Thirty-nine years ago tomorrow, on October 19, 1987, the Dow Jones Industrial Average fell 22.6 percent in a single session. It is still the largest one-day percentage drop in the index's history. The popular memory of the day is panic: traders shouting, phones ringing unanswered, a crowd losing its nerve. The official finding was something colder. The Brady Commission, the task force that investigated, concluded the crash was driven by structural flaws in market mechanisms, not by a collapse in economic fundamentals. The machines that sold were not broken. They were doing exactly what they had been told to do, all at once, by everyone.
Quick disclaimer before I get into it. The trading I mention in this post is a job I did about a decade ago. It does not reflect my current role or my current employer, and nothing here is about either.
What Portfolio Insurance Was
The strategy at the center was called portfolio insurance. The idea was sensible on its own: a large fund would protect itself against a falling market by selling index futures as prices dropped, so that losses on the stocks were offset by gains on the short futures. The rule was mechanical. Prices fall this much, sell this much. Every fund running it had a similar rule.
| Step | Locally rational action | Aggregate effect |
|---|---|---|
| Prices fall | The model says sell futures to stay hedged | Futures fall |
| Futures fall below the index | Arbitrage desks sell stocks and buy the cheap futures | Stocks fall |
| Stocks fall | The model says sell more futures | Futures fall further |
| Volume overwhelms the floor | Specialists and the futures market cannot absorb it | Prices gap, and the models sell into the gap |
Every row is a reasonable decision by a reasonable actor. Nobody in the loop had the aggregate. The insurance worked exactly as designed for any one fund, and the design did not contemplate every fund executing it in the same hour against the same market.
Loops, From the Inside
I traded United States Treasury and commodity futures at high frequency from 2013 to 2016, and I have written about that job before. The thing it teaches about automation is simple and unpleasant. Every automated strategy has a trigger. When many strategies share a trigger, the market stops being a market of many strategies and becomes one strategy, very large, with no one in charge of it. The Treasury market had its own version of a mechanical stampede in October 2014, a flash move that nobody ordered. The lesson I took was not that automation is dangerous. It was that the thing that saves you is the kill switch, and the person who owns it, and the observable condition that tells them to pull it.
Circuit Breakers Are a Gate
The regulatory answer to 1987 was the circuit breaker: a rule that halts trading for a fixed period when the index falls by a fixed amount, and closes the market for the day at a larger one. Notice the shape. It does not ask anyone whether the selling is reasonable. It fires on an observable signal and produces a fixed action. That is the same shape as every gate I have built into my own work, and for the same reason. A preference does not survive a stampede. A rule does.
The Loop Is Being Rebuilt
Here is why I am writing about a crash I know only from the record. The systems I now build and use are chains of artificial intelligence (AI) agents acting on each other's output. One summarizes, one decides on the summary, one executes the decision. They are trained on similar data, they respond to similar signals, and they run fast. That is portfolio insurance with a different asset class.
| 1987 | Agent systems today |
|---|---|
| A portfolio insurance model with a mechanical sell rule | An autonomous agent with a policy it applies without judgment |
| Index arbitrage acting on the futures discount | A downstream agent acting on an upstream agent's output as if it were fact |
| Specialists overwhelmed by volume | One human reviewer with a queue of two hundred agent-generated items |
| The circuit breaker, adopted after | The halt condition, which should be designed in before |
This is not a prediction of a crash. It is a design note. Any system where automated actors respond to each other needs three things the equity market did not have on that Monday: an observable signal that means stop, a fixed action when it fires, and a named person who owns the switch. In my own practice that is the preflight gate before any irreversible action and the human at the merge. The Brady report's lesson was never that the automation was wrong. The automation was fine. The system had no view of itself, and no way to stop.