Execution algorithms

Smart order routing

∞structural
By . Reviewed 22 October 2026. As of 2026: a permanent feature of the market, not an edge that decays.

One instrument, many venues: a router decides where each child order goes to get the best price net of fees, latency and fill probability. The defensive twin of latency arbitrage.

The idea

One child, many venues: the cheapest quote is not the best fill annotated diagramDG-ROUTER
best bid 50.00 (Venue A) best ask 50.01 NBBO is assembled across venues, not one place child order sell 5,000, mid 50.00 Venue A (lit) – 2,000 bid 50.00 taker fee 0.30 bps → net 49.9985 Venue B (inverted) – 1,500 bid 50.00 rebate 0.20 bps → net 50.0010, beats A Dark pool – hidden size at 50.005 mid no display, less leakage · toxic? markout down-weights it rest dark + take rebate first Naive "hit displayed best" → ~49.997 net, leaked Fee + toxicity-aware → ~50.003 net (~1 bp better) Reg NMS Rule 611

What this shows. The same instrument trades on many venues, so one child order is routed, not just sent. The venue with the cheapest displayed quote (Venue A) is neither the cheapest net price (Venue B, after its rebate) nor the best outcome (the dark mid, if the fills are clean). A router that prices fees, toxicity and hidden liquidity reaches about 50.003 net here, roughly a basis point better than naively hitting the displayed best, and leaks less – all within Reg NMS order protection.

Why routing exists: one stock, many venues

A stock trades on many venues at once, so the "market" is an aggregate, not a place. The SEC's list of national securities exchanges showed 29 registered exchanges in January 2026 (equities and options together), and its Form ATS-N page listed 35 operating alternative trading systems for NMS stocks in March 2026, among them Goldman Sachs SIGMA X2, UBS ATS, JPM-X, Instinet CBX and Liquidnet H2O. Most of those systems do not display their orders. The best bid and the best ask can sit on different venues, so there is no single place to send an order; the router has to find where the liquidity, and the best net price, actually is.

The NBBO is assembled, not located. The National Best Bid and Offer is the best bid across all venues and the best ask across all venues, frequently on two different exchanges.

The NBBO is the maximum bid and the minimum ask taken across every venue's top of book: a synthesised quote, not the quote of any one venue.
NBBO=(max⁡v bv,    min⁡v av)\text{NBBO} = \Big(\max_{v}\, b_v,\;\; \min_{v}\, a_v\Big)

The router's first job, per child order, is to know where the opposite-side liquidity lives. The mechanics of that synthesis are covered under market fragmentation, whose widget lets you build the NBBO across venues by hand. Crypto fragments too: the same coin trades on many centralised and decentralised venues with no consolidated tape and no order-protection rule, so routing there is entirely the executor's problem (see market making in crypto). Fragmentation cuts both ways. Competition between venues can tighten the touch, and Foucault and Menkveld (2008) study how that competition for order flow interacts with smart routing, but it also means more places to be slow and more routing complexity.

Lit and dark liquidity

Lit venues display their book, so you can see the depth, but so can everyone else, and resting there leaks your intent. Dark venues hide resting orders and typically execute at the midpoint, so you can trade size without showing it, at the cost of uncertain fills and the risk of meeting informed flow.

On lit venues you get displayed depth and price–time priority; the price is the signalling cost described in why execution algorithms exist. In the dark you can rest a large child without display, but you fill only when a counterparty arrives, and you may be picked off by informed traders or probed by pinging and sniffing orders hunting for hidden size. The research on dark trading is mixed in a useful way. Comerton-Forde and Putniņš (2015), using Australian data, find that dark trades are less informed than lit trades, that low levels of non-block dark trading are harmless or even helpful, and that high levels harm price discovery.

The router therefore sprays small slices across lit venues to take displayed liquidity and rests selectively in dark venues for hidden liquidity, continuously judging which dark pools give clean fills.

Spray or sequential?

Spray routing fires slices at several venues simultaneously; sequential routing tries one venue, then the next. Spray grabs fleeting displayed size before it moves but advertises you widely and can over-fill; sequential is stealthier and fee-aware but slower, so the liquidity may vanish while you work through the first venue.

With spray, orders go to every venue showing size at your price. You win the race against rival routers chasing the same quote, but if several hit at once you can fill more than intended. With sequential routing you try the cheapest or best venue first and cascade, which lets you prefer rebate venues, but the size you saw on venue 2 may be gone when venue 1 is done, the stale-quote problem behind latency arbitrage. In practice routers blend the two: spray the displayed top of book to take what is fleeting, then work the residual sequentially, posting passively where rebates and fill quality justify it.

Venue toxicity and fill quality

A venue is toxic for you if its fills are systematically followed by the price moving against you, meaning informed flow is picking you off there. Smart routers measure realised fill quality per venue (markout, fill rate, reversion) and route away from toxic venues even when the displayed price looks best.

Markout is the standard measure: track the mid a few hundred milliseconds and a few seconds after each fill. If your buys are consistently followed by a falling price, that venue's flow is picking you off, which is the adverse selection a market maker fears, seen from the executor's seat.

