The Model Had Never Seen It

Securitisation is a good idea. Take a thousand mortgages, pool them, and sell claims on the resulting cash flow. Because the loans are in different places to different people, some will default and most will not, and the pool as a whole is far more predictable than any individual loan.

Then slice the claims into tranches. The senior tranche is paid first; the junior tranche absorbs the first losses. A conservative investor buys the senior tranche and a risk-seeking one buys the junior, and both get what they want.

None of this is a trick. It is the same logic as insurance, and it genuinely reduces risk: under one condition, which is that the loans are not correlated.

That was the assumption, and it was calibrated on data. The models were built on decades of American mortgage performance, in which defaults clustered regionally (an oil bust in Texas, a factory closing in Michigan) and a default in Florida told you very little about a default in Nevada.

There was no national house price decline in the data, because there had not been one since the 1930s.

If house prices fall everywhere at once, the loans stop being independent. Every mortgage in the pool is exposed to the same event, the diversification disappears, and the senior tranche that was rated as safe as government debt starts taking losses. A model that has never observed a stress will confidently report that the stress cannot happen, and the confidence is generated by the absence rather than by evidence.

The ratings failure compounded this in a way that was entirely visible at the time.

Credit rating agencies are paid by the issuers of the securities they rate. An agency that rated conservatively lost the business to one that did not, and the issuers could shop between three of them. Internal emails later disclosed show analysts describing deals they were rating in terms that make clear they understood what was happening.

This is a straightforward conflict of interest that everybody knew about and nobody fixed, because the arrangement worked for all the parties present. The issuer got the rating, the agency got the fee, and the investor got a rating that satisfied their own mandate to hold highly rated paper. The party who would have objected was the investor a decade later.

Add originate-to-distribute. A mortgage broker who sells a loan on within weeks bears no default risk, so the incentive to check whether the borrower can repay collapses. Add capital rules that made highly rated tranches very cheap for banks to hold, which created enormous demand for exactly these instruments. Add leverage of thirty to one at several investment banks, so a 3% fall in asset values wipes out equity.

Now the mechanism by which it became a systemic crisis rather than a large loss, which is the part with the most transferable content.

A great deal of the financial system funded itself in the repo market: overnight secured borrowing, rolled over continuously. You post collateral, receive cash, and repay tomorrow. It is cheap because it is short and secured, and it means an institution holding thirty-year assets can be funding them with money it must re-borrow every morning.

When confidence in mortgage collateral broke, lenders began demanding more collateral for the same cash. A higher haircut means an institution must either find more collateral or sell assets. Selling into a falling market pushes prices down, which raises haircuts further, which forces more selling.

That is a run, with exactly the structure of the Diamond-Dybvig model, conducted between institutions rather than at a branch counter, on an asset class rather than a bank. Gary Gorton's phrase (a run on repo) is the clearest description of what actually happened, and it explains why the crisis spread to institutions holding no subprime mortgages at all.

Which brings up the framing question, because it determines what anybody does afterwards.

The popular explanation is greed. Bankers were greedy, and the greed caused the crisis.

Greed is a constant. Bankers in 1995 were not less interested in money than bankers in 2005, and a constant cannot explain a change. What changed was a set of structures: compensation paying out on annual revenue with no clawback when the position later blew up; ratings paid for by issuers; originators with no skin in the game; capital rules that made the instruments cheap to hold; and a regulatory perimeter drawn around institutional form while the risk migrated to entities that were not banks and were doing what banks do.

Every item on that list is a design choice that could be made differently, which is why the structural framing is more useful than the moral one. It also explains why "too big to fail" became the central post-crisis concern.

The phrase is slightly wrong: what matters is interconnection rather than size. An institution at the centre of a web of obligations can be more systemic than a larger one sitting at the edge. And the expectation of rescue is itself a subsidy: creditors lend more cheaply to a firm they believe will be bailed out, which lowers its funding costs and encourages precisely the growth and leverage that made it systemic in the first place.

The regulatory answer has been resolution regimes and bail-in debt: making it possible to impose losses on a failing institution's creditors over a weekend without the payments system stopping. Whether that works is a question that will only be answered by a crisis, which is an uncomfortable way to hold a policy.