Market Structure Is Policy Written in Microseconds

This issue was meant to carry a measurement. The obvious test of the claim that microstructure rules become executable incentives is whether trading costs moved when the rules moved, and there is a published estimator that recovers bid-ask spreads from daily high and low prices alone. It was built, run over eight long-listed large caps back to 1996, and rejected.

It returns a median of about 20 basis points for securities whose true quoted spreads are on the order of one to three, roughly 30 per cent of its monthly variation is explained by realised volatility, and its estimate around decimalisation rises rather than falls. It is measuring the daily range. The script is preserved in the repository with that verdict printed in it, and this article carries a table of the actual regulatory parameters instead.

Section 01Rules become incentives

An electronic market is a rule system before it is a stream of prices. The matching engine defines priority; the tick constrains admissible quotes; transparency rules determine what is published and when; access fees determine what a fill actually costs. Private optimisation then happens inside that feasible set, and the visible behaviour of the market is the solution to the problem the rules posed.

The cleanest demonstration is the speed race. Budish, Cramton and Shim argued that the high-frequency arms race follows from continuous-time matching itself rather than from any participant's conduct. If orders are processed in the sequence they arrive and price information is correlated across venues, then being first to react to the same public signal is mechanically valuable, and the value gets bid into latency. Batch the auctions into discrete intervals and first-to-arrive stops being decisive. Same information, different allocation rule, different equilibrium.

Figure 1 sketches the journey of an order through the resulting fragmented market: a signal becomes a parent order, a router splits it into children and decides continuously where each should go, and the fills come back to be measured against a benchmark. Every arrow in that diagram has a rule behind it.

Section 02The parameters, and when they bind

Figure 2 sets out what the 2024 Regulation NMS amendments changed and, above the chronology, when each piece of that change actually takes effect. Every value and every date on the plate is machine-read from the release that set it by the script that draws the figure, which refuses to publish if one of them stops matching its source. That is a small guard against the most common failure in writing about regulation: a number that was right in a draft and drifted.

Three entries carry most of the economic weight. A half-cent quoting increment applies to NMS stocks priced at or above a dollar whose time weighted average quoted spread is 0.015 dollars or less. The access fee cap for protected quotations at or above a dollar falls from thirty mils per share to ten, and the sub-dollar cap from 0.3 to 0.1 per cent of the quotation price. Odd-lot information, including the best odd-lot order, enters consolidated market data. No minimum increment was adopted for trades: the increment governs quoting alone.

The dates are the part a reader is most likely to be carrying around wrong. As adopted, six of the seven requirements were to bind on the first business day of November 2025 and the seventh on the first business day of May 2026. What happened instead was litigation. Petitions for review were filed in the D.C. Circuit within six weeks of adoption, and in December 2024 the Commission stayed the minimum pricing increment, the access fee caps and the associated regulatory-data indicator pending judicial review, while declining to stay the requirement that fees be determinable at execution, the round lot definition, or the odd-lot definitions. The petition was denied in October 2025.

Three exemptive orders have since moved the schedule further. The increment and the fee caps went to the first business day of November 2026, then to November 2027. The odd-lot depth-of-book requirement, which asks the exclusive processors to publish odd-lot quotations aggregated at each price level, went to May 2028 at the plans' own request, on grounds of concurrent implementation load rather than any objection to the requirement. The best odd-lot order, a separate paragraph of the same definition, was not deferred. Two obligations adopted in the same passage of the same release are now almost two years apart.

That gap between a rule and a date is the durable lesson here. The rule text has not been amended. Each order suspends an obligation temporarily and says so; none concedes anything about the merits, and the Commission said in the stay order itself that it would continue to defend the amendments in court. But a routing model, a fee schedule or a tick-size assumption keys off the compliance date rather than off the rule, and that date has now moved four times.

The Commission's own estimate in the same release is that up to 74.3 per cent of the share volume transacted in NMS stocks in 2023 may have had spreads constrained by the one-cent increment. Whether one accepts that figure or not, it is the scale the amendment is addressing, and it is a great deal larger than the topic's usual billing as a technical adjustment suggests. It is also, on the current schedule, a constraint that stays in place for another year.

Section 03When the design fails visibly

Stress episodes make the design legible. On 6 May 2010 US equity markets fell and rebounded violently within minutes. The official SEC and CFTC reconstruction identified a large automated sell programme in E-Mini futures executed against thinning liquidity, and traced the propagation through cross-market arbitrage into individual equities. No participant had to behave badly for the outcome to occur; the interaction of a sizing algorithm, a fragmented venue structure and withdrawing liquidity was sufficient.

The lesson concerns design rather than blame. A market's behaviour under stress is a property of the rules it runs on, and those rules are a policy artefact. A different set of rules would have produced a different episode, and the counterfactual is unknowable, which is precisely why the parameters in Figure 2 are worth stating precisely rather than describing loosely.

continuous matching, price-time priority:

    public signal arrives at t
    N participants react; the first to arrive captures the stale quote
    => the value of the arbitrage is bid into latency, not into price

    the race is mechanical: it recurs on every signal, and it is
    a property of the allocation rule rather than of any participant

discrete batching:

    orders arriving within an interval are matched at one price
    "first to arrive" is no longer decisive within the interval
    => competition moves from speed back toward price

Section 04Policy catching up

If microstructure is policy, supervision is the slow loop that watches the fast one and adjusts. That loop is visibly active. In February 2026 ESMA published a supervisory briefing on algorithmic trading covering governance, testing, outsourcing and pre-trade controls, non-binding material intended to converge supervisory practice rather than to create obligations.

The behaviour you see at the touch is the policy you wrote, executed at machine speed.

The operational implication is model risk of an unusual kind. A strategy calibrated to queue length, venue fill probability, fee schedules or transparency waivers contains regulatory parameters whether or not the code labels them as such. The access fee cap has not yet fallen from thirty mils to ten; on the schedule as it stands it does so in November 2027. A model calibrated on thirty-mil economics is therefore not wrong today, and it carries a dated expiry that nothing in its backtest will produce.

What this issue does not establish is whether any of these changes improved execution. That is the measurement that failed, and the honest close is to leave it open. The parameters are documented; their effect is an empirical question that requires data this journal does not hold, and a weaker proxy would have produced a confident answer that happened to be about volatility.

This research is analysis and commentary for general information. It is not investment advice, legal advice, an offer, or a solicitation, and it contains no price forecasts. Regulatory parameters are quoted from the release cited and are verified against its text by the script that draws Figure 2.