Case Study 1 — The 2008 Mortgage Crisis as a Story About Models and Their Assumptions

In the summer of 2007, the largest financial crisis since 1929 began to unfold. By the fall of 2008, the U.S. financial system was on the brink of collapse. Lehman Brothers — a 158-year-old investment bank — filed for bankruptcy on September 15. The Federal Reserve and the Treasury improvised emergency interventions on a scale not seen in living memory. The S&P 500 lost roughly 40% of its value over the next six months. Unemployment, which had been 4.4% in May 2007, eventually peaked at 10.0% in October 2009. Millions of Americans lost their jobs, their homes, or both.

How did this happen? There are many ways to tell the story — the Wall Street greed story, the regulatory failure story, the consumer-debt story, the global-imbalances story. Each is partially true. This case study tells one specific version: the story about models and their assumptions. It is not the whole story. But it is a true story, and it illustrates almost everything Chapter 2 has been trying to teach you about how economic models work and how they fail.

The setting

Mortgages are loans for buying houses. For most of the 20th century, mortgages worked roughly like this: a bank loaned you money to buy a house, you paid the bank back over 30 years, and the bank held the loan on its books and bore the risk of you failing to pay. If too many borrowers defaulted, the bank lost money — so the bank was careful about whom it lent to.

Starting in the 1990s and accelerating in the 2000s, this model was largely replaced. Banks began to securitize mortgages — bundling thousands of individual mortgages together into a single financial instrument called a mortgage-backed security (MBS), which they could then sell to investors. The bank originated the mortgage and collected a fee, but the mortgage itself ended up on the books of pension funds, hedge funds, foreign banks, and other investors. The risk had been spread out.

This was a clever financial innovation. By spreading risk across many investors, it should have made the mortgage market more efficient and more resilient. And for many years, it appeared to work.

The pricing of mortgage-backed securities required a model. The model needed to estimate the probability that the underlying mortgages would be repaid. To do this, the modelers used historical data on mortgage defaults from the 1980s and 1990s and built a statistical model to predict future defaults under various economic conditions.

The model had assumptions. The most important one — the one that mattered most for the eventual crisis — was about correlation.

The fatal assumption

The assumption was: house prices in different regions of the United States are not strongly correlated. Specifically, the model assumed that even if one region experienced a housing downturn, others would not — because the U.S. is a large, diverse economy with regional housing markets that respond to different local conditions.

Why did the modelers make this assumption? Because the historical data — going back several decades — supported it. In the 1980s, when oil prices crashed, the housing market in Houston suffered while other parts of the country were fine. In the early 1990s, when the New England technology economy went into recession, housing in Boston suffered while housing in California was fine. There was no precedent in the historical data for all regions of the country experiencing a housing downturn at the same time. The assumption of regional independence was not a wild guess. It was the historical record.

If the assumption was right, mortgage-backed securities were extremely safe. Even if individual mortgages defaulted at high rates in one region, the diversification across many regions meant that the overall pool would still receive most of its expected payments. The triple-A credit ratings the securities received from the rating agencies (Moody's, S&P, Fitch) reflected this assumption: the securities were rated as safe as government bonds.

If the assumption was wrong — if all regions could experience downturns simultaneously — the securities were much more dangerous than they looked. A nationwide housing downturn would cause defaults across all regions at once, the diversification benefit would disappear, and the security would lose much more value than the model predicted.

In 2007, the assumption was wrong.

What happened

The U.S. housing market peaked in early 2006. Prices began to decline gently across the country. By 2007, the decline had accelerated. By 2008, almost every major metropolitan area in the United States was experiencing falling home prices simultaneously. The Case-Shiller national home price index, which tracks home prices across 20 major metropolitan areas, fell by roughly 30% from its 2006 peak to its 2012 trough. Every one of the 20 metro areas saw price declines. The assumption of regional independence — built into the pricing models for mortgage-backed securities — was violated.

When the assumption was violated, the consequences cascaded. Defaults rose across the country at once. Mortgage-backed securities lost value much faster than the models predicted. Investors holding those securities had to sell other assets to meet redemptions and capital requirements, which pushed down prices for other assets, which forced more selling, which... the dynamic is called a fire sale, and it is what financial crises are made of.

Some institutions were holding such concentrated positions in mortgage-backed securities that they became insolvent when those securities lost value. Bear Stearns. Lehman Brothers. AIG (which had sold credit default swaps insuring those securities and now owed payouts it couldn't make). The Federal Reserve and the Treasury intervened to prevent a complete collapse. The crisis spread from financial institutions to the real economy as credit froze, businesses cut back, unemployment rose, and the deepest recession since the 1930s began.

