The genius fund that trusted its models too far

Long-Term Capital Management (LTCM) was founded in February 1994 by John Meriwether, a former Salomon Brothers vice chairman. Its team included the economists Myron Scholes and Robert Merton, who shared the 1997 Nobel Prize in economics for their work on pricing options. The fund bet on small price gaps between similar bonds closing over time, and returned about 43% in 1995 and 41% in 1996.

To make small gaps pay, it borrowed heavily. According to the US government's later report, it held over $125 billion in assets on about $4.8 billion of its own capital at the start of 1998, a ratio of more than 25 to 1.

On 17 August 1998, Russia devalued its currency and stopped paying some of its debts. Investors everywhere rushed to the safest assets, and the price gaps LTCM expected to shrink grew wider instead. It lost $1.8 billion in August alone, and by the end of that month more than half the capital it started the year with was gone.

Fearing that a forced sale of its huge positions could disrupt markets, the Federal Reserve Bank of New York brought its lenders together. On 23 September 1998, 14 banks and brokers put in about $3.6 billion for 90% of the fund. No Federal Reserve money was used. The fund's positions were wound down, and the group got its money back by the end of 1999.

LTCM's models learned how bond prices had moved in the past and assumed they would keep moving in similar ranges. Testifying to Congress in October 1998, Fed chairman Alan Greenspan said the market turmoil was far outside the experience built into its models. A government report the next year said plainly that LTCM's models had underestimated the risk it was taking.

The danger was not one wrong forecast. It was confidence: a model that had worked for years made rare, extreme events look almost impossible, and heavy borrowing turned those rare events into a near collapse. Economists call this tail risk, the small chance of a very large loss.

AI systems share the weakness. A model can give a precise, confident answer about a situation it has never seen, because its confidence comes from past data, not from knowing its own limits. The worst failures tend to come exactly when conditions are new.

The idea behind this caseLesson: Why AI makes things up →

1. 1Treat a model's confidence score as a guess about the past, not a guarantee about the future.
2. 2Ask what conditions the model was trained on, and list the situations it has never seen.
3. 3Test the system on rare, extreme scenarios on purpose, not only on typical days.
4. 4Never let one model's output carry so much weight that a single bad week can sink you.
5. 5Keep human judgement in charge when the world changes suddenly.