The Cheapest Insurance Ever Sold: The 2007 Subprime Trade

The trade, accurately

In 2006, John Paulson set up two funds at Paulson & Co with a single purpose: buying protection against subprime mortgage debt.

One correction worth making at the outset, because the story is frequently told loosely. The instrument was credit default swaps, not put options. A credit default swap is an insurance contract on a bond: the buyer pays a periodic premium, and if the underlying debt defaults, the seller pays out.

That distinction matters for accuracy and not much for the lesson, because economically a CDS on a bond behaves very like a deeply out-of-the-money put. Small recurring premium, no payoff in almost every period, enormous payoff in the rare event. Both are insurance. The market pricing them was making the same error.

The results are documented: the Credit Opportunity fund returned roughly 590% net in 2007, with a second fund at about 352%. The firm made in the region of $15 billion, and Paulson's personal payday that year was reported at around $3.7 billion — among the largest single-year totals in the industry's history.

Why it was so cheap

The premiums were tiny because everyone selling them was confident nothing would happen.

Three widely shared assumptions supported that confidence.

House prices had never fallen nationally in the post-war period. Regionally, yes. Nationally, in the data anyone was using, no. Models built on that history assigned a near-zero probability to the event that occurred.

Diversification was assumed to work. A bond pooling mortgages from many states was thought to be safe because those regional markets had historically been weakly correlated. The assumption held right up until the driver became national rather than regional, at which point every pool moved together.

The ratings said it was safe. Large parts of these structures carried the highest available ratings, which let institutions hold them without asking further questions.

Each of those was reasonable given the data available and wrong in the same way: they described a distribution built from a period that had not contained the event.

Who was on the other side, and what they were thinking

This is the part that matters for anyone selling premium.

The sellers were not reckless. They were institutions with quantitative teams, risk models and every incentive to get the price right. They looked at contracts that paid a steady premium and had, on their models, an almost negligible chance of paying out.

That is a good trade on those numbers. Collect a small amount reliably, pay out almost never. It is the same shape as selling far out-of-the-money puts, and it looked like the safest income in the market.

They were wrong about one input: the probability of the rare event, and how correlated the outcomes would be when it arrived. Everything else in the analysis was competent.

Being wrong about that one input was sufficient to destroy the position, because the payoff was so asymmetric. Years of premium against a payout many multiples larger. The ratio does not forgive a small error in the tail estimate.

The lesson for a put seller

You are on the same side of the trade those institutions were on. Not in scale or complexity, but in shape: collecting premium for accepting a rare, large obligation.

Three things follow, and none of them argues against selling premium.

Your edge, if you have one, is in the tail estimate. Not in the frequent outcomes — those take care of themselves. Everything depends on whether the rare event is rarer than the premium implies. If your assessment of that comes from a model whose assumptions exclude the event, the model will look excellent until it does not.

Correlation is where the damage compounds. The subprime sellers did not lose one position; they lost all of them at once, because the diversification they relied on evaporated when the driver became common. A put-selling book has the same property in miniature — sector concentration shows it in real settled data, on a single expiration date where more than half a book finished below its strike across thirty different companies.

A record of quiet success is not evidence about the tail. The sellers had years of clean premium collection. It told them nothing about the event that mattered, because that event had not happened yet. Why a 90% win rate can still lose money is the same point in arithmetic.

Where the analogy stops

Two limits, stated so the comparison is not overdrawn.

The mispricing was extraordinary and identifiable. Paulson found a case where the market's model excluded the outcome entirely, and the protection was therefore close to free. That is not the ordinary condition of the options market. Equity index puts are, if anything, persistently overpriced relative to subsequent movement — which is why selling them has been a viable business rather than a trap. The subprime case was an outlier, and treating every premium sale as though it were the same setup would be the wrong conclusion.

The sellers had leverage that a cash-secured seller does not. Much of the damage came from positions financed rather than fully collateralised. A cash-secured put is bounded by the strike and the cash is already set aside. That single structural difference is why the same shape of error is survivable at retail scale and was not at institutional scale.

The useful takeaway

Not "do not sell insurance". Selling insurance is a sound business and the buyers usually overpay — who buys put options covers why.

The takeaway is narrower: the entire outcome of an insurance-shaped strategy lives in the part you have the least data about. Frequent small wins tell you almost nothing. Every question worth asking is about the rare case — how deep, how correlated, how survivable at your size.

Which is why size and correlation are the controls that matter, and why a floor built from how things have actually moved is worth more than one derived from a model that has never seen the event.

How StratosIQ approaches it

The floor is built from behaviour rather than from a pricing model, which is a direct response to this failure mode. ShieldIQ measures each strike against how the underlying has actually moved rather than against the distance a model implies, and reports Fortified, Secure, Tight or Exposed. A contract failing that floor does not publish however rich its premium.

There is also a ceiling. Contracts with unusually wide buffers do not publish either — the band that looks safest by model-implied distance was measured carrying a materially higher loss rate than the baseline, which is exactly the kind of result a model-derived view would miss.

What the platform does not do is manage correlation across your book, which is the other half of what went wrong here. That remains yours.

The full published record, contract by contract and including every loss, is on the performance page.

Worth reading next

Short puts in a market-wide drawdown covers what happened to sellers in 2018, 2020 and 2025. The $4.9 billion bet covers the same trade from the seller's side, done well.