The lessons (for this chapter)

This story is interesting for many reasons. For Chapter 2's purposes, the lesson is about the role of model assumptions.

1. Every model has assumptions, and assumptions are where models can fail. The mortgage-backed securities model was a real model. It used real data. It made real predictions. It also made an assumption (about regional correlation in house prices) that turned out to be wrong, and the failure of the assumption was the failure of the model. The model was not obviously wrong before 2007. It became wrong only when reality moved in a way that the model had assumed reality didn't move.

2. Historical data is not always a guide to the future. The assumption of regional independence had survived decades of historical testing. It still failed. The reason is that the historical period the modelers were drawing from — roughly 1950 to 2005 — did not include any period in which the U.S. experienced a nationwide housing bubble. The 2006 housing market was different from any housing market in the historical data, and so the historical data was not a reliable guide. This is a deep problem with statistical models in general: they are reliable when the future looks like the past, and unreliable when it doesn't. The hard part is knowing which case you're in.

3. The same model can look brilliant for years and then catastrophically fail in one year. From the mid-1990s to early 2007, the mortgage-backed securities model looked like a triumph of financial engineering. It correctly priced an enormous number of securities; it survived the dot-com crash; it survived 9/11 and the recession of 2001. The track record was so good that institutions used the model with confidence and built leverage on top of its predictions. When the model failed in 2007–2008, the failure was not gradual; it was sudden and severe. This is the failure mode of models that work most of the time but break catastrophically in tail events.

4. The right response to a failed model is not to abandon all models — it's to ask what assumption broke and whether other models share the same assumption. After 2008, it was tempting (especially in popular commentary) to conclude that the entire enterprise of mathematical financial modeling had been discredited. The honest professional response was different: which models had the regional-independence assumption and what should replace it? Which institutions held positions sized for a model that the data could not support? What stress-testing procedures should regulators require to catch this kind of mistake earlier? The reformulated models that came out of the post-crisis work have their own assumptions, which will themselves someday fail. That's not a fatal indictment; it's how the field works. The discipline learns; the discipline updates; the next failure will be different.

5. The same story applies to economic models in general. Macroeconomic models predicted the early-2000s economy reasonably well — and then largely failed to predict 2007–2008, just like the financial models did. The reasons are similar: the models had assumptions about how the financial system worked (the financial sector was largely absent from the standard macro models of that era), and when the financial sector turned out to matter enormously, the models were caught flat-footed. Macroeconomics has spent the years since the crisis adding the financial sector back into its models, with mixed results. We will see this in Chapter 26 (where we look at the crisis from a money-and-banking perspective), Chapter 30 (where we look at it as a business cycle event), and Chapter 39 (where we ask what economists agree and disagree about in the wake of it).

Two things this case study is not saying

It is not saying that the modelers were stupid or evil. Most of them were doing the best they could with the information available. The assumption that broke was a defensible choice given the historical record. The failure was a failure of model construction and a failure of human judgment about how much to trust the model when the world started to look unusual.

It is not saying that we should never trust models. Modern life would be impossible without models. Engineering models are why bridges don't collapse. Medical models are why drugs are tested before approval. Economic models — used carefully, with attention to their assumptions and their failure modes — are how policymakers think about complicated problems. The lesson is to use models with humility, not to abandon them.

Discussion questions

1. The mortgage-backed securities model worked for years and then failed catastrophically. How should an institution that relies on a model decide when to stop trusting it?
2. Some critics argue that the assumption of regional independence wasn't really a failure of the model — it was a failure of the modelers to take seriously the possibility that the historical record was not representative of the future. What's the difference between those two framings? Does the difference matter?
3. The chapter said that "every model has assumptions, and every assumption is a place where reality could prove the model wrong." How can a careful user of a model know which assumptions are most likely to break? Are there warning signs to watch for?
4. The 2008 crisis prompted a major rethinking of how economists model the financial sector. Search for what some of those changes have been (look up "financial frictions in DSGE models" or "macroprudential policy" for starting points). Do the new models seem better positioned to predict the next crisis, or are they just adapted to predict the last crisis?
5. In your own field of interest (whatever it is — sports, biology, public policy, software, etc.), can you think of a model that has a similar profile: works most of the time, embeds an assumption that could break catastrophically, and could be the next "regional independence" failure?