Milliseconds Are Not a Strategy: When Being Faster Matters, and When It Does Not

Speed is the most visible feature of modern electronic markets, and the most misread. The folk theory holds that the fastest participant wins, that every microsecond removed from a trading system converts into return, and that a firm which is not racing for the last nanosecond is not a serious participant. None of that follows from how markets actually reward speed. Latency is a constraint on which opportunities a system can act on. It is not, on its own, a source of return. A faster system captures more of an opportunity only when the opportunity is disappearing on a timescale comparable to the system's reaction time, and most opportunities are not.

This piece sets out a deliberately simple model of that relationship and draws four conclusions from it. First, the value of speed for any given opportunity is a cliff, not a slope: above a certain reaction time the trade is gone, and below it further speed buys nothing. Second, the marginal value of an increment of speed is a hump concentrated within about a decade of the opportunity's decay horizon. Third, only a small share of a diversified systematic inventory sits inside that band. Fourth, latency is an engineering budget with a physical floor, spent in identifiable places. Together these say that milliseconds are not a strategy. They are a precondition for a narrow class of strategies, an irrelevance for most, and an expense in every case.

A Clock on Every Opportunity

Every trading opportunity has a decay horizon: the time over which its value disappears as prices adjust, as other participants act on the same information, or as the event that created it is absorbed. The horizons differ enormously. A price discrepancy between two venues quoting the same instrument closes within microseconds, because the participants who arbitrage it are co-located and competing with one another. A quote that is stale after a scheduled announcement survives for milliseconds to seconds. An intraday mean-reversion pattern unwinds over minutes. A cross-sectional valuation signal takes weeks or months to be reflected in prices, and nobody is racing to be first.

The simplest model that captures this is exponential decay. Let the value of an opportunity at reaction time τ, measured from the triggering event to the arrival of an order at the venue, be V(τ) = V₀ · exp(−ln 2 · τ / h), where h is the half-life. Expected capture is the fraction V(τ) / V₀ that remains when the order arrives. Figure 1 plots that fraction against reaction time on a logarithmic axis running from one microsecond to ten seconds, for four stylized opportunity types whose half-lives span six orders of magnitude: a cross-venue arbitrage at 50 microseconds, short-horizon liquidity provision at 5 milliseconds, a reaction to a public event at half a second, and an intraday statistical signal at one minute.

Each curve has the same shape: a plateau, a cliff, and a floor. Left of the cliff, the system is fast enough and further speed changes nothing. Right of the cliff, the system is not in the trade at all. The cliff itself spans roughly two decades of reaction time centered on the half-life, and that is the whole region in which speed is a live variable. A participant whose systems react in about a millisecond is comfortably on the plateau for the event-reaction and intraday curves, sits on the cliff for liquidity provision, and is at the floor for the arbitrage. Nothing about that participant's engineering changes the ordering. The requirement is set by the opportunity, not by the ambition.

The Speed Premium Is a Hump

The practical question for anyone deciding whether to invest in speed is not how much of an opportunity is captured but what one more increment of speed is worth. The model answers that directly. The gain from halving reaction time is V(τ / 2) − V(τ), which is exp(−ln 2 · τ / 2h) − exp(−ln 2 · τ / h). That expression has a maximum of exactly one quarter of the opportunity, reached when the reaction time is twice the half-life, and it falls toward zero on both sides. Figure 2 draws it for the same four opportunity types.

The picture explains a good deal of what is otherwise puzzling about the economics of speed. An engineering program that halves reaction time from two milliseconds to one is worth a great deal to a strategy whose opportunities have millisecond half-lives, next to nothing to one with microsecond half-lives, because those trades were already gone, and next to nothing to one with minute half-lives, because those trades were never at risk. The same investment is decisive, irrelevant, and irrelevant, depending only on where the inventory sits. Whether speed pays is therefore not a property of the firm but a property of the match between the firm's reaction time and its opportunities' horizons.

