The team and the capital base
John Meriwether founded LTCM in 1994 after leaving Salomon Brothers. The staffing list reads like a graduate seminar. Myron Scholes, co-author of the Black-Scholes formula, 1997 Nobel laureate. Robert Merton, continuous-time finance, co-laureate that same year. David Mullins Jr., former Vice Chairman of the Federal Reserve Board. Behind them, a cohort of PhDs out of MIT, Harvard, and Chicago, most of whom had sat on Meriwether’s arb desk at Salomon.
Initial fundraising closed at roughly $1.25 billion. Minimum ticket was $10 million, with a three-year lock-up. The fee schedule was 2 and 25 — a 25% performance cut where the industry standard was 20. The lock-up matters for what happened later: the hedge fund structure locked the investor base in, which means the funding crisis that killed LTCM came from prime-broker margin calls and bilateral counterparty exposure, not from investor outflows. Different failure mode than most of the hedge-fund failures of 2008.
Convergence trading, mechanically
LTCM’s core strategy was convergence (or relative-value) arbitrage. Identify two instruments whose cash flows are substitutable and whose prices have temporarily diverged. Go long the cheap leg, short the rich leg, collect the spread as it narrows. The per-trade edge is small, on the order of a few basis points. Leverage turns small edges into double-digit returns.
The canonical example is on-the-run versus off-the-run US Treasuries. A 30-year bond issued last week and a 29.5-year bond issued six months earlier have nearly identical duration and identical credit quality. Their yields should differ by the liquidity premium on the newer issue, typically 3–5 bp. If the gap widens to 8 bp, you short the rich (on-the-run) and buy the cheap (off-the-run). When the spread reverts, the basis points are yours. On a $100 million notional, a 5 bp move is $50,000. On $10 billion, $5 million. The math is linear in size, which is exactly the property that justifies leverage to a desk that believes the spread is mean-reverting on a known half-life.
The mathematics here is uncontroversial. The trade is duration-neutral to first order. The credit leg is identical. What you are being paid for is bearing the liquidity differential and the carrying cost of the short. In calm markets this is close to riskless, in the technical sense that the P&L variance is low and the reversion half-life is short. The qualifier — in calm markets — is doing almost all the work.
Returns, 1994–1997
| Year | LTCM (net of fees) | S&P 500 |
|---|---|---|
| 1994 (partial) | +20% | +1% |
| 1995 | +43% | +38% |
| 1996 | +41% | +23% |
| 1997 | +17% | +33% |
The 1997 number is the signal. As assets grew (peaking near $7 billion before partners returned $2.7 billion to outside investors, which mechanically raised effective leverage on the remaining capital), the obvious convergence trades became crowded and the spreads compressed. The response a risk manager would expect is straightforward: reduce gross exposure, hold net capital constant, accept a lower absolute return. What actually happened was the reverse. Gross leverage went up. The fund expanded into equity volatility, merger arbitrage, and emerging-market sovereign debt — asset classes where the partners’ in-sample data was thin and where the convergence argument depended on assumptions the Treasury book had not required.
This is the point at which the strategy drifted from a statistical-arbitrage book to something closer to a correlated portfolio of short-volatility positions across unrelated markets. The partners understood this. The question is whether they priced the joint tail correctly. They did not.
August–September 1998
The trigger was external. The Asian crisis of 1997 had already raised global risk aversion. Then, on 17 August 1998, Russia defaulted on its domestic ruble-denominated debt and devalued the currency. A sovereign default from a nuclear power was not in anyone’s base case. LTCM’s models, calibrated on post-1980s developed-market data, assigned this outcome a probability that rounded to zero.
What followed was a classical flight to quality. Capital moved into US Treasuries, specifically on-the-run Treasuries, and out of everything else. The asymmetry matters here: LTCM was short on-the-run and long off-the-run. The spread the fund was being paid to close instead widened. Simultaneously:
- The corporate bond – Treasury spread widened from roughly 50 bp to over 200 bp.
- Mortgage-backed securities, held in size, repriced as investors demanded higher liquidity premiums.
