Quant Analytics: from scratch to pro · Professional practice and failure case studies · Pro
Case study: Knight Capital, 1 August 2012
In 45 minutes a market maker's trading software sent millions of unintended orders and lost about $460 million. The SEC's order traces the loss to a deployment error, old code left on a server and missing controls. The case is the standard example of operational and model risk.
In this lesson
- Describe the sequence of the Knight Capital incident from the SEC's findings
- Identify the operational controls whose absence turned a software fault into a large loss
- Compute the rate of loss and of order flow during an incident to size a kill switch
- Explain why this is a model-risk case without any bad forecast
Read first: Model risk and live monitoring, Orders, bid, ask and liquidity
On 1 August 2012 Knight Capital Americas, a US market maker, sent millions of orders it never intended to send. It lost more than $460 million. No forecast was wrong. The loss came from software that did the wrong thing at scale, and from controls that did not stop it. The SEC order of 16 October 2013 is the source for most of what follows.
What happened
Knight's router, called SMARS, sent orders to the market on behalf of customers. In the days before, the firm rolled out a new function for a US retail liquidity programme. The release went to seven of eight servers. The eighth kept an old routine, Power Peg, which had not been used for years. The new release reused the flag that had once switched Power Peg on, so orders carrying it on the eighth server woke the old code.
While Knight processed 212 small customer orders, its router sent more than 4 million orders in the first 45 minutes. They produced more than 4 million executions in 154 stocks for more than 397 million shares. Knight lost more than $460 million on the unwanted positions.
An internal system sent 97 automated e-mails about the problem before the open. They were not acted on.
The mechanism
Deployment error. One of eight servers was missed, and no one noticed. A second person checking the roll-out would have found it.
Dormant code. Unused code stayed on a live server. A repurposed flag meant a message intended for new code could reach old code.
Testing and removal. The old routine was not removed and not covered by the testing of the new release.
Missing controls. The SEC found that Knight lacked adequate controls immediately before orders reached the market, such as one comparing orders leaving the router with orders entered. Limits that halted the flow automatically did not stop it.
Alerts without action. The automated e-mails were not acted on, so the fault ran on.
Worked example: what the numbers say
Take a loss of $460 million, the reported figure being "more than" this and the 45 minutes. The average loss was 460 / 45, which is $10.2 million a minute. On more than 397 million shares the average was 460 / 397, which is $1.16 a share, and the flow was about 8.8 million shares a minute. Using 4 million executions as a floor, that is at least 1,481 executions a second, or 18,868 executions per customer order.
Each unintended share cost little. The scale and speed made it large, and each minute without a stop added about $10.2 million. A control that halts trading when order count, position size or loss per minute passes a limit works at those rates. The chart shows the constant-rate approximation, which is only an average.
Lessons for a quant
- Operational and deployment controls: roll out changes to every server or none, verify what version is running, and have a second person sign off. Remove dead code.
- A kill switch: automatic limits on order rate, gross position and loss, set before the system goes live and tested in a rehearsal.
- Real-time monitoring beyond profit and loss: order count, order-to-fill ratio and position size warn earlier. See live monitoring.
- Alerts need owners: an alert that nobody must answer is a log entry.
- Model risk without a bad forecast: SR 11-7, the US guidance in force at the time, treats a model used outside its design as model risk. Software that executes decisions is a model in this sense, and it needs effective challenge from people independent of its builders.
- Liquidity and leverage: a firm can only absorb a fault of a given size. Size limits to what capital can survive.
- Rehearsed response: know in advance who can stop trading and how.
What changed afterwards
On 16 October 2013 the SEC ordered Knight to pay a $12 million penalty and retain an independent consultant to review its controls. It was the SEC's first enforcement action under the market access rule, Rule 15c3-5, which requires broker-dealers to have risk controls on orders before they reach the market.
Common mistakes
- Treating the case as bad luck. Each failure was a known type of control gap.
- Testing the new code and not the old code it sits beside.
- Checking profit and loss only. By then the damage is done.
- Assuming limits exist because the software is old. Write them down and test them.
