweep is preferred to joining unrelated prints.

Cross-Side Sweeps Are Two Signals

A trader simultaneously buying calls and selling puts (a synthetic long) shows up as two independent same-side groups by design. Each leg is scored separately. The combined intent is unclear from the tape alone, so the scorer does not assume a structure. Multi-leg structure inference is a candidate for future work but is not in scope at the coalescer layer.

Comparing With Other Vendor Approaches

Most unusual-activity vendors do not publish their sweep-detection rule. The few that do hint at it use one of three patterns:

Our approach is closer to how the order books actually work than the exchange-tag method, and far stricter than time-only joins. The whole point is that "what was the trader trying to do" is recoverable from the tape with a careful enough rule, and a 500ms same-contract same-side interval is the rule we settled on.

Frequently Asked Questions

Related Reading

• Flow Signals API: Scoring Unusual Options Activity - the full methodology pillar that this sweep-detection rule feeds into
• Opening vs Closing Bias on Options Flow - how the OI simulator's bias tag determines whether a sweep is a directional signal or just an unwind
• Build an Unusual Options Activity Scanner (Python) - end-to-end example combining sweep detection with the full Flow Signals API
• Effective Open Interest Methodology - the same OPRA tape and side classifier that feed the coalescer
• Options Flow API: Recent Trades, Blocks, Leaderboard - the raw-trade surface where individual child prints live