LTCM.
Two Nobel laureates, the trade everyone wanted to copy, and a $4.6B four-month unwind. The cleanest evidence for why a model without a combinator is a single-regime bet.
The smartest room in finance, betting on convergence.
Founded 1994 by John Meriwether (ex-Salomon arb desk) and an all-star bench: Myron Scholes and Robert Merton (both Nobel 1997), plus a row of PhDs and former Fed officials. The investor pitch was simple: they had a model nobody else had, and the math said convergence was inevitable.
The strategy: relative-value statistical arbitrage. Find two assets the model said should trade at the same price (off-the-run vs on-the-run Treasuries, Italian vs German bond spreads, Royal Dutch vs Shell). Go long the cheap one, short the expensive one. Hold until they converge. Lever the position up because each trade looked low-variance.
In year one this paid 21 % net of fees. Year two: 43 %. Year three: 41 %. By 1997 the fund had grown so big they returned $2.7B to outside investors so the partners could keep a bigger share. The combinator layer didn't exist. There was no ensemble, no fusion across uncorrelated strategies, no scheduler that could turn the firm off. Just one model, levered.
Six years up. Four months down.
The model wasn't wrong about long-run convergence. It was wrong about what happens when liquidity disappears in every spread at once.
LTCM opens with $1.25B
Largest fund launch on record. Investor base includes Banco di Sicilia, Sumitomo, UBS, and dozens of central banks looking for safe arbitrage exposure.
First full year · +21 % net
Off-the-run / on-the-run Treasury spread trades, Italian-German bond convergence, equity-pair stat-arb. All low-variance, all heavily levered.
21 % net · Sharpe ≈ 4.3Compounding · +43 %, +41 %
Capital base swells to $7B. Banks start replicating the trades (Goldman, Bear, Bankers Trust all running adjacent books). Spreads tighten across the entire convergence menu.
capital ×6 in three yearsReturns capital to outside investors
$2.7B sent back to keep the partner share larger. The fund is now harder to scale: spreads are too narrow to absorb the AUM. Leverage climbs from 25:1 to compensate.
First losses · −16 %
Salomon Smith Barney shuts its US fixed-income arbitrage desk, dumping similar positions. LTCM's convergence trades move against them as a peer unwinds the same book.
Russia defaults on ruble debt
The "flight to quality" everyone modeled separately suddenly hits every spread at once. Off-the-run spreads blow out, EM-DM correlations go to 1, the model's covariance assumption breaks in a single morning.
monthly loss · $1.85B in 12 trading daysCapital down to $400M
From $4.7B at the start of the year. Leverage ratio mechanically hits 250:1 as denominator collapses. Derivatives notional ~$1.25T, larger than the GDP of all but a handful of countries.
−$4.6B cumulative · 90 % of capitalFed-organized rescue · 14 banks · $3.6B
The New York Fed convenes 14 major banks (Goldman, Merrill, JP Morgan, Morgan Stanley, ...) to inject $3.6B in exchange for 90 % equity. Not a bailout, a controlled unwind to prevent a counterparty cascade.
first time the Fed coordinated a hedge-fund rescueLTCM is liquidated
Positions unwound, fund closed. Meriwether returns to launch JWM Associates with the same playbook and the same blind spot. JWM closes after 2008 losses. The combinator-less model fails twice.
The fund was not undone by a single bad signal. It was undone by the absence of a layer above its signals, a layer that could have noticed every spread moving the same direction at once, weighted the model's confidence accordingly, and turned the leverage off.From the StratCraft signal-fusion thesis · "combinators are moats"
What LTCM didn't have.
Read at the combinator layer: LTCM is a one-recipe, no-scheduler architecture. The arb model produced one read on the market and the firm sized into it with leverage that assumed the read was independent across positions. It wasn't.
One model · no combinator.
A single statistical-arb recipe, sized by a covariance matrix calibrated on 5 years of post-1990 data. No layer above it watched for regime shifts or weighted competing reads.
Many recipes · fused & ranked.
Multiple uncorrelated signals (indicator, HMM, ML, factor) fused by a combinator with confidence weighting. A scoreboard ranks across symbols and turns leverage off when consensus disappears.
Three things LTCM taught the next generation.
Correlation isn't a parameter. It's a regime.
LTCM's risk model treated the covariance matrix as a fixed object calibrated on historical data. In a flight-to-quality, every spread it traded correlated to 1 in a single morning. The "5-sigma event" was a regime change the recipe couldn't see.
↪ Read the taxonomyA model that scales linearly isn't a strategy.
LTCM scaled the position size, not the architecture. There was no second recipe, no fusion across families, no scheduler that could dial the firm down. Every additional dollar bought more of the same exposure.
↪ See the architectureThe Nobel laureate isn't the moat.
Two Nobel prizes inside the firm didn't help. The mathematics of Black-Scholes and continuous-time finance describes derivative pricing; it doesn't ship a combinator. The moat lives in the layer that fuses signals, not the layer that generates them.
↪ View the infographicFour firms. One pattern.
Build above the model.
The combinator is what survives 1998.