- Deploying changes by hand with no checklist.
- Using the loss figure without its caveats. Accounts differ in how they round the number.
In practice
Before any rule trades live, write the kill criteria: the order rate, position and loss at which it stops itself, who can stop it and how fast. Rehearse the stop. Trading on FinUties starts on a demo account, where ideas like this can be practised without real money.
Cumulative loss at a constant rate over the incident
Worked numbers
Each result is recomputed from the formula shown.
| What | Formula | Result |
|---|---|---|
| loss in millions of US dollars | 460 | $460 million |
| minutes of trading | 45 | 45 minutes |
| loss per minute | loss_m / minutes | 10.2 |
| shares in millions | 397 | 397 million |
| loss per share | loss_m / shares_m | 1.16 |
| shares per minute in millions | shares_m / minutes | 8.8 |
| executions per second, using 4 million as a floor | 4000000 / (minutes * 60) | 1,481 |
| executions per customer order, using 4 million as a floor | 4000000 / 212 | 18,868 |
Check yourself
Roughly how fast was Knight losing money during the incident?
About $460 million over about 45 minutes is about $10.2 million a minute.
What does the loss per share say?
$460 million over more than 397 million shares is about $1.16 a share on average. Each unintended share was cheap; the volume was the problem.
Name two controls that could have limited the loss.
A limit on gross position or order count that blocks orders automatically (a kill switch), and a check comparing orders leaving the router with those entered. A second-person review of the deployment would also have caught the missing update.
Why is this a model-risk case without a bad forecast?
SR 11-7 counts a model used outside its intended design as model risk. Old code ran on live orders under a repurposed flag, so the software did something it was never meant to do.
What should be monitored in real time besides profit and loss?
Order count, order rate, position size, fill ratio and message errors. These move before profit and loss does.
Key terms
Checked facts
Every figure below carries how it was checked: Primary checked (read in the cited primary, official, academic or data source), Two sources agree, or Approximate (an order of magnitude only).
- Primary checked
The incident took place on 1 August 2012 at Knight Capital Americas LLC, a US market maker.
- Primary checked
In the first 45 minutes after the open, while attempting to fill 212 customer orders, Knight's router sent more than 4 million orders into the market.
- Primary checked
The orders produced more than 4 million executions in 154 stocks, for more than 397 million shares.
- Primary checked
Knight lost more than $460 million from the unwanted positions.
- Two sources agree
New code was installed on seven of Knight's eight servers; the eighth kept old code (Power Peg), unused for years, which the new release's repurposed flag woke up.
- Two sources agree
An internal system sent 97 automated e-mails about the problem before the market opened; they were not acted on.
- Two sources agree
On 16 October 2013 the SEC ordered Knight to pay a $12 million penalty and retain an independent consultant, in its first enforcement action under the market access rule (Rule 15c3-5).
- Primary checked
The SEC found that Knight lacked adequate controls immediately before orders reached the market, such as one comparing orders leaving the router with orders entered.
Sources
- Order Instituting Administrative and Cease-and-Desist Proceedings against Knight Capital Americas LLC, Exchange Act Release No. 70694. U.S. Securities and Exchange Commission, 16 October 2013, 2013. Accessed 9 Nov 2026
- SEC News Digest, 16 October 2013 (Knight Capital Americas LLC). U.S. Securities and Exchange Commission, 2013. Accessed 9 Nov 2026
- SEC Charges Knight Capital With Violations of Market Access Rule. U.S. Securities and Exchange Commission, press release 2013-222, 16 October 2013, 2013. Accessed 9 Nov 2026
- Knight Capital Settles Rule 15c3-5 Violations with SEC, Agrees to Pay $12 Million. WilmerHale client alert, October 2013, 2013. Accessed 9 Nov 2026
- SEC order explains Knight Capital systems failure. J. R. Varma, blog, 2013, 2013. Accessed 9 Nov 2026
- SR 11-7: Supervisory Guidance on Model Risk Management. Board of Governors of the Federal Reserve System, with the OCC, 2011. Accessed 9 Nov 2026