Long-Term Capital Management — the fund that was diversified until it turned out everyone owned the same book

Filed under interpretation in canon/proposals.md — the entry sits inside that file's ## interpretation block. I am filing it as a moment, and the reason is not pedantry.

The proposal is right that the reason this belongs in the canon is a failure mode: "The models were not wrong about the historical distribution. They were wrong because the strategy's own success drew imitators, so a great many books held the same positions and the diversification was notional." That is a reading, and readings are interpretation in this canon — the test model-laundering-2008 applied to itself, and passed.

But the same test, applied honestly here, comes out the other way, for three reasons.

First, the canon already holds the idea. algorithmic-monoculture-2022 exists, it carries Kleinberg and Raghavan's theorem and Bommasani et al.'s metrics, and it is the entry a reading cites when it needs the mechanism by which shared components produce correlated outcomes. What LTCM adds is not another statement of the mechanism. It is that the mechanism ran, in a market, with real money, and put a central bank in a room on a Wednesday afternoon. That is what moment is for, and filing this as a second interpretation of the same idea would duplicate an entry that is already better at the job.

Second, the causal reading here is contested and the entry has to carry the contest. Donald MacKenzie lays out five explanations of the 1998 crisis and argues for the fifth; the fifth is not established the way the ratings finding in model-laundering-2008 is established (there, all ten commissioners of a 6–4-split commission recorded it). An entry that must present five hypotheses, grade them, and say which evidence discriminates between them is an entry about an event whose causes are argued about. expert-systems-collapse-1987 is filed moment on exactly those terms.

Third, the citation occasions the proposal itself names both want an event. "Correlated automated decisions are argued to be a systemic risk" — a reading wants the precedent, not the theory. "Credentials are offered as evidence a model is sound" — a reading wants the GAO's finding about a specific set of credentials on a specific set of loan documents. Both are this happened, on this date.

idea fails: there is no loop and no formalism to implement. limit fails: nothing here is proved and nothing is forbidden. prediction fails: LTCM made no dated public forecast, though Alan Greenspan made several about it a week afterwards and section 3 grades them.

Three corrections to the proposal, made here because a later reading will otherwise carry them. Merton and Scholes were founding principals of the management company, not board members. The rescue is better described as a consortium the Federal Reserve Bank of New York convened and hosted than as one it organised — no Federal Reserve money was committed and Greenspan said so under oath, which matters for the misuse in section 4. And the distinction model-laundering-2008 draws — "LTCM was sophisticated users wrong about the tails" — is very nearly right and is worth sharpening, because it is the single most useful thing in this file: LTCM was not principally wrong about the tails. It was wrong about the independence of its positions, and the reason it was wrong was social rather than statistical. The distributions were about right. The correlation matrix was not, and it was not wrong because the mathematics was bad. It was wrong because other people had read the same papers, hired the same kind of person and put on the same trades.

The proposal's parenthesis — "Not Bridgewater, which is Dalio's firm, systematic, still the largest, and never bailed out" — is correct and I have nothing to add to it beyond confirming that nothing in this file is about Bridgewater.

On descends_from, which is one id and has to be argued for

goodharts-law-1975 is listed, and the descent is conceptual and not historical. Nobody at LTCM was reading Goodhart, and I found no evidence that anyone in the 1998 literature reached for him. The link is that Goodhart's sentence, in its original form, describes what happened with unusual precision:

> "Any observed statistical regularity will tend to collapse once pressure is > placed upon it for control purposes."

LTCM's most important observed statistical regularity was that the correlations between its many unrelated-looking positions were about 0.1 or lower over five years of data. In August and September 1998 they went to roughly 0.7. The regularity collapsed, and it collapsed because enough capital had been placed on it.

And here is the mismatch, which a reading should not paper over. Goodhart says "for control purposes." The pressure on LTCM's correlations was not control; it was exploitation — many firms trying to earn the same spread, not one authority trying to steer it. goodharts-law-1975 itself insists that the difference between Goodhart's sentence and Strathern's popular rendering is load-bearing, and by the same discipline the difference between control and exploitation is load-bearing too. A reader who prefers to treat this as a rhyme rather than a descent is not making a mistake. I have listed it because the entry is genuinely downstream of the shape of the thought, and because the alternative was an empty list that would have hidden a real relation.

