The Flow Nobody Measures

What the Retail Record Supports

Last month, across the six largest wholesalers, 65.5 billion shares of marketable retail orders were executed in 12,643 symbols. We can tell you almost anything you want to know about them.

Clients received $673 million in price improvement against the prevailing quote. The brokers who routed that flow were paid $115 million. We can decompose that by symbol, by order size, by order type and by venue, fit an elasticity to it, measure how unevenly it is distributed, and cluster the symbols into archetypes to see who actually benefits.

None of it required a vendor, a survey, or a single conversation with a market maker. It is all sitting in files those firms are required to publish every month.

Now ask the same question about your own flow.

Improvement Falls Sharply With Order Size

Exhibit 1

Click to expand

A share-weighted log-log fit across 48 size-by-type-by-center cells gives an elasticity of -0.78, with a standard error of 0.067 and an R-squared of 0.75. A tenfold increase in order size is associated with roughly an 84% fall in price improvement per share.

The third column is why a cents-per-share gradient cannot be read on its own. Order size and share price move in opposite directions in this flow: the share-weighted average price falls from $115 in the smallest bucket to $14 in the largest, an eightfold drop, and the share-weighted correlation between log size and log price is -0.32. Large orders are overwhelmingly cheap stocks, where a penny of quoted spread is a much larger fraction of the price and there is correspondingly less room inside the quote. Some of the gradient above is that composition rather than size itself.

Controlling for price directly, on the unit-free measure used in the next section, the size coefficient falls from -0.075 to -0.058 and stays highly significant. It retains 77% of its magnitude. Both effects are real and separable, but the absolute measure confounds them, which is why the next exhibit matters more.

So Does the Share of the Spread Returned

Price improvement and effective spread are complementary measures. Improvement is the distance from the quote the client would otherwise have paid. Effective spread over two is the distance from the midpoint. Their sum approximates the half quoted spread the client actually faced, which lets us ask a sharper question than the last one: of the spread that was available, how much did the client keep?

Exhibit 2

Click to expand

The largest orders in the reported range do not merely face a narrower spread. They are returned less than half the fraction that the smallest orders receive, 34% against 73%. Size is penalised twice: once through the narrower spread, and again through the smaller share of it.

Computing this requires joining price improvement and effective spread across the full symbol-level tape rather than reading the summary rows, which is why it does not appear in the usual commentary.

And the Benefit Is Extraordinarily Concentrated

Exhibit 3

Click to expand

Across 12,643 symbols, the Gini coefficient on price improvement dollars is 0.878. For reference, that is a more unequal distribution than income in any country on earth.

Half of all price improvement in the month came from 131 symbols. The top ten symbols account for 20% of it, and the top hundred for 46%.

Improvement tracks spread width, and spread width lives in volatile mid-caps and leveraged products rather than in the mega-caps and index funds most portfolios actually hold. The aggregate figure is real. The typical investor's share of it is not what the aggregate implies.

Where the Improvement Actually Comes From

K-means over five standardised symbol-level features, namely improvement per share, cost versus midpoint, price impact, improvement rate and log volume, separates the universe cleanly.

Exhibit 4

Click to expand

One cluster is 2.2% of executed volume and produces 29.4% of all price improvement dollars, at a median of nearly 17 cents per share. Another is 95.7% of volume and produces 63.8%, at 0.80 cents. The headline number is an average across two populations that barely resemble each other.

Now the Same Question About Your Own Flow

Exhibit 5

Click to expand

If you are an adviser routing through an algorithm, there is no answer. Not a poor answer. No answer at all.

Start with the public order routing report every broker files quarterly under Rule 606. Open the XML and the root element is named heldOrderRoutingPublicReport. The scope is in the schema. Held orders are those a broker must attempt to execute immediately. Not-held orders, where the customer grants discretion over price and time, which is to say essentially all algorithmic execution, are outside it.

The natural response is that Rule 605 will cover it, particularly now. Rule 605 was overhauled in 2024, the compliance date arrived on August 1 of this year, and for the first time brokers introducing or carrying 100,000 or more customer accounts must publish execution quality for their own customers rather than sitting inside a wholesaler's blended tape. The first reports are due by the end of September.

Read the covered order definition in the adopting release:

"Covered order shall exclude any order for which the customer requests special handling for execution, including... orders submitted on a 'not held' basis."

The exclusion survived the overhaul. The most significant expansion of retail execution disclosure in twenty years does not touch discretionary institutional flow.

That leaves exactly one disclosure that does. Rule 606(b)(3) requires a broker, on request from a customer who has placed not-held orders, to produce a customer-specific report within seven business days covering the prior six months by calendar month, including all child orders derived from the parent. It is not published. It goes to the requesting customer and nowhere else.

The Commission understood what it was building. From the adopting release, the rule would "likely result in more Rule 606(b)(3) disclosures for order flow that is typically characteristic of institutional customers, not retail customers."

Every exhibit above exists because retail flow is measured twice over. None of them can be built for adviser flow, at any price, by anyone.

Why This Lands Hardest on Advisers

An adviser's duty of care under the Advisers Act includes, in the Commission's own 2019 interpretation, "the duty to seek best execution of a client's transactions where the adviser has the responsibility to select broker-dealers to execute client trades."

