2.5.5 Failure Modes and Practice

Basic Prerequisites: 02 - Fragmentation & the NBBO and 03 - SOR Logic.

1. Intuition & Practical Objective

The routing logic of pages 03–04 assumes the router sees the venue as it is when the order lands. In production, three things go wrong, and each is a first-principles failure with a dollar number attached:

Stale NBBO → trade-through. The router snapshots the consolidated best quote, then sends the order; by the time it executes, that venue’s quote has moved and a better protected quote sits elsewhere. Executing at the stale, worse price trades through a protected quote - a Reg NMS Rule 611 violation and a real cost.
Latency arbitrage (pick-off). A resting quote is a free option written to anyone faster than the quoter. The expected adverse move over latency $L$ scales like $σ L$ , so a venue that looks cheap can be the most expensive once arrival delay is priced.
Fee-driven misrouting. Optimizing the displayed price rather than the all-in price systematically sends flow to the venue that shows best and charges most (page 04 made this precise).

The practical objective: attach a dollar number to each failure, then state the defense. Every number below is re-executed and reproduced.

The one-sentence essence. “SOR fails structurally in three ways: it acts on a quote that has already moved (trade-through), it loses a race it did not price in (pick-off), or it optimizes the displayed price instead of the all-in price (fee misrouting).“

2. Mathematical Ground Truth & Derivations

Stale NBBO and the trade-through cost. Let the router snapshot the NBBO ask $a_{snap}^{*}$ on venue $L$ at time $t_{0}$ and route there. At execution $t_{1}$ venue $L$ shows $a_{L} (t_{1}) > a_{snap}^{*}$ , while a fresh re-read would find $a_{exec}^{*} = min_{i} a_{i} (t_{1})$ . The per-share loss is

Δ_{TT} = a_{L} (t_{1}) - a_{exec}^{*} \geq 0,

positive exactly when the router trades through a venue whose quote is now better. The Reg NMS Rule 611 constraint forbids this: no execution may occur at a price worse than a protected (automated, displayed) quote. A correct SOR re-validates the NBBO atomically at execution, or sweeps the best protected prices in order.

Latency pick-off. Suppose the reference price diffuses with per-millisecond scale $σ$ and the router/quoters need $L$ ms to react. The expected absolute move over $L$ is

E [∣Δ m ∣] \approx σ L (Brownian scaling),

so the expected loss on $q$ shares is $\approx q σ L$ - sublinear in latency. This is the key scaling law: the pick-off loss grows like the square root of latency, so a 1,000× cut in latency buys ~32× less loss, which is why the arms race has diminishing returns but never stops.

Latency as an effective-price term. The router should not rank on $p_{i} + f_{i}$ but on $p_{i} + f_{i} + σ L_{i}$ - the latency-aware effective price (page 06 develops the full optimizer). A venue with a 1-cent better quote but 8 ms of lag can be strictly worse once $σ L$ is added.

Fee-driven misrouting. The per-share misrouting cost is

Δ_{fee} = i min p_{i} + (fee of the raw-best venue) - i min (p_{i} + f_{i}),

which is positive whenever the price-gap/fee-gap reorder condition $f_{i} - f_{j} = p_{j} - p_{i}$ is crossed. Its magnitude scales with the take fee (up to the SEC’s $0.0030/share cap).

3. Computational Implementation - three failures, three dollar numbers

Quantify the trade-through, the pick-off, and the fee misroute. Standard library only; re-executed and reproduced.

Read the numbers. (1) The stale router locks onto venue L at 100.010 and fills 1,000 shares at 100.050 after the move - $30 worse than the fresh NBBO (100.020 on venue D), and a textbook trade-through of D’s protected quote. (2) Pick-off loss scales as $L$ : going from 0.05 ms to 50 ms of staleness raises the loss from $4.47 to $141.42 on a single 1,000-share quote - ×31.6 for a ×1,000 latency, the unmistakable Brownian signature. (3) The fee-blind router picks the venue displaying 99.996 and pays an all-in 100.0110 versus 100.0040 - $7.00 burned on 1,000 shares by optimizing the wrong number.

4. Failure Modes & First-Principles Breakdowns

Acting on a snapshot. The NBBO is a past observation. Any router that does not re-validate the protected quote atomically at execution is one quote update away from a trade-through and a Reg NMS breach. Defense: atomic re-check, or sweep the best protected prices in strict order.
Underpricing latency. Pick-off loss $\propto σ L$ means a venue can be cheap on the screen and expensive in reality. Latency must be an explicit term in the effective price (page 06), not an afterthought.
Optimizing the displayed price. The fee misroute is silent, uniform, and scales with volume - the most dangerous failure precisely because it never looks like a failure. Rank on the cum-fee price.
The failures compound. A stale-NBBO router that is also slow and fee-blind experiences trade-through × pick-off × fee loss on the same order. Defense is layered, never a single fix. This is the routing analogue of the layered defenses in 06 · Advanced Extensions.

5. References

O’Hara, Maureen & Ye, Mao - “Is market fragmentation harming market quality?” JFE 100(3), 2011.
Hasbrouck, Joel & Saar, Gideon - “Low-latency trading,” JFM 16(4), 2013. Order lifetimes and cancel rates that make latency a first-order routing cost.
Biais, Foucault & Moinas - “Equilibrium fast trading,” JFE 116(2), 2015. *When speed investment is privately profitable but socially wasteful
Colliard, Jean-Edouard & Foucault, Thierry - “Trading fees and efficiency in limit order markets,” RFS 25(11), 2012.
Foucault, Pagano & Röell - Market Liquidity (2013)

6. Connected Graph Bridges

Back: 04 · Fees & Venue Selection · Index Hub
Continue: 06 · Advanced Extensions
Root causes: Adverse Selection & Glosten–Milgrom · Toxic Order Flow & VPIN
Where the losses bite: Queue Position & Fill Probability · Low-Latency Systems Architecture

Kwant Atlas

Explorer