There is a further complication for the fastest class. For opportunities whose decay is driven by information arriving in the market, the half-life is a property of the world: an announcement is absorbed at whatever rate it is absorbed. For opportunities whose decay is driven by competition, the half-life is set by the fastest competitor, and it shrinks every time someone gets faster. Speed in those trades is a relative quantity. The hump in Figure 2 moves left as the field improves, and a participant who stands still slides down the cliff without having changed anything. This is the mechanism behind the well-documented latency arms race, in which the return to being first is competed toward the cost of the infrastructure required to be first.^[1]

Where a Systematic Inventory Actually Lives

Against that backdrop, it matters where the strategies in a diversified systematic inventory sit. Figure 3 shows a stylized inventory sorted by decay horizon. The shares are hypothetical, but the shape is the point: the great majority of strategies, by count, have horizons of minutes to days, where a reaction time of a second, or even a minute, leaves expected capture essentially intact. The sub-millisecond bucket, the only one in which the full sweep of Figure 1 is in play, is a small minority.

That distribution is not an accident of preference. The set of opportunities that decay in microseconds is narrow because it is bounded by physical facts: the number of venues quoting related instruments, the distance between them, and the speed of light in glass. It is also crowded, for the reasons in the previous section. Horizons of minutes and longer, by contrast, are limited by how well one can forecast, how cheaply one can trade, and how carefully one can construct a portfolio, none of which is exhausted by the arrival of one more fast competitor. The ratio of how good the forecast is to how quickly it is acted on shifts by orders of magnitude as one moves across the inventory.

Figure 4 makes the comparison explicit. For each of six stylized strategy families it lists a representative half-life, the reaction time at which the model retains about ninety percent of the opportunity, the capture gained by a tenfold speed-up from a baseline of 100 microseconds, and a summary of what constrains the strategy more than speed does.

Two columns deserve attention. The ninety-percent column is what "fast enough" means for each family: eight microseconds, under a millisecond, a tenth of a second, nine seconds, most of a day. Being faster than fast enough is spending, not investing. The gain column shows the same thing from the other side. Below the first row, a tenfold improvement in reaction time from a baseline that any competent co-located system already meets is worth about a percentage point of the opportunity or less, and for most of the inventory it rounds to zero. The single-parameter model does understate the case for speed in liquidity provision, where reaction time also governs how often a resting quote is picked off by better-informed flow, a cost that does not appear in the capture curve.^[2] Even there the value is bounded, and the binding constraint is the quality of the inventory and quoting logic rather than the last microsecond.

Where the Microseconds Go

If latency is an engineering budget rather than a strategy, it helps to know what it is spent on. Figure 5 traces a single round trip for a co-located participant: an event on the venue's book, the venue's publication of that event, transit to the participant, decoding, decision, transit back, the venue's own validation, and the matching engine's response. The stage times are illustrative orders of magnitude, not measurements, but their relative sizes are set by physics and by the venue's architecture rather than by anything a participant chooses.

Three things stand out. The first is that the stages a participant controls, decoding and decision, are a small fraction of the total for a co-located system. The venue's own processing at both ends of the trip, which no participant can shorten, sets a floor that is typically an order of magnitude larger than a well-built decision path. The second is that distance dominates everything once a system is not co-located. Light in fiber covers roughly 200 kilometers per millisecond, so a thousand kilometers of separation costs about five milliseconds each way, three orders of magnitude more than the decision step. The first question about any latency-sensitive strategy is therefore geographic, not computational. The third is that the budget tells an engineering team where not to work. Shaving a microsecond from decision logic that is bookended by tens of microseconds of venue processing changes nothing for opportunities with millisecond horizons. The exception is competition for queue position among co-located participants, where relative order is decided by the participant-controlled stages alone and nanoseconds can matter. That is, once again, the case in which the horizon is set by other people's speed.

How We Think About Speed

Oak St. treats latency as a property of each strategy rather than of the firm. Every strategy in the inventory is classified by its decay horizon and by the reaction time at which its capture curve flattens, and the infrastructure it runs on is chosen to meet that requirement with a margin, not to exceed it. Speed is bought to the point where the inventory stops rewarding it, and the budget is spent where the hump in Figure 2 actually sits for the strategies we run, rather than where it sits for the fastest participants in the market.

This makes us deliberately skeptical of any strategy whose entire edge is being first. When the horizon is set by competitors' speed, the edge is competed to the cost of maintaining it, and a firm that wins the race has usually paid for the prize in advance.^[3] We would rather own opportunities whose horizons are set by the world, act on them fast enough, and put the remaining effort into forecasting, portfolio construction, and cost control, where the returns to being better do not evaporate the moment someone else buys a faster switch.

Milliseconds are not a strategy. They are a precondition for a narrow class of strategies, an irrelevance for most, and a cost in every case. The job is to know which is which, strategy by strategy, and to spend accordingly.