Latency & the HFT layer

This is the lesson most likely to be distorted by things you’ve already read, so it’s worth being precise. Speed is not a general-purpose advantage. It buys two specific things, and outside those two it buys nothing at all.

What speed actually buys

1. Queue position. M0.2’s price-time priority means that at a given price level, orders fill in arrival order. If 40,000 shares rest at 50.00 and you are behind 30,000 of them, you fill only if the level trades more than 30,000 before it is cancelled or the price moves away.

   the bid at 50.00, in arrival order

   ┌────────┬────────┬─────────┬─────────┬──────────────┐
   │ 12,000 │ 8,000  │ 10,000  │  YOU    │  later ...   │
   └────────┴────────┴─────────┴─────────┴──────────────┘
     ◀── fills first                       fills last ──▶
     30,000 shares ahead of you

   a 25,000-share market sell arrives  →  you get NOTHING
   a 45,000-share market sell arrives  →  you get 15,000

Being a microsecond earlier moves you up the queue, and queue position is close to binary in its consequences. This is why the top of the book is not a “price” so much as a race result.

2. Picking off stale quotes. A maker’s quote reflects its estimate of value. When something that determines that value moves — the futures contract, an ETF’s underlying basket, the same stock on another venue — the quote is briefly wrong, and stays wrong until the maker can cancel it.

   t = 0 μs    ES futures tick up. SPY is now worth more.
   t = ?       fast TAKER hits every stale SPY offer still resting
   t = ?       slower MAKER cancels — too late

   whoever’s message arrives first, wins. Same information,
   available to both. The entire contest is transmission time.

Look carefully at what that is: adverse selection where the private information is nothing but speed. The maker loses exactly as it loses to a fundamental informed trader in M1.2 — it sold just before the price rose. Glosten–Milgrom’s μ, at microsecond resolution, is partly just how many people are faster than me.

This is why the arms race is symmetric and why it doesn’t end. Makers invest in speed defensively, to cancel before being picked off; takers invest offensively, to arrive before the cancel. Both are buying the same microseconds from the same suppliers.

The physical layer

The constraints are embarrassingly concrete for a financial market.

   Chicago ⇄ New Jersey, ~1,200 km each way

   light in vacuum          299,792 km/s   →  8.0 ms round trip
   light in optical fibre   ~200,000 km/s  →  12.0 ms  (glass is slow,
                                                and routes aren’t straight)
   Spread Networks fibre (2010)            →  ~13 ms — a $300m tunnel
                                                through the Alleghenies
   microwave through air (2012→)           →  ~8.5 ms — air beats glass,
                                                weather permitting

Microwave beat a $300m fibre build within about two years, largely because the refractive index of air is closer to 1 than glass. Latency is now measured in nanoseconds inside a matching engine’s colocation hall, where the binding constraint can be the length of your cross-connect cable — which is why exchanges sell every customer the same cable length regardless of rack position.

Where speed does not matter

This is the part usually omitted, and it matters more to most readers than everything above.

If your holding period is longer than seconds, latency is not your cost. A fund buying £50m of a stock to hold for six months is not losing to microseconds. It is paying spread (M0.4) and impact (M1.5) — costs measured in tens of basis points, orders of magnitude larger than anything latency does to it. Optimising the wrong one is a common and expensive mistake.

Retail is not the victim of latency. Retail orders are mostly internalised by wholesalers (M0.3) at or inside the NBBO, and never touch the race. The genuine questions about retail execution are about PFOF, segmentation and price improvement — not speed.

The measured spread has fallen. Whatever else HFT did, spreads and explicit costs are dramatically lower than in the quarter-fraction era. The serious critique isn’t “spreads got worse”; it’s about who captures the remaining spread, and whether the resources burned racing for it are socially wasted.

The mechanism-design critique

The sharpest academic argument (Budish, Cramton & Shim, 2015) isn’t that fast traders are villains. It’s that the continuous limit order book is a badly designed mechanism: by processing messages in continuous time and serially, it turns every public information event into a race, and races are won by speed rather than by price. The rents are then competed into infrastructure spending rather than into tighter spreads.

Their proposed fix is elegant: frequent batch auctions. Discretise time into (say) 100ms intervals and clear all orders in each interval at a single uniform price. Arriving 5 μs earlier within a batch is worth exactly nothing, so the race collapses and competition returns to price. Related, live experiments: IEX’s 350 μs speed bump (a physical coil of fibre) and various asymmetric delays that slow takers but not maker cancellations.

None of this is settled policy, and batching has real costs of its own. But it reframes the debate usefully — from “are HFT firms good or bad people” to “what behaviour is this mechanism paying for?”

And Flash Boys

Read it, but read it after the above, which is the roadmap’s advice and worth following. It is a genuinely good piece of narrative journalism, and it is also a book with protagonists — the IEX founders — whose venue it is effectively arguing for. Held up against the material in this module, you can see clearly which parts are accurate (the races are real, the stale-quote picking is real, the infrastructure spending is real) and which are overdrawn (that this constitutes systematic theft from ordinary investors, whose actual costs are spread and impact).

Being able to make that separation is roughly what M1 was for.

Source: Budish, Cramton & Shim (2015), “The High-Frequency Trading Arms Race” — the central paper, and readable. Irene Aldridge, High-Frequency Trading (2nd ed.) for practitioner mechanics. Michael Lewis, Flash Boys, last and with the above in mind.