Information Leakage and Signalling Risk

At 09:45 you are handed 400,000 shares of a $50 stock to buy — about 10 percent of a normal day's volume. Nobody outside your firm knows. Two minutes later a dozen strangers have a better-than-even guess that a large buyer is working, and the price you will pay has already gone up. You never told anyone. Your algo did, one child order at a time.

Think of bidding at an auction while raising your paddle in a steady rhythm. No single raise gives you away. The rhythm does.

The book, and what four clips reveal

The top of the book at 09:45:00.

Side	Price	Size
Ask	50.03	4,100
Ask	50.02	2,600
Ask	50.01	1,900
Bid	50.00	2,400
Bid	49.99	3,300
Bid	49.98	5,000

Your algo is set to 10 percent participation. It slices the parent into 5,000-share children, posts them passively at the bid, and works only the primary exchange.

• 09:45:00 — post 5,000 at 50.00. The displayed bid goes 2,400 → 7,400.
• 09:45:38 — sellers take 4,900 at 50.00. Within 300 ms the bid is back at 7,400: your algo reposted the instant it filled.
• 09:46:21 — same shape. Level chewed down, refilled to 7,400 inside 300 ms.
• 09:47:04 — and again.

No single event is unusual. The pattern is: a 5,000-lot arriving at one venue within 300 ms of the level clearing, three times, on a 45-second beat. A detector that starts from "one refill in a hundred is a large parent" is, after three matching refills, extremely confident — and that confidence is worth money.

• 09:47:20 — the detector lifts the whole visible offer: 1,900 at 50.01, 2,600 at 50.02, 4,100 at 50.03. That is 8,600 shares at an average of 50.0225, about $430,200. He then posts 12,000 for sale at 50.06.

You still have 385,000 to buy and the cheap offers are gone. Over the next hour your average fill comes in at 50.055 instead of the 50.020 the same schedule would have produced unnoticed.

Extra cost: 385,000 × $0.035 = $13,475. On $20.0m of notional that is 6.7 basis points, paid to somebody with no view on the stock at all — only a view on you.

Leakage is the impact that does not come back

Ordinary market impact is mostly temporary: you consume liquidity, the book refills, the price drifts back. Leakage is not. Somebody who has learned something about future demand holds the position, and the price stays where they pushed it. That difference is measurable.

Worked example: separating the two

• Arrival mid 50.000, average fill 50.055. Shortfall = 5.5 cents = 11.0 bps.
• Mid 30 minutes after your last fill: 50.048. So 4.8 cents stuck and only 0.7 reverted. Almost none of your cost was the book refilling — it was people repricing.
• The control: on comparable days when the same size ran through a randomised, multi-venue schedule, the mid 30 minutes after completion averaged 50.021 — 2.1 cents of permanent move. That residual is real information plus drift, and it is unavoidable.
• Leakage = 4.8 − 2.1 = 2.7 cents per share, or 5.4 bps — about $10,800 on this $20m order.

Tracked per algo and per venue, that number is the only honest scorecard for anti-gaming work.

What actually leaks

• Size regularity. Identical clips are the loudest tell — so are clips of exactly 10 percent of displayed depth.
• Timing regularity. A fixed interval, or reposting within a fixed few hundred milliseconds of a fill, is a fingerprint.
• Venue concentration. Sitting only on the primary makes you trivially observable; so does pinging thirty venues in a fixed sequence.
• Price footprint. Always joining the touch and never stepping back says "I have more to do".
• Shopping the order. Every RFQ, indication of interest and dark ping tells one more counterparty your side and rough size, traded or not.

The fixes mirror the tells: randomise size and interval, split across venues without a fixed rotation, mix passive and aggressive placement, size children against current displayed depth rather than a constant, and delay reposts randomly. Randomising Execution To Avoid Detection covers the mechanics; Anti-Gaming Logic In Execution Algos covers what to do once you suspect you have been found.

In interviews

You will be handed a schedule and asked what is wrong with it. Point at the constants first — same size, same interval, same venue — then say how you would measure the damage rather than assert it: shortfall against arrival, minus the portion that reverts within thirty minutes, benchmarked against a randomised control. The usual follow-up is "so why not just trade in the dark?" Leakage does not disappear when you hide; it changes counterparty. See Market Impact for the reverting half.