The bank that borrowed too much on house prices

Lehman Brothers was a 158-year-old US investment bank. In the mid-2000s it grew fast by buying and packaging home loans, including risky subprime mortgages, and by investing in commercial property. It did this mostly with borrowed money: by 2007 it held roughly 30 dollars of assets for each dollar of its own capital.

When US house prices started falling in 2006 and 2007, those assets lost value, and lenders grew nervous. A later court-appointed examiner found that Lehman had used an accounting move called "Repo 105" to make its borrowing look smaller in its quarterly reports.

In March 2008 another bank, Bear Stearns, was rescued with US government help. Lehman looked for a buyer through the summer. Over the weekend of 13 to 14 September talks failed, and officials declined to put public money behind a deal. On 15 September 2008 Lehman filed for bankruptcy with about $600 billion in assets, the largest bankruptcy in US history. Credit markets froze around the world.

Heavy borrowing meant a small fall in asset values could wipe out Lehman's own capital. Its risk models, like those at other banks, leaned on years when house prices had mostly risen across the country at the same time, so a nationwide fall looked very unlikely. Because banks were tangled together through loans and trades, Lehman's failure spread fear to everyone.

The AI angle: models learn from the past, and a long calm period can make them confident about risks they have never seen. Ask what happens if the rare bad case arrives, and keep a buffer for it.

The idea behind this caseLesson: How machines learn →

1. 1Ask how much a plan relies on borrowed money, and what happens if values drop by 5 or 10 percent.
2. 2Test plans against bad scenarios that have not happened recently, not only against recent history.
3. 3Be suspicious of accounting moves that make a report look better without changing the real risk.
4. 4Keep enough cash or slack to survive a few bad months without needing a rescue.