Long-Term Capital Management: The Nobel Laureates Who Broke the Global Economy

In September 1998, the chairman of the Federal Reserve Bank of New York, William McDonough, summoned the heads of fourteen of the world's largest banks to a conference room in lower Manhattan and told them they had a weekend to figure out how to rescue a single hedge fund, or the global financial system was going to seize up the following Monday.

The hedge fund was called Long-Term Capital Management. It had been founded four years earlier in Greenwich, Connecticut, by a former Salomon Brothers trader named John Meriwether and a roster of partners that included Myron Scholes and Robert Merton — two economists who, in October 1997, had jointly won the Nobel Prize in Economics for their work on options pricing. The fund had, at its peak, $4.7 billion in capital. Through derivatives and leverage, it controlled positions worth approximately $1.25 trillion. Its trades touched almost every bond market on Earth.

In four months, between May and September 1998, the fund lost $4.6 billion. It came within hours of triggering a cascade of counterparty defaults that would have, in the assessment of every major bank involved in the rescue, taken down at least four of the world's largest financial institutions and possibly the global derivatives market entire.

The story of how a hedge fund staffed by the most credentialed traders in financial history blew up so completely, so fast, is one of the most instructive case studies in the limits of quantitative modelling that has ever been documented.

The premise of the fund

LTCM's strategy, in its original form, was simple to describe and apparently sensible. The fund's traders identified small pricing inefficiencies in bond markets — moments when two essentially identical bonds were trading at slightly different prices — and bet that the prices would eventually converge. The bets were small individually, but the fund used enormous leverage to multiply the returns. A trade that would, on its own capital, have produced a return of 0.1 percent could, with thirty-to-one leverage, produce a return of 3 percent.

The mathematical models the fund used to identify these inefficiencies were, by the standards of 1994 finance, the most sophisticated in the world. Scholes and Merton had spent their academic careers building the theoretical foundation for options pricing. Meriwether's team had spent years at Salomon Brothers refining the models for actual trading use. The combination of academic rigour and practical experience was, by any honest assessment, world-class.

The fund's track record, in its first three years, was extraordinary. It returned approximately 21 percent in 1995, 43 percent in 1996, and 41 percent in 1997. These returns were achieved with what the fund's risk models said was very low volatility. The investors — including most of Wall Street's largest banks, several European central banks, and a long list of wealthy individuals — were ecstatic. The fund's partners were, by 1997, among the wealthiest people in the financial industry.

What the models did not include

The models LTCM used to size its positions were based on the historical relationship between the bonds it was trading. The models assumed, in effect, that the future would behave like the past. Specifically, they assumed that the correlations between different bond markets would remain stable, that price movements would be roughly normally distributed, and that even in stressed markets, the inefficiencies the fund was exploiting would not move very far before correcting.

There were two problems with these assumptions, both of which were known to the fund's partners and both of which the partners chose to manage by other means.

First, the historical data the models were based on covered approximately five years. Five years of bond-market history, in 1994, did not include any episode of severe global financial stress. The data set was clean, well-behaved, and produced models that worked beautifully on the period it had been calibrated on. What it did not capture was what happens to bond markets in genuine crisis.

Second, the leverage the fund used meant that even small adverse price movements would produce enormous absolute losses. A trade that the model said had a 1-in-100 chance of moving against the fund by 5 percent would, if it happened, wipe out a substantial fraction of the fund's capital. The partners knew this. Their response was to argue that the diversification across many trades, combined with the historical pattern of these trades, made the risk acceptable.

This is the critical assumption that broke. The diversification across many trades was real only if the trades were genuinely independent — if, for example, a Russian government bond moving against the fund had no relationship to a Danish mortgage bond moving against the fund. The historical data said they were independent. The historical data was wrong.

The four months that broke the fund

The trigger was geographically unrelated to anything the fund was specifically trading. In August 1998, the Russian government — facing a fiscal crisis it could no longer paper over — defaulted on its rouble-denominated debt and devalued its currency. The default itself was small in dollar terms. But it produced, in the days that followed, a global flight to safety in bond markets.