Two real ancestors have no id and cannot be listed. The Black–Scholes–Merton option pricing model (Black and Scholes, Journal of Political Economy, 1973; Merton, Bell Journal, 1973) is the actual intellectual parentage — two of its three authors were LTCM principals, and the third, Fischer Black, had died in 1995. It deserves an id and does not have one, and it is a stronger candidate than most of what is on the proposals list. The second is Salomon Brothers' bond arbitrage group, where Meriwether, Rosenfeld, Haghani, Hilibrand, Hawkins and Krasker all worked before LTCM, and which is the institutional rather than intellectual parent. Neither is invented into the header.

One entry that should probably descend from this one, and which I have not touched. algorithmic-monoculture-2022 postdates this by twenty-four years and carries descends_from: [transformer-2017]. Whether ltcm-1998 belongs in that list is a judgment for a session working on that file; a canon job writes exactly one file, and this is it.

---

What it is

The fund

Long-Term Capital Management began trading in February 1994, out of a low-rise office block in Greenwich, Connecticut, with offices in London and Tokyo and its primary registration in the Cayman Islands. Its founder was John Meriwether, who had built and run the bond arbitrage group at Salomon Brothers and was, in MacKenzie's phrase, "widely acknowledged as the most talented bond trader of his generation."

The partnership was small: initially 11 partners and 30 employees; by September 1997, 15 partners and about 150 employees; 16 partners by 1998. Among them were Robert C. Merton and Myron S. Scholes, who on 14 October 1997 were awarded the Sveriges Riksbank Prize in Economic Sciences in Memory of Alfred Nobel "for a new method to determine the value of derivatives" — the option pricing formula they had developed with Fischer Black. Also among them was David W. Mullins Jr., formerly Vice Chairman of the Federal Reserve Board and before that Assistant Secretary of the Treasury.

This was not a fund for rich individuals. Individuals supplied less than 4 percent of LTCM's capital; the rest came from financial institutions. It launched with roughly $1.25 billion — reported at the time as the largest initial hedge-fund raise ever, from about eighty investors at a $10 million minimum — and it required investors to leave their capital in for a minimum of three years. As MacKenzie notes, that lockup "was in part this restriction that made the fund Long-Term Capital Management."

The returns, net of fees, were extraordinary and are worth having exactly right, because the popular account tends to round them upward:

| year | gross | net of fees | |---|---|---| | Feb–Dec 1994 | 28.1% | 19.9% | | 1995 | 59.0% | 42.8% | | 1996 | 61.5% | 40.8% | | 1997 | — | about 17% |

By August 1997 the fund held $126 billion in assets against $6.7 billion of its own capital. At the end of 1997, finding fewer opportunities, it returned about $2.7 billion of capital to investors — reducing its equity by over a third, to a net asset value of $4.67 billion, while leaving the balance sheet roughly where it was. Balance-sheet leverage went from 17-to-1 at the end of 1994 to 28-to-1 at the end of 1997.

What it actually traded, which was not what people think

LTCM was a fixed-income relative-value fund. The GAO sorts its book into four kinds: convergence trades (long an "old" two-year Treasury, short the corresponding "new" one, where the two must converge by a specifiable date), relative-value trades (same idea without a fixed convergence date), conditional convergence trades (arbitrages that hold if a pricing model is right), and directional trades.

Concretely, the thirteen major positions in August 1998 included: equity index volatility (LTCM had sold large quantities of long-dated index options across the major markets, on the view that their implied volatilities — over 20 percent per annum on five-year S&P 500 options — were far above the 10–13 percent the indices had actually delivered); swap spreads in the US, UK, France, Germany and Japan; Royal Dutch/Shell, a pair of stocks entitled to the same income stream in a fixed 60:40 ratio and persistently not priced that way; Italian government bonds; commercial and residential mortgages; merger arbitrage; and Japanese bank preference shares.

By the time of the crisis its swap book alone ran to some 10,000 swaps with a total notional value of $1.25 trillion; the BIS called LTCM "perhaps the world's single most active user of interest rate swaps." As of 31 August 1998 the GAO puts total off-balance-sheet derivatives notional at about $1.4 trillion, of which more than $500 billion was exchange-traded futures and at least $750 billion OTC. LTCM was counterparty to over 20,000 transactions with over 75 counterparties.

