In early 2012, a portfolio inside JPMorgan Chase’s Chief Investment Office was breaching its risk limits. Then a new value-at-risk model arrived, and the problem seemed to vanish. Under the new model, the portfolio’s reported risk fell by roughly half.
Within months, the trades later known as the “London Whale” had generated losses of more than $6 billion. When the bank investigated, its own review found that the model ran on spreadsheets fed by manual copy-and-paste steps, and that its calculations contained errors.
The trades themselves were dangerous. Far more unsettling was the figure that called them safe.
This contradiction sits at the heart of algorithmic trading risk management. Every risk system is also a model: designed by people, written in software, and capable of error. Institutions that survive have learned to ask a question most traders never consider: who measures the measurer?
Why VaR Is Not Enough
Value at risk won its popularity by reducing a portfolio’s danger to a single figure. That simplicity is also its weakness. It describes what losses look like on ordinary days and says almost nothing about how bad the worst days can get.
Regulators absorbed the lesson. The Basel Committee’s revised market risk framework moved bank capital models from VaR to expected shortfall, a measure that asks what the average loss looks like once you are already in the tail.
This principle extends well past the banking sector. Serious quantitative trading research treats risk metrics with the same skepticism it applies to alpha signals. Metrics are tested against historical crises, their assumptions are documented, and any methodological change is independently reviewed. A risk model that cannot be challenged is just an opinion with decimals.
Limit the Execution, Not Just the Exposure
The Flash Crash struck U.S. markets on May 6, 2010. The joint SEC and CFTC report identified a key trigger: a sell program of 75,000 E-mini S&P 500 futures contracts, executed by an algorithm that targeted a share of trading volume without regard to price or time.
Nothing about the position was exotic. The execution simply had no constraints.
Institutions now write execution limits directly into AI trading system development. These include price collars that block orders far from the prevailing market, participation caps, order-rate throttles, and checks that pause an algorithm when conditions change faster than it was designed to handle. Exposure limits set how much you can lose. Execution limits set how fast.
Layer the Defenses
Mature risk architectures rarely rely on one control. They stack them.
Before an order leaves the firm, pre-trade checks reject it if it violates size, price, or credit limits. Intraday monitoring follows positions and losses in real time, using hard limits that stop trading and soft limits that trigger escalation. Post-trade review then compares what happened with what was website supposed to happen.
Every layer assumes the previous one may fail. That is not pessimism. It is engineering.
Treat Risk Code as Critical Code
The London Whale model failed partly through ordinary software problems: manual processes, insufficient testing, and weak change control. Institutional algorithmic trading software development now treats risk logic with the rigor usually reserved for trading engines. Changes are version-controlled and peer-reviewed, and a new model runs in parallel with the old one before it takes over.
Historical crises are replayed through the systems as well, to verify that the controls really trigger. Here AI trading infrastructure consulting often adds the most value, since an outside team with no attachment to the existing design can test whether safeguards hold under conditions their builders never imagined.
Give Breaches Authority
A limit is only as strong as the response when it is crossed. In disciplined institutions, a breach triggers escalation to people independent of the trading desk, and the default response is to reduce risk, not to revisit the limit.
The riskiest moment in any risk framework comes when someone explains a breach away.
The Four-Question Audit
Before you trust a risk framework, ask four questions:
Was the risk model validated by people independent of its owners?
What does the system do in the tail, not just on an average day?
Can an algorithm lose money faster than a human can intervene?
When a limit is breached, which changes: the portfolio or the limit?
The Number That Feels Like Safety
Risk numbers hold an odd emotional power. They help people sleep and let committees approve. They make uncertainty look like knowledge.
Institutions surround these numbers with doubt for exactly that reason.
The London Whale trades did not hide. They sat in plain view and were measured every day. The danger was not invisible. It was written into a number everyone wanted to believe.