When the Commission has charged advisers with best-execution failures, it has proved the case through the absence of documented, benchmarked, periodic review. That is a reasonable standard when the data exists. For algorithmic orders it creates an awkward position: the obligation is to evaluate execution quality systematically, and the public dataset that would let you do it explicitly excludes your orders.

The academic record confirms this is structural. The only broad institutional transaction-cost dataset ever available to researchers stopped being provided in 2017, and it never captured principal internalization by the firms operating the algorithms. There is no successor.

Note also what Exhibit 2 implies for institutional-sized orders. Within the reported range, the share of the spread returned to the client is already falling steeply by ten thousand shares. Institutional parent orders begin where the tape ends. The measured trend does not tell you what happens beyond it, and no public data does either.

Meanwhile the flow itself is worth more to whoever handles it. A parent order carries size and urgency, which is the most valuable information a manager holds, and it is handed over in full at submission. Retail child orders carry almost none. The market measures the flow that reveals the least and leaves unmeasured the flow that reveals the most.

The Label Is Not the Test

There is a second issue inside the word "not-held," and the SEC has addressed it twice.

The staff FAQ on Rule 606 works through four examples. A fifteen-share order where the customer reasonably expects immediate execution is held. A $100 cash order that produces a fractional component of 15.625 shares is also held, "since the customer reasonably expects Broker-Dealer A to attempt to execute its orders in NMS stocks immediately." A 150,000-share order is not held. A $1,000,000 cash order, also with a fractional component, is likewise not held.

The fractional component appears on both sides of the line. It moves nothing. What moves the answer is the customer's reasonable expectation of immediate execution.

The staff returned to the same principle under Rule 605 this April. Asked how an ATS should treat an immediate-or-cancel order arriving with a not-held instruction attached, the answer was that the subscriber and its customer should expect immediate matching without price or time discretion, and "therefore the ATS should categorize the order as held for purposes of Rule 605 reporting." An explicit not-held instruction did not make the order not held.

Two staff FAQs under two different rules reach the same conclusion. Not-held describes how an order is actually handled and what the customer actually expects. It is not a field a system populates.

That matters because the classification decides whether an order appears in public data at all. FINRA has noticed. Its examination reports have repeatedly flagged firms failing to provide Rule 606(b)(3) reports on time, failing to provide accurate and properly formatted six-month disclosures, and lacking the supervisory systems to verify eligibility when claiming one of the rule's de minimis exceptions. This year's report goes further and flags inaccurate order classification where firms did not consider whether the order actually involved a customer request for special handling.

One thing FINRA has not done is bring a named public action charging a Rule 606(b)(3) failure specifically. We searched and could not find one. Enforcement in this area has run under 606(a), the public report. The customer-facing not-held obligation appears to be examined but never publicly charged, which tells you roughly how often anyone asks.

What To Actually Do

The rule has been on the books since 2018 and almost nobody exercises it. That is the opportunity.

Any customer placing not-held orders can demand the report, and the broker has seven business days. The two exceptions are narrow: a firm is excused if not-held orders were under 5% of the shares it received from all customers over six months, which almost never describes an agency broker or an algorithm provider, or if the requesting customer traded under $1,000,000 average monthly notional, which almost never describes an institution. If a broker asserts an exception, make it identify which one and produce the supporting figure. FINRA already examines whether firms can.

Three fields in that report are worth the effort on their own.

• Total shares the broker executed as principal for its own account, which is how much of your flow the firm traded against itself.
• Orders exposed through an actionable indication of interest, and the venues they went to. This is the only measurement in United States disclosure of where your parent order was shown before it traded.
• The split between shares executed on the more and less favorable side of the spread, read against the midpoint percentage. It is the closest thing to a like-for-like quality comparison you can construct across brokers.

Send the requests to several brokers on the same day, because the reports are only comparable when the periods match.

Two dates are worth marking. The first amended Rule 605 reports arrive at the end of September and will, for the first time, let anyone compare execution quality broker by broker for held retail orders using a true effective-over-quoted spread. The price improvement statistics benchmarked to the best displayed odd-lot price are deferred and do not begin until November data published at the end of December. Neither wave covers not-held flow.

The measurement gap is not closing. It is being formalized. The one lever that exists is a seven-business-day request that costs an email, and the reason it produces nothing today is that almost nobody sends it.

Method and Sources

Retail figures are computed from the Rule 605 monthly files published by six wholesalers for June 2026, comprising 395,112 reported cells, 12,643 symbols and 65.5 billion executed shares of marketable flow, together with the complete FINRA Rule 606 quarterly archive. Market centers are pooled throughout and no firm-level result is reported. Rule 605 publishes pre-aggregated cells rather than individual orders, so every fit is on aggregated cells weighted by executed shares and the standard errors are descriptive rather than order-level inference. All relationships are descriptive; nothing here identifies a causal effect. The legacy tape excludes odd lots and orders of 10,000 shares or more, so the size relationship is estimated only over the reported range. Rule text is quoted from the relevant SEC adopting releases and staff guidance.