None of this was a bet on markets going up. Most of LTCM's positions were, as MacKenzie puts it, "almost completely insulated from overall increases or decreases in the prices of stocks or bonds," and its involvement in high-default markets — junk bonds, Russia, Thailand, Argentina — was limited.

The risk model, and the number in it

The model was value-at-risk, and it was run with a deliberate thumb on the scale toward caution:

• The observed standard deviation of the US dollar swap spread was about 12 basis points a year. The model used 15.
• The observed correlations between the components of the portfolio, on five years of data, were "very small — typically 0.1 or lower." LTCM ran its analyses at correlations of around 0.3.
• The model therefore predicted annual volatility of net asset value of 14.5 percent. Realised volatility was 11 percent. Investors had been told to expect 20 percent.

They also stress-tested: a huge stock-market crash, an Italian sovereign default, a Chinese devaluation, a failure of European monetary union. They bought insurance against an Italian default. They set aside cash for a sudden increase in repo haircuts.

Hold on to the 0.3, because it is the pivot of this entry. LTCM's principals looked at an observed correlation of 0.1, did not believe it, raised it by a factor of three as a safety margin, and were still wrong by a factor of more than two in the event, when correlations across the portfolio "leapt upward from their typical level of 0.1 or less to around 0.7."

A coincidence that has to be named so it is not mistaken for a finding. The number 0.3 also sits at the centre of model-laundering-2008, where it is S&P's assumed correlation between asset-backed securities in the same sector, chosen to reproduce the answers of the procedure it replaced. The two numbers are unrelated. Different models, different assets, different decades, no causal or documentary link, and they were arrived at in opposite spirits — one raised from observation out of caution, one reverse-engineered to preserve a result. The coincidence is worth stating precisely because it invites an inference that the record does not support.

The eight weeks

The chronology is the GAO's, from its Table I.1, with MacKenzie's account filling in the market mechanics.

• 6 July 1998 — Salomon Brothers, by then owned by Travelers, disbanded its US bond arbitrage unit. The unit had been losing money — about $200 million by June — and Sanford Weill closed it. (MacKenzie dates the announcement to 7 July; the GAO chronology says 6 July. The discrepancy is a day and I have not resolved it.)
• 17 July 1998 — Salomon announced it was selling out of certain trades. The liquidation was handed to Salomon's customer desk, which was not accountable for the losses it generated and therefore moved fast. Costas Kaplanis, Salomon's head of global arbitrage, told MacKenzie: "Not only did we lose money as the positions went against us as we were selling them, but all the other funds that also had these consensus trades also started losing money."
• 17 August 1998 — Russia devalued the rouble by about 25 percent and declared a moratorium on debt payments, defaulting on debt denominated in its own currency, which almost nobody had expected. The Dow rose nearly 150 points that day. LTCM's own Russian exposure was small.
• 21 August 1998 — LTCM lost $550 million in a single day as US and UK swap spreads widened and the planned Ciena/Tellabs merger, in which it held a large position, was cancelled.
• 31 August 1998 — Net asset value $2.3 billion. The fund had lost 44 percent in the month of August alone and 52 percent for the year.
• 2 September 1998 — Meriwether faxed investors the August number and an invitation to put in more money: "the opportunity set in these trades at this time is believed to be among the best that LTCM has ever seen." The fax was on the internet within five minutes. Assets LTCM was believed to hold became unsellable except at fire-sale prices; the price of catastrophe ("hurricane") bonds fell 20 percent on 2 September with no change in hurricane risk.
• 18 September 1998 — The flight to quality had driven the yield on the longest-maturity US Treasury bond to a thirty-year low. LTCM told the New York Fed its capital raising had failed and invited officials in.
• 20 September 1998 — A team from the New York Fed and the Treasury, including Assistant Secretary Gary Gensler, met LTCM's partners in Greenwich. In the GAO's words, "the team realized the impact LTCM's positions were having on markets around the world and that the sizes of the positions was much larger than market participants imagined."
• 21–22 September 1998 — Goldman Sachs, Merrill Lynch, J.P. Morgan and UBS concluded the positions could not be lifted out; thirteen more firms joined the discussion. Bear Stearns, LTCM's clearing agent, demanded collateral against settlement exposures.
• 23 September 1998 — An outside offer from Goldman Sachs, Berkshire Hathaway and American International Group to buy the fund was on the table with a 12:30 pm expiry; it lapsed, LTCM saying it could not accept without shareholder approval in the time available. Fourteen firms agreed to recapitalise.
• 28 September 1998 — Closing. The fourteen contributed about $3.6 billion for 90 percent of the fund, through a vehicle called Oversight Partner I LLC, on a 1 percent management fee and a 15 percent incentive fee over a Libor hurdle. The fourteen: Barclays, Bankers Trust, Chase Manhattan, Credit Suisse First Boston, Deutsche Bank, Goldman Sachs, Lehman Brothers, Merrill Lynch, J.P. Morgan, Morgan Stanley Dean Witter, Paribas, Société Générale, Travelers, UBS.
• 30 September 1998 — Net asset value $3.81 billion.
• November 1998 — The stake of the sixteen original partners was worth $30 million, against $1.6 billion at the start of the year.

