Discussion of

"Best Execution Puzzles"

by Indira Puri, Mohan (Fred) Sun, and Mao Ye

Discussant: Katya Malinova

DeGroote School of Business, McMaster University

53rd Annual Meeting of the European Finance Association · August 19–22, 2026
Vlerick Business School, Ghent, Belgium

Some venues are both slower & more expensive. Puzzle: why do they exist?

Why do apparently dominated venues get flow?

Typically ranked on two dimensions: cost & speed.
This paper: a third dimension: timing reliability.

Is timing uncertainty priced?

Yes. And the tail (non-execution) matters most.

How to address?

Rule 605 data. 2013–2016.
Timing histogram to get the execution time vol
80 off-exchange venues.

Two approaches

Reduced form

Execution time dispersion from Rule 605 bins: <9sec, 10-29, 30-59, 1-5min, 5-30min

Regress effective spread on speed, speed volatility, non-execution

Is execution timing uncertainty priced?

Venue-choice model

Heterogeneous trader preferences over venue attributes

Fitted to observed venue shares

Can we rationalize coexistence?

Data note: 22.6m venue–stock–month cells → 2.6m used. Most cells have zero measured dispersion (single bin); this also removes exchanges, mostly in the <9sec bin.

→ A venue enters only through cells with non-zero dispersion. Conditioning on "hard to fill"?

Q1-A: how much of a puzzle is this?

Effective spread (bps) costlier ↑

Time to fill (seconds) slower →

Q1-B: might we expect slow to be expensive?

605 data: “slow” = very slow → over 9 seconds
Orders are marketable on arrival.

Either fill on arrival

→ at or inside the quote →low spread

Or don't fill, and rest

→ wait for the market to come to your limit
→high spread (relative to mdpt @entry)

→ slow can be mechanically more expensive.
→ why not execute? low depth? a liquidity effect?
→ is the bigger puzzle fragmentation/existence of illiquid venues (many explanations

As an aside: once resting, the order supplies liquidity — execution time is now about limit price and queue position (Lo, MacKinlay & Zhang; Yueshen).

Q1-A: how much of a puzzle is this?

Effective spread (bps) costlier ↑

Time to fill (seconds) slower →

Q1-B: might we expect slow to be expensive?

605 data: "slow" = very slow → over 9 seconds
Orders are marketable on arrival.

Either fill on arrival

→ at or inside the quote → low spread

Or don't fill, and rest

→ wait for the market to come to your limit
→ high spread (relative to mdpt @entry)

→ slow can be mechanically more expensive.

As an aside: once resting, the order supplies liquidity — execution time is now about limit price and queue position (Lo, MacKinlay & Zhang; Yueshen).

And: why didn't it execute? Low depth? Is speed volatility partly a liquidity measure?

And: is the broader puzzle the existence of illiquid venues?

Q2: what is being priced?

Speed vs dispersion ✓

Strongly co-move; the coarse bins make it challenging to separate. BUT: the paper's ETF results help here.

Dispersion vs tail risk

The paper does distinguish these. BUT: the strongest evidence is for extreme delay or non-execution.

→ Is the object execution speed volatility — or non-execution risk?

Good news, bad news: amended Rule 605 collection began August 1

much finer time-to-execution buckets
median and 99th percentile of time to execution reported directly

Q3: bifurcation of trading venues?

Low-fill venues look faster and cheaper (conditional on execution; Table 1, Panel C — raw means).

"Unfilled" here ≠ never executed: non-execution at the receiving venue can include routing elsewhere.

Is the low spread partly a consequence of selective execution and/or venue business model?

Q4: what explains the pattern?

Less-predictable execution timing ↔ lower effective spreads.

Table 6: ATSs are more predictable & costlier → architecture matters. But through what channel?

Paper: rule-based ATS matching vs greater non-ATS execution discretion.

Demand (the paper)

Traders pay more for predictable timing.

Screening

Uncertain fills can screen out informed traders. (also, Zhu, 2014)

Selection / access

Different mechanisms allow & attract different orders. Often the broker routes.

Citadel and Knight internalize; Instinet, Level, UBS ATS match (dark).
Not freely routing to all 80: PFOF, affiliation, connectivity.

Heterogeneous preferences — or heterogeneous choice sets

→ does this affect the venue-choice model?

Would Rule 605 realized spread help tell some of these apart?

One puzzle solved: execution timing uncertainty is priced.
Would love to know more detail on the next one: why.

Fun paper.

Go read it.