Investors everywhere stopped wanting to hold anything except the safest possible assets — US Treasury bonds, German bunds, Japanese government bonds. Everything else — corporate bonds, emerging-market bonds, mortgage bonds, the various spread trades LTCM specialised in — sold off simultaneously. The pricing inefficiencies the fund had been profiting from did not converge. They diverged. They diverged together, across markets that the historical correlations had said should be uncorrelated. The fund's positions began losing money in every market simultaneously.

The leverage now worked in reverse. A 5 percent adverse movement in the fund's positions, on thirty-to-one leverage, was a 150 percent movement against the fund's capital. The fund did not have 150 percent of capital to lose. It had 100 percent. The other 50 percent was money that had been borrowed from the fund's lenders, which the fund could not repay.

By the second week of September, LTCM had lost approximately $4 billion of its $4.7 billion in capital. The fund's lenders — primarily the major Wall Street investment banks — were beginning to make margin calls on their positions with LTCM. The fund could not meet the calls. Under normal conditions, a defaulting hedge fund would simply be liquidated and the lenders would seize the collateral. The problem was the size of LTCM's positions.

LTCM, at this point, controlled derivatives positions worth approximately $1.25 trillion, distributed across virtually every major financial institution in the world. If the fund was forced to liquidate its positions in distressed markets, the liquidation itself would push prices further against every counterparty that had similar positions. The cascade of losses would, by the Federal Reserve's analysis, force at least four of the world's largest investment banks into emergency capital raises and possibly into outright failure.

The bailout that was not, technically, a bailout

The Federal Reserve had no legal authority to use public money to rescue a private hedge fund. What it had was convening power. McDonough, the New York Fed chairman, summoned the heads of the fourteen banks most exposed to LTCM and told them they would, collectively, recapitalise the fund — putting in approximately $3.6 billion of new capital, in proportion to their exposure — to allow the fund's positions to be unwound in an orderly way over the following months.

The banks agreed, reluctantly. They had no real choice. The alternative, which several of them had calculated independently, was to lose far more than $3.6 billion if the fund were forced into immediate liquidation. The deal was struck over a weekend. The recapitalisation was announced on Monday, September 28, 1998. The global financial system did not seize up. The fund was wound down, slowly, over the following two years. The positions were eventually closed, mostly at small losses or break-even relative to the rescue valuations. Almost all of the original investors, including the Nobel laureates, lost everything they had invested.

What the case actually demonstrates

LTCM is sometimes told as a story about hubris — Nobel laureates who thought they had figured out the markets and got humbled. The hubris was real. But the deeper lesson is more uncomfortable, because it applies to many situations where the participants are not particularly arrogant.

The lesson is that mathematical models, however sophisticated, are calibrated on the data the modellers have access to. When that data does not include the kind of stress event the model is being asked to handle, the model will produce confident answers that are exactly wrong at the moment they matter most. The fund's models said certain trades had effectively zero probability of all moving against the fund simultaneously. In the historical data the models had been trained on, that was true. In the actual market of August 1998, it happened.

This is not a problem unique to hedge funds, or to bond markets, or even to finance. Any system that uses historical data to project future risk faces the same structural limitation. Insurance companies face it. Climate models face it. Pandemic preparedness models face it. Public-policy planning faces it. The model captures the regularities of the data set; the catastrophes it cannot predict are, almost by definition, the events that fall outside the data set.

LTCM's specific contribution to the literature is to demonstrate, with unusual clarity, what happens when the model's blind spot is met with leverage. The blind spot itself is universal. The leverage is what turned a blind spot into a near-collapse of the global financial system.

The pattern recurs. Every major financial crisis since 1998 — the 2008 mortgage crisis, the 2010 European sovereign-debt crisis, the various crypto-derivative blowups of the 2020s — has involved, at its core, the same structural mistake. Sophisticated participants used historical data to size positions that, when stressed by events the historical data did not contain, produced catastrophic losses. The mathematics gets more sophisticated. The blind spots do not go away. They just get hidden in models that look more impressive.

LTCM is the cleanest example of the pattern because it involved the most credentialed possible group of participants making the mistake in its purest form. If Nobel laureates with the best models money could buy could blow themselves up this completely, the lesson is not that other people should be more humble. It is that any model resting on a finite historical data set is, structurally, a confident answer about a future the model has not actually seen.