- Long-dated S&P implied volatility, which LTCM had sold as a short-vol position, spiked into the high 30s. The fund had been short at σ closer to 19.
In August alone the fund lost $1.8 billion, roughly 45% of capital. By 21 September, cumulative losses were $4.6 billion against $4.7 billion of starting capital, a drawdown of 98%. Because of the 30:1 leverage, that $4.6 billion evaporation was sitting on top of $125 billion of notional positions that still needed to be closed out against counterparties who, by then, understood the fund was insolvent.
The Fed-brokered rescue
On 23 September 1998 the Federal Reserve Bank of New York convened fourteen banks in a conference room at 33 Liberty Street. The case made to them was narrow and empirical: if LTCM is forced to liquidate $125 billion of positions into a market already short of liquidity, the mark-to-market on your own inventory — because you are holding similar positions — will fall by more than the cost of recapitalising LTCM and unwinding it slowly. This was not a moral argument. It was a cost comparison.
The banks agreed to inject $3.6 billion in exchange for 90% of the fund. LTCM’s partners, individually worth nine figures weeks earlier, were effectively wiped out; most had reinvested prior profits into the vehicle. No public money was used, which distinguishes this from the 2008 financial crisis, but the mechanism — a regulator coordinating private balance sheets to prevent a disorderly unwind — set the template.
The positions were unwound over eighteen months. Most of them did converge, eventually. The banks that absorbed the book earned modest profits on it. LTCM was directionally correct and temporally insolvent. That distinction is the entire lesson. Leverage collapses the horizon on which you are permitted to be right.
Why the models failed
The failure was not in the equations. It was in the inputs and in the independence assumptions.
Correlation inputs. The book was constructed as if the Treasury basis trade, the mortgage-spread trade, the equity-vol short, and the emerging-market positions were weakly correlated. Under the historical data available to 1994, that was approximately true: pairwise ρ ranged from roughly 0.1 to 0.3. Under stress, what the covariance matrix actually did was collapse toward rank one. Every position was correlated with every other position through a single latent factor: global risk appetite. A portfolio that looked diversified in the unconditional basis was a single directional bet in the basis that actually mattered. The right way to phrase the failure mode is this: the covariance matrix LTCM ran the book on was not the covariance matrix the market traded on August 17. In a crisis, correlations converge to one, and a 16-asset portfolio becomes a 1-asset portfolio.
Liquidity. Pricing models typically assume you can transact at or near the mid. That assumption breaks when you are the marginal seller in every market you are in. LTCM’s positions were large enough in several markets that the act of exiting moved the price against the remaining exposure. This is the adverse-selection version of market impact: the more urgent your need to close, the worse your fill, the more you need to close. It is a positive-feedback loop the Black-Scholes framework cannot see, because it assumes frictionless hedging.
Tail thickness. The internal risk system used Gaussian-tail assumptions. Empirical bond-spread returns are not Gaussian; their kurtosis κ is materially above 3. A move LTCM’s VaR system flagged as a 6σ event — probability roughly 1 in 500 million per trading day under the normal — occurred inside the fund’s first four years of operation. That is not a once-in-the-age-of-the-universe coincidence. It is evidence that the probability was never 10-9. The real distribution had tails perhaps five orders of magnitude heavier than the model admitted. Mandelbrot had shown this on cotton prices in 1963. The literature predated LTCM by roughly three decades. The failure was not absence of available knowledge. It was a deliberate trade of realism for tractability, made because Gaussian returns let you write closed-form expressions and Pareto-stable returns do not.
The engineering frame
I think about this episode in the language I came up in, which is software. A risk system is a piece of code. Its inputs are price histories, position sizes, and assumed distributions. Its output is a number, the VaR, which determines how much capital sits behind the book. The engineer’s question is not whether the function compiles. The question is whether the test coverage exercises the regions of the input space the function will actually be called on in production.
