Skip to content
Rabin Poudyal
Rabin Poudyal
← All models

Decision-making and strategy

Margin of safety

Leave a wide gap between what you expect to need and what you build or pay, so that your errors and bad luck do not sink you.

Benjamin Graham, The Intelligent Investor (first published 1949), whose final chapter distils sound investment into three words: "margin of safety". Engineers had long built the same idea into structures as a safety factor; Warren Buffett drew the comparison in his 1984 talk "The Superinvestors of Graham-and-Doddsville".

Graham told investors to buy only when the price sat well below their estimate of what a business was worth. The gap is the margin of safety. Buffett put it more bluntly: you do not buy a business worth $83 million for $80 million. When you build a bridge, he said, you insist it can carry 30,000 pounds, but you only drive 10,000-pound trucks across it.

Every estimate is wrong, and you cannot know in advance which way or by how much. A margin does not need to know. It only assumes that some error will come, and makes sure that being wrong costs an inconvenience rather than a collapse. It is insurance against your own model, and it is cheap compared with the cost of being exactly right and then unlucky.

A margin costs something: idle capacity, deals you pass on, slower delivery. Too much is waste. The more common failure is a margin that quietly disappears, eaten by later changes that nobody checks against the original assumptions. A margin also covers errors of size, not errors of kind. If the model leaves out a whole class of load, a factor of three on the loads it includes is false comfort.

Example

Citicorp Center in New York was finished in 1977. Its design called for welded joints in the wind bracing, but on 1 August 1974 the structural engineer William LeMessurier's New York office had accepted Bethlehem Steel's proposal to use bolts. In June 1978 a call from an engineering student led LeMessurier to check winds striking the building at its corners. Such quartering winds raised the strain in four of the eight chevron braces in each tier by 40 per cent, a 160 per cent increase in load on the bolts. If the tuned mass damper lost power, the weakest joint could fail in a storm expected once every 16 years. From August to October 1978, with Hurricane Ella briefly forecast to reach New York, welders fixed two-inch-thick steel plates over more than 200 bolted joints.

Where it applies

Plan servers, queues and database connections to run at well under full capacity, so a traffic spike or a slow dependency is absorbed rather than cascading. Give an estimate with slack you can explain, keep a large language model (LLM) prompt well inside its context window and rate limits, and keep personal savings that cover months, not weeks.

Think about it

Where in your work are you running at close to full capacity, and what happens on the day your estimate turns out 30% too optimistic?

Applied it to something real? Log it.

Test yourselfWhy is a margin of safety useful even when you cannot say which of your estimates is wrong?

It only assumes that some error or bad luck will come, and leaves enough room that being wrong is survivable. You do not need to predict the error to be protected from it.

Related models

Sources and further reading