The Federal Reserve cut the federal funds rate three times that autumn — 29 September, 15 October (an intermeeting cut, which MacKenzie identifies as the moment confidence began to return), and 17 November — from 5.5 percent to 4.75 percent. The portfolio was wound down in an orderly rather than a sudden liquidation, substantially complete by December 1999; the consortium's $3.6 billion was repaid, reportedly with a small profit.

What the official record concluded

Two investigations. The President's Working Group on Financial Markets reported on 28 April 1999; the GAO reported on 29 October 1999.

The PWG's headline is that "the central public policy issue raised by the LTCM episode is how to constrain excessive leverage more effectively." The GAO's is that LTCM got that big "primarily because the banks and securities and futures firms that were its creditors and counterparties failed to enforce their own risk management standards."

Three findings from that record matter more than the headlines.

One: the halo. Among the reasons the GAO gives for why firms did not apply their own standards:

> LTCM benefited from a "halo" effect; that is, creditors and counterparties > appeared to base their credit decisions for LTCM on the credentials of its > principals — among whom were a former vice chairman of the Federal Reserve > Board and two Nobel laureates — rather than on traditional credit analysis.

Greenspan said the same thing under oath on 1 October 1998, eight days after the rescue:

> To be sure, sometimes decisions are based on judgments about the soundness of > borrowers that are accepted from third parties or, possibly in this case, that > are founded on the impressive qualifications of LTCM's principals. In some > cases, such truncated risk appraisals may be accurate, but they are not a > substitute for a rigorous analysis by the lender of the borrower's overall > credit worthiness and risk profile.

Two: the diversification was geographic, not strategic. This is the sentence of the whole official record with the most transfer value, and it is the GAO's:

> LTCM's diversification was geographic rather than based on different > strategies; thus, it had replicated similar bets in markets around the world > rather than having different strategies. This lack of true diversification > resulted in LTCM experiencing losses on positions in numerous markets.

And, in the GAO's summary of the PWG's finding about LTCM's counterparties:

> The firms apparently shared LTCM's view that its risks were widely diversified > because its positions were spread across markets around the globe. However, > LTCM's worldwide losses in August and September showed that although its risks > were spread across global markets, LTCM had replicated similar strategies in > each market. As a result, when its strategies failed, they failed across > markets.

Three: nobody's job was the system. The PWG, quoted by the GAO, on why market discipline could not have prevented this:

> [market participants'] motivation is to protect themselves but not the system > as a whole … No firm … has an incentive to limit its risk-taking in order to > reduce the danger of contagion for other firms.

And one thing the official record did not say. Neither report identifies the imitators. Both diagnose LTCM's own diversification as notional. Neither says that a great many other books held much the same positions, which is a different and larger claim. That claim is MacKenzie's, and the next section is about how well it stands up.

---

The part the proposal was commissioned for: five explanations, and the evidence

MacKenzie sets out five accounts of the 1998 crisis and takes them in turn.

1. Greed and reckless gambling. He finds it "entirely inadequate": the partners believed they were running the fund conservatively, the modest volatility of the returns was evidence for that belief, and "it is hard to find anyone inside or outside LTCM who can be proved to have expressed that view before the crisis." 2. Blind faith in mathematical models. Also rejected: "Models were much less critical to LTCM's trading than commonly thought." Many of the anomalies — Royal Dutch over Shell, the swap-spread trades — could be identified without sophisticated modelling; and it was LTCM's own traders who had built the models and who deliberately inflated the risk parameters. 3. Over-leverage. "Almost tautologically correct" and therefore nearly empty: it explains vulnerability, not the event. MacKenzie's line is the best one in the chapter — "Blaming LTCM's crisis on leverage is similar to attributing a plane crash to the fact that the aircraft was no longer safely in contact with the ground." 4. The flight to quality after the Russian default. Real, and not disputed by anyone. 5. The superportfolio. "LTCM's success led to widespread imitation, and the imitation led to a 'superportfolio' of partially overlapping arbitrage positions. Sales by some holders of the superportfolio moved prices against others, leading to a cascade of self-reinforcing adverse price movements."

How the imitation happened

Not by conspiracy and mostly not by theft. LTCM's basic strategy had to be disclosed to potential investors. Its trades could be inferred: being on one leg of an LTCM trade "would greatly narrow the range of likely other legs," and Eric Rosenfeld told MacKenzie it "[didn't] take a rocket scientist" to work out what LTCM was doing. Some trades were famous before LTCM existed — the Royal Dutch/Shell discrepancy had been in the academic literature since 1990. Richard Roll, running mortgage research at Goldman Sachs in 1985–87, describes the same process a decade earlier and calls it mimicry: "If you saw Meriwether going long … with an option-adjusted spread you thought was five basis points, you knew that his model said it's 100 basis points."

And success advertised itself. Traders were told "LTCM made $2 billion last year. Can't you?" In January 1998 a new fund, Convergence Asset Management, "raised $700 million in a single month purely from disgruntled investors denied a chance to buy into LTCM." Kaplanis, who was running Salomon's global arbitrage book at the time:

> the arbitrage philosophy … had been disseminated, well disseminated by August > '98; it was there in quite a few hedge funds, it was there in quite a few > firms. So Salomon [and LTCM] lost their uniqueness in doing these things.

And:

> What happened by September '98 is that there was a bunch of arb trades that … > became consensus. People knew that the U.K. swap spreads was a good trade, > people knew that U.S. swap spreads was a good trade. No other market participant > would have had the same portfolio as LTCM did … but, collectively, much of > LTCM's portfolio of positions was also being held by others.

The evidence that discriminates

The two explanations that survive — flight to quality and superportfolio — mostly predict the same thing, because both push prices away from what an arbitrageur holds. MacKenzie finds the cases where they disagree.

Case type one: LTCM held positions in only some of a set of otherwise similar spreads. Flight to quality should move them all; the superportfolio should move LTCM's more. Case type two: the minority of positions where LTCM was long the more liquid instrument. Flight to quality predicts the spread rises; the superportfolio predicts it rises more slowly or falls.

Equity index volatility is a type-one case, and the numbers are clean. Average implied volatilities on five-year index options, June–July 1998 against September 1998, from JWM Partners' data: S&P 500 up 7.3 points, FTSE 100 up 9.5, CAC up 7.1, SMI up 9.4, DAX up 9.0 — and the Nikkei 225 up only 4.7. LTCM had sold long-dated options on all the major indices except the Nikkei, and there is no flight-to-quality story in which confidence in the future stability of the Japanese stock market rose in September 1998.

Swap spreads give the same shape. June–July against September 1998: UK up 40 basis points, US up 23 — the two markets where LTCM was short the swap spread — against France up 6 and Japan up 7, where it was long or neutral.

Across all thirteen major positions, MacKenzie's verdict is that "in all the cases for which data are available, the relative price movements of the crisis are consistent with the 'superportfolio' explanation, while in five cases they are inconsistent (and in a further four, possibly inconsistent) with the flight-to-quality explanation."

The strongest single piece of evidence is a non-event. The attacks of 11 September 2001 produced a flight to quality of comparable severity. By then Meriwether's successor fund, JWM Partners, was running a broadly similar portfolio, and so were other arbitrageurs — but with far less capital behind it, possibly a tenth as much in total. There was a flight to quality and no superportfolio, and there was no crisis: JWM's returns in September 2001 were "basically flat," the autumn of 2001 saw no big hedge-fund failures and few major losses, and investors added to their hedge-fund holdings rather than withdrawing as they had in 1998.

That is as close to a controlled comparison as financial history offers: same strategy, same kind of shock, different amount of correlated capital, and only one of the two produced a crisis.

What made the unravelling self-feeding rather than self-limiting

LTCM's assumption had been that widening spreads attract arbitrage capital, which restricts the widening. In August 1998 the reverse happened, and the reasons are mostly institutional rather than mathematical, which is why they transfer:

• Managers hit internal loss thresholds and got out. Kaplanis on a British bank that exited its sterling swap-spread position: "if you hold [it] until its maturity you're probably going to make money. But if there are managers out there that can't stand the daily volatility … then that's when you're in trouble."
• Investors in funds without LTCM's three-year lockup pulled money. "They knew that if they showed big losses a lot of investors would want to get out." The fourth quarter of 1998 saw net hedge-fund withdrawals of about $6 billion.
• Value-at-risk limits forced liquidation mechanically. "A proportion of the investment bank[s] out there … are managed by accountants, not smart people, and the accountants have said 'Well, you've hit your risk limit. Close the position.'" MacKenzie flags that Jorion disputes how large this channel was.
• Counterparties marked against LTCM deliberately, choosing unfavourable prices from a wide plausible range in an illiquid market, because doing so extracted collateral before a possible default. In the index-option book this cost LTCM around $1 billion — nearly half of the September losses that pushed it to the brink. One banker who had bought index options from LTCM: "it was clearly in our interest to mark it at as high a volatility as possible. That's why everybody pushed the volatility against them, which contributed to their demise in the end."

Meriwether's own summary, to MacKenzie: "The nature of the world had changed, and we hadn't recognized it." Scholes's: "Maybe the error of Long-Term was … that of not realizing that the world is becoming more and more global over time."

---

Why a reading would cite it

Four occasions, and a boundary.

First: when a supervisor argues that correlated AI decisions are a systemic risk

This is now a live regulatory position rather than an academic one, and the readings will meet it. The Bank of England's Financial Stability in Focus: Artificial intelligence in the financial system (April 2025) warns that advanced AI-based trading strategies could lead to firms taking increasingly correlated positions and acting in a similar way during a stress, and that common reliance on a limited number of open-source or vendor-supplied models and datasets may produce correlated trading behaviour. Deputy Governor Sarah Breeden has since floated market-wide "kill switches" for agentic trading systems, against a survey figure of roughly half of finance firms already running them. The Financial Stability Board named third-party and pretrained-model concentration as a monitored vulnerability in November 2024 and again in October 2025, and Gary Gensler used the word monoculture about AI in January 2024. (Those last three are established in algorithmic-monoculture-2022, which carries their citations.)

LTCM is the base case for that worry, and it is a base case with numbers. Cite it for four specific things a supervisor's warning usually leaves abstract:

1. What the correlation actually did. Not "went to one" — from about 0.1 to about 0.7, in a portfolio deliberately built out of assets with nothing in common at the level of economic fundamentals. 2. How long it took. Six weeks from the trigger to a central bank convening fourteen firms; five days from the Fed seeing the book to the agreement. 3. What the correlated capital was worth. The 2001 comparison is the number that matters: roughly a tenth as much capital in the same strategies, and the same class of shock produced nothing. 4. That nobody had to intend it. No communication was required. Kaplanis: "even had there not been communication, many traders would still have identified the same opportunities."

The transfer to a monoculture argument about models is not automatic and a reading should not pretend it is. What LTCM establishes is that correlated positions arising from independently-arrived-at similar analysis are sufficient to produce a systemic event. Whether many institutions running the same foundation model produces correlated positions is the open empirical question — and algorithmic-monoculture-2022 is explicit that the one deployment study that exists is about a vendor's own models in hiring, not about a shared frontier model.

Second: when credentials are offered as evidence a model is sound

This is the cleanest occasion and it needs no analogy at all. The GAO's "halo" finding and Greenspan's "impressive qualifications … are not a substitute for a rigorous analysis" are about a specific, documented failure: sophisticated lenders, with their own written standards, extended zero-initial-margin credit and rehypothecation rights against a balance sheet they had not sized, because of who was in the room. Eleven months after a Nobel Prize.

Cite it when a safety case, an evaluation, or a capability claim is being carried by the reputation of the people who produced it rather than by anything a third party could check — and cite it with the second half, which is the part people drop: the credentials were not fake and the models were not stupid. Merton and Scholes had earned the prize; the risk model was conservative by construction. The halo did its damage precisely because there was something real underneath it. A reading that uses this entry to say "experts were wrong, therefore distrust experts" has inverted it.

Third: when a diversification claim is made across deployments

The GAO sentence — diversification that was "geographic rather than based on different strategies" — is a portable test, and this project's ground supplies the cases. Many vendors, one base model. Many agents, one scaffold. Many evaluations, one benchmark suite. Many jurisdictions, one compliance framework derived from the same standard. In each, the count of instances is offered as evidence of independence, and the question the 1998 record says to ask is not how many but what would have to be true for these to fail separately.

This is also where the entry pairs with algorithmic-monoculture-2022's individual-versus-group finding: aggregate diversity can look fine while the thing that matters is correlated underneath it.

Fourth: when a state is asked to intervene in a failure, or is accused of a bailout

The LTCM rescue is the canonical instance of a central bank producing an outcome without spending anything: a room, a phone call to fourteen firms, and no public money. That structure is exactly what would be proposed if a large AI-dependent firm or a model provider failed disruptively, and the arguments on both sides were made in public in October 1998 by the person who did it. Greenspan conceded the moral hazard rather than denying it — "any time that there is public involvement that softens the blow of private-sector losses — even as obliquely as in this episode — the issue of moral hazard arises" — and then argued it was worth it. A reading covering a state intervention in an AI incident can cite this entry for what that trade actually looks like when the person making it writes it down.

Where a reading must not take it

It is not evidence about AI, and it cannot move the needle. It is an event in a market in 1998. It establishes that a mechanism is possible and once was decisive; it establishes no rate, no probability, and nothing whatever about whether any current AI deployment is correlated. This canon's rule that history is not evidence exists for cases exactly this seductive.

It also cannot convert a concentration finding into a monoculture finding. "A few companies own the frontier models" and "many decision-makers are running correlated positions derived from the same source" are different claims with different evidence, and the distance between them is most of the content of this entry.

---

What it got right, and what it got wrong

Not required for moment, and done under the full prediction discipline anyway, because this episode generated an unusual number of dated public claims by people in a position to know, and because most of what a reading will have absorbed about LTCM comes from the ones that were wrong.

Claim 1 — "such techniques would be emulated." Alan Greenspan, 1 October 1998. Right about the mechanism, wrong about the consequence, and the error is the interesting part.

Eight days after the rescue, Greenspan told the House Committee on Banking and Financial Services:

> But it is the nature of the competitive process driving financial innovation > that such techniques would be emulated, making it ever more difficult to find > market anomalies that provided shareholders with a high return. Indeed, the very > efficiencies that LTCM and its competitors brought to the overall financial > system gradually reduced the opportunities for above-normal profits. Indeed, > LTCM acknowledged this when returning $2-3/4 billion of capital to investors at > the end of 1997.

Graded: the mechanism is right, in the right place, at the right time, and it is read as an efficiency story rather than a fragility story. Greenspan had imitation, and he used it to explain diminishing returns — the anomalies get competed away. He did not draw the second consequence, which is that emulation also makes the emulators' books resemble each other, so that a shock which used to hit one desk now hits all of them at once. Everything MacKenzie later established was available to him: he names the emulation, he names LTCM's 1997 capital return as evidence of it, and he stops one inference short.

This is the single most useful grade in the file, because it is the error a reading is most likely to repeat about AI. "Everyone can now do what only the best lab could do last year" is currently read, almost universally, as a story about diffusion and efficiency. It is the same fact as "everyone's systems now behave alike under stress," and the second reading is not more pessimistic than the first — it is the other half of the first.

Claim 2 — "a strategy that was destined to fail." Alan Greenspan, 1 October 1998. Not supported by the record.

> To counter these diminishing opportunities, LTCM apparently reached further for > return over time by employing more leverage and increasing its exposure to risk, > a strategy that was destined to fail.

Graded: wrong on the leverage, and the correction matters because this became the standard account. LTCM's leverage did rise after the December 1997 capital return, but the GAO records that its year-end 1997 ratio "was consistent with its 1995 and 1996 ratios," and MacKenzie puts the early-1998 figure at about 27-to-1 — which was, per the PWG's own report, the average ratio of the five largest investment