LTCM’s risk engine was unit-tested on the 1990–1994 sample. It was not tested on a Russian sovereign default with concurrent flight-to-quality and margin-call cascades. That input vector did not exist in the test suite. When it arrived in production, the function returned a value that was numerically correct under its assumptions and operationally meaningless under the regime that was actually unfolding. This is what fuzzing exists for. The 1987 crash, the 1994 Mexican peso crisis, the 1997 Asian crisis — those were available adversarial inputs. Run the covariance matrix through each window, observe what the rank does, size the book so the worst observed rank-collapse leaves you solvent.
I have caught myself making the smaller version of the same mistake. Backtested a pairs strategy in 2021 on five years of clean post-GFC data, deployed at three-quarter sizing, watched the residual-spread distribution behave nothing like the in-sample histogram during a six-week vol regime in 2022. The Sharpe was 0.4 ex-post against 1.6 in-sample. Discontinued. The lesson, miniature version: any backtest that has not been run through the regimes it has not seen is, structurally, a demonstration that the strategy worked on the data you happened to fit. Nothing more.
What the episode teaches, operationally
Leverage is a multiplier on conviction and on error, symmetrically. There is no version of LTCM’s strategy at 30:1 that survives an August 1998, regardless of how correct the partners were about convergence over a one-year horizon. The horizon is a hard constraint, not an assumption. This is the core of honest position sizing: the worst-case path matters more than the expected-value destination.
Diversification defined on peacetime data is not diversification. The correct test is whether positions are independent conditional on stress, not conditional on the last five years of unconditional returns. Run the covariance matrix through the 1987, 1994, and 1998 stress windows separately. If the rank drops by more than two, the diversification is cosmetic.
Time is a dimension of risk. A trade profitable over 12 months that cannot survive a 3-month drawdown is, operationally, a losing trade. Survival is a binary constraint that multiplies everything upstream of it.
Roger Lowenstein’s When Genius Failed (2000) remains the best narrative account. For the quant side, Philippe Jorion’s 2000 post-mortem in European Financial Management is the document worth reading. Jorion walks through what LTCM’s own internal VaR system was reporting in the weeks before the collapse and shows that the numbers were defensible by the system’s own lights and still worthless. His core observation is that VaR computed on a five-year window using Gaussian tails is not a measure of risk; it is a measure of recent volatility, and any model that confuses the two will eventually be embarrassed by a regime that is not in the window. That is the right frame.
The 2008 rhyme
The parallels to 2008 are not coincidental. Investment banks going into that year ran balance-sheet leverage in the 25–35:1 range, the same neighbourhood as LTCM. Their risk models, like LTCM’s, used correlation inputs estimated on the prior decade of data. Their positions, like LTCM’s, were large enough and overlapping enough that a forced unwind by any one of them imposed mark-to-market losses on all of them. The rescue structure — regulator-coordinated balance-sheet support — was the same playbook, scaled up by three orders of magnitude and this time with taxpayer money attached.
The ten-year gap between the events did not produce meaningful leverage regulation on hedge funds or on bank proprietary desks. That is the policy conclusion, to the extent there is one. The quant conclusion is simpler: any risk system whose worst-case scenarios are drawn exclusively from realised history is, by construction, unprepared for the first event of a new kind.
What I draw from it
The received moral of LTCM — “genius failed” — reads to me as off the mark. The partners were not undone by intelligence. They were undone by a specific, identifiable modelling choice, which was to price joint-tail risk using assumptions the academic literature had already shown were wrong by 1963. The failure was not of genius. It was of the engineering discipline that should have forced the question: what does this portfolio look like when the covariance matrix is not the one I fit on? That question has an answer, and the answer is a number, and the number would have told them to run roughly one-third the leverage they actually ran. The right lesson is not that brilliant people make bad trades. It is that a model’s output is only as credible as the assumption you would least want to defend in public. Assumptions, on the record for this article: Gaussian tails do not describe financial returns; correlations are not stationary across regimes; liquidity is a state variable, not a constant. In the tails, everything you diversified into is the same trade.
See also: The Asian Financial Crisis of 1997 · The 2008 Financial Crisis · What Is a Hedge Fund? · What Is Margin Trading? · Position Sizing and Risk Management