Per-fill markout: the signed mid move a short horizon Δt\Delta t after a fill, averaged over a venue. A persistently negative markout for your side means the venue is toxic for that flow.
markout(Δt)=ε (mt+Δt−mt),ε=+1 buy,  −1 sell\text{markout}(\Delta t) = \varepsilon \,\big(m_{t+\Delta t} - m_{t}\big), \qquad \varepsilon = +1\ \text{buy},\; -1\ \text{sell}

Price and cost are different things. The cheapest displayed price is not the cheapest net price once maker–taker fees and rebates are included, and chasing rebates can backfire: Battalio, Corwin and Jennings (2016) find that brokers route limit orders to the highest-rebate venues and that limit-order execution quality is lower where fees and rebates are highest. A good router minimises net-of-fee, toxicity-adjusted expected cost rather than the headline quote. Since the 2018 amendments to Rule 606, institutional clients can ask their broker for a six-month report on not-held orders, venue by venue, including average net fees or rebates, which makes this behaviour auditable (SEC press release 2018-253).

The constraint: order protection in US equities

In US equities routing is not a free choice. Rule 611 of Regulation NMS, the Order Protection Rule, requires each trading center to maintain written policies and procedures "reasonably designed to prevent trade-throughs" of protected quotations (17 CFR 242.611(a)). A trade-through is a trade at a price below a protected bid or above a protected offer, and a protected quotation is the automated best bid or offer of an exchange, so the rule covers the top of each exchange's book, not its depth (17 CFR 242.600(b)). It was adopted in the SEC's 2005 release on Regulation NMS (34-51808); the wider rule set is described under Reg NMS.

A compliant router therefore consults every protected top of book before executing. If venue B shows a better bid than the price at which you are about to sell, you must reach that quote or trade at a price at least as good. The standard tool is the intermarket sweep order (ISO): a limit order marked as such, sent together with orders that take the full displayed size of every better-priced protected quote, so the router can hit several venues at once. Rule 611(b) also lists exceptions, including single-priced opening and closing transactions, crossed markets, system failures and a one-second window for flickering quotes. Order protection constrains where a child can legally execute; fees, toxicity and dark availability decide which of the legal options is best.

Other regimes differ. The EU has no trade-through rule of this kind; MiFID II Article 27 instead requires firms to take all sufficient steps to obtain the best possible result for the client (see MiFID II). In the EU, the double volume cap on dark trading (4% per venue, 8% EU-wide) ended on 28 September 2025 and a single volume cap applies from October 2025 (ESMA).

Worked example: selling 5,000 shares across venues

The figures are illustrative. Child order: sell 5,000 shares in a stock quoted 50.00 bid, 50.01 ask, mid 50.005. The book is fragmented: Venue A shows a 2,000 bid at 50.00 with a taker fee of $0.0015 a share (0.3 bps on a $50 stock); Venue B, an inverted venue, shows a 1,500 bid at 50.00 but pays a rebate of 0.2 bps ($0.001 a share) to take; and a dark pool is believed to hold midpoint liquidity at 50.005.

Naive "hit the displayed best". Sweep A and B for 3,500 at 50.00, pay A's fee, and push the residual 1,500 down to the next lit level at 49.99. The net realised price is around 49.997 after fees, and your selling has been shown on two lit venues.

Fee-and-toxicity-aware routing: rest in the dark at the 50.005 mid, take the rebate venue B first (its net price beats A despite the same quote), and only sweep A for the residual, about 1 bp better net, with less leakage.
pBnet=50.00+0.20 bps  >  pAnet=50.00−0.30 bps,pˉnet≈50.003p^{\text{net}}_{B} = 50.00 + 0.20\,\text{bps} \;\gt\; p^{\text{net}}_{A} = 50.00 - 0.30\,\text{bps}, \qquad \bar{p}^{\text{net}} \approx 50.003

If the dark pool fills cleanly you realise about 50.003 net, roughly 1 bp better than the naive route (0.006 on a $50 stock is 1.2 bps), and you leaked less. If the pool is toxic, with fills followed by the bid dropping, the router's markout history down-weights it and the lit route wins. The cheapest displayed price (A) is neither the cheapest net price (B, after the rebate) nor the best outcome (the dark midpoint, if clean). Routing is a net-of-fee, toxicity-adjusted optimisation, re-solved for every child, and its inputs move whenever a venue changes its fee schedule.

Sources

  1. Foucault, T. and Menkveld, A. J. (2008). "Competition for order flow and smart order routing systems". Journal of Finance 63(1), 119–158. doi.org
  2. Battalio, R., Corwin, S. A. and Jennings, R. (2016). "Can brokers have it all? On the relation between make-take fees and limit order execution quality". Journal of Finance 71(5), 2193–2238. doi.org
  3. Comerton-Forde, C. and Putniņš, T. J. (2015). "Dark trading and price discovery". Journal of Financial Economics 118(1), 70–92. doi.org
  4. SEC (2005). Regulation NMS, final rule, Release No. 34-51808. sec.gov
  5. 17 CFR 242.611, "Order protection rule", and 17 CFR 242.600, definitions. law.cornell.edu
  6. SEC. "National Securities Exchanges" (updated 7 January 2026) and "Form ATS-N Filings and Information" (updated 30 March 2026). sec.gov
  7. SEC press release 2018-253 (2 November 2018), "SEC Adopts Rules That Increase Information Brokers Must Provide to Investors on Order Handling". sec.gov
  8. ESMA. "[Discontinued] Double Volume Cap Mechanism". esma.europa.eu

Where this fits