_BACKGROUND

The Smartest Guys in the Room

In this edition, we are doing one more legacy finance story as we come off of the Labor Day holiday here in the states. Honestly, I was expecting the Anthropic S-1 to drop by the time this issue came out, but no dice yet so we are taking it back to Greenwich, CT.

In 1994, John Meriwether, the former head of bond arbitrage at Salomon Brothers, founded a hedge fund in Greenwich, Connecticut staffed by two Nobel laureates, a former Federal Reserve vice chairman, and the sharpest quantitative minds Wall Street had ever produced.

The fund was - ironically - called Long-Term Capital Management.

For four years, it was the most successful hedge fund in history. Annualized returns of 21% in year one, 43% in year two, 41% in year three. (Net of fees)

The minimum investment was $10 million and there was a three-year lockup. People lined up down the harbor to get in.

In August and September of 1998, the fund lost $4.6 billion in less than five months.

Its equity fell from $4.7 billion to $600 million. Its notional exposure at peak was $1.25 trillion, roughly 5% of global GDP. The Federal Reserve Bank of New York convened an emergency meeting of fourteen Wall Street institutions on a Sunday afternoon and told them they needed to put in $3.6 billion by the end of the day or global credit markets would seize up on Monday morning. They put up the money.

So what went wrong? The trade, the timing, or the leverage? Keep reading.

_{PRESENTED BY TOLTIQ}

Uber Burned Through Its Full-Year AI Coding Budget by April

Uber exhausted its full-year 2026 AI coding-tool budget by April, with its COO saying it is "very hard to draw a line" from AI usage to consumer-feature output. ServiceNow drew down its annual AI coding budget early over the same stretch. And, this is becoming a common theme.

79% of finance leaders reported AI cost overruns in the past twelve months. The major vendors, like OpenAI and Anthropic, moved enterprise AI to metered billing this spring. The default enterprise pricing model is now variable with usage, not headcount.

Optimal model architecture has become vital for usage scalability across active deals, and ToltIQ is built for it. Their research team continually tests frontier models to optimize the platform, but more importantly, their system is designed to retrieve information efficiently to help PE firms (and their PortCos) avoid unanticipated costs.

Some of the most respected firms use ToltIQ for their diligence. Will yours?

Schedule a demo with them today.

_{THE PEOPLE}

The Top Minds on Wall Street

Listen, the book is called "When Genius Failed" for a reason. The people behind LTCM were the smartest guys in any room, at least if you're counting book smarts. The partner roster was an all-star list, and sometimes too many geniuses in one room doesn't play out the way you'd expect.

John Meriwether had built Salomon Brothers' bond arbitrage desk into the most profitable trading operation on Wall Street in the 1980s. He left under a cloud in 1991 after a Treasury auction scandal involving one of his traders. LTCM was his comeback.

Myron Scholes and Robert Merton shared the 1997 Nobel Prize in Economics for developing the Black-Scholes options pricing model, the foundational framework for pricing derivatives that the entire industry still runs on. They were not figureheads. They were active partners who contributed directly to the fund's trading strategies.

David Mullins was the former vice chairman of the Federal Reserve. He had the relationships, the regulatory knowledge, and the institutional credibility that let LTCM borrow from counterparties on terms no other fund could access.

The reputation of the partners was a trading asset in itself. LTCM could borrow at near-sovereign terms because no bank believed this particular group of people could be wrong enough to default. That belief was the foundation of the leverage. When it broke, everything broke with it.

_{THE TRADE}

When it Rains, it Pours

LTCM's core strategy was convergence arbitrage. The basic idea is straightforward: if two similar instruments are mispriced relative to each other, buy the cheap one, short the expensive one, and wait for the spread to close. The trade makes money when the world returns to normal. It loses money when the world moves further away from normal.

The specific trades took many forms. The one that defined LTCM was the on-the-run versus off-the-run Treasury spread. A freshly issued 30-year Treasury bond ("on the run") trades at a small premium to a Treasury issued six months earlier with the same maturity ("off the run") because the newer bond is more liquid.

The premium is irrational. Both bonds are backed by the same government and mature to the same value. LTCM bought the off-the-run bond, shorted the on-the-run, and waited for the liquidity premium to compress.

The trade worked consistently. The problem was that the spread was small, just a few basis points, so the returns were also small unless you applied enormous leverage…yes, enormous.

LTCM's balance sheet leverage ran at approximately 25:1 on disclosed positions and significantly higher on a risk-adjusted basis. For every dollar of equity, the fund controlled roughly $25 of assets. On notional derivatives exposure the ratio was north of 250:1.

By 1997 the fund had grown to $7 billion in equity. The partners concluded the fund was too large relative to the opportunities available and returned $2.7 billion to outside investors. This allowed them to keep the capital they had invested themselves, reducing the outside equity base while maintaining the full portfolio size.

The leverage ratio rose from approximately 25:1 to 28:1 as a result.

They made the fund riskier by giving investors their money back. The portfolio did not change, but the cushion did.

_{THE RUSSIAN DEFAULT}

The Straw That Broke the Camel’s Back

On August 17, 1998, the Russian government defaulted on its domestic ruble-denominated debt and devalued the ruble.

Russia itself was a small part of the portfolio and in isolation would not have taken LTCM down. The fallout reached Chernobyl-levels quickly, melting down the LTCM portfolio in a manner of months.

When Russia defaulted, global investors panicked and fled to safety simultaneously. Every player in every market made the same decision at the same time: sell risky, illiquid, complex positions and buy safe, liquid, simple ones. The flight to quality was not a gentle drift, it was a stampede.

Every spread LTCM had bet would converge instead diverged, dramatically and simultaneously. 

The off-the-run Treasury spread blew out instead of compressing. European swap spreads blew out. Mortgage-backed security spreads blew out. Emerging market spreads blew out. Equity volatility positions that should have been uncorrelated with credit moved in lockstep.

LTCM's risk models were built on historical correlations between markets. Those correlations assumed that different asset classes moved somewhat independently. In a crisis, they do not. In a crisis, everything correlates to one.

The diversification that the model showed as risk reduction vanished precisely when it was needed most.

The fund lost $1.9 billion in the last two weeks of August alone. By late August its equity was under $2.5 billion. By September 22, it was $600 million against a portfolio that had barely shrunk. The leverage ratio was now in excess of 100:1.

_{THE BAILOUT}

A Sunday Morning at the Fed

On September 23, 1998, William McDonough, the President of the Federal Reserve Bank of New York, convened a meeting at 33 Liberty Street in lower Manhattan.

In the room were the senior executives of fourteen Wall Street institutions: Goldman Sachs, Merrill Lynch, JPMorgan, Morgan Stanley, Lehman Brothers (RIP), Bear Stearns (RIP), Deutsche Bank (a bulge bracket at the time), Société Générale, and others.

McDonough's message was not complicated. LTCM had approximately $1.25 trillion in notional derivatives exposure across virtually every major financial market in the world.

If LTCM defaulted, its counterparties would be forced to immediately close out those positions. Every major bank in the room was a counterparty. 

The resulting fire sale would hit markets already in crisis. Spreads that had blown out in August would blow out further. Liquidity would disappear. The cascade could not be modeled with confidence.

The Fed was not going to use taxpayer money. It was explicit about that, but McDonough made clear that the alternative (doing nothing) was not obviously better for the institutions in the room.

Fourteen firms put in ~$3.6 billion for a 90% stake in the fund. Bear Stearns, LTCM's clearing broker, refused to contribute a dollar, a decision the street remembered ten years later when Bear needed help of its own.

The existing partners kept 10%. Meriwether kept his job managing the wind-down. The positions were unwound over the following months at prices that recovered most of the capital.

That same morning, Goldman Sachs, AIG and Warren Buffett had offered to buy the partners out for $250 million, essentially nothing, inject ~$3.75 billion of new capital, and take over the fund entirely. The offer carried a fuse of about an hour. Meriwether questioned whether he had the authority to accept it. By the end of the day he had a worse outcome and no choice.

That $250 million offer would have been the best deal of John Meriwether's life.

_{THE POST MORTEM}

What Went Wrong

The models assumed normal correlations. LTCM's framework was built on the assumption that different markets move with historical degrees of independence. In August 1998, every market moved in the same direction simultaneously. Correlation went to one. The diversification the model showed as safety disappeared at exactly the moment it was needed. LTCM's models had assigned a 1-in-6.4-trillion-year probability to losing 40% in a month. They lost 44%.

The leverage left no room for being early. The trades were correct. Spreads did eventually converge. The fund ran out of capital before they did. At 28:1, a 4% adverse move eliminates all equity. Being right on a six-month delay is indistinguishable from being wrong.

They returned $2.7 billion of outside capital at the worst possible moment. They kept the positions. They reduced the cushion. The effect was to concentrate risk in the partners' own money at precisely the moment that risk was about to be tested. In late 1997, with the Asian crisis building and early credit stress visible, LTCM's smartest people in the room concluded the fund was too large and gave investors their money back. It was the most expensive act of modesty in financial history.

Counterparty risk became a real concept. Before LTCM, banks did not think systematically about simultaneous default across hundreds of derivatives positions with a single counterparty. After LTCM, they did. For a while.

The Fed set a precedent it could not unset. No public funds were used in 1998, but the principle that the Fed would intervene to prevent the disorderly collapse of a systemically important private institution was established explicitly. It was cited during Bear Stearns, AIG, and the 2008 banking system interventions. The LTCM Sunday was the rehearsal; 2008 was the show.

Nobody changed their models. The lesson was written up in academic papers, Congressional testimony, and the President's Working Group on Financial Markets report in 1999. Ten years later, the same failure mode at ten times the scale produced the 2008 crisis. The models continued to assume diversification would protect institutions in exactly the conditions under which diversification stops working.

RJR Nabisco failed because leverage amplified a macro outcome nobody fully modeled. AOL-Time Warner failed because a bubble currency bought real assets at the peak. LTCM failed because the model was excellent and the market was not the model. The trades were right, but the timing went wrong and the leverage left no room for the gap between those two things.

The LTCM partners were not stupid. Obviously, they were the smartest people in the room. 

That was part of the problem. When you are certain enough in your model to borrow 28 dollars for every dollar of equity, you have stopped asking what happens when the model is wrong. 

The market is not interested in how good your model is. It will take your equity while you wait to be proven right.

If you want to learn more, check out When Genius Failed by Roger Lowenstein. If you work in credit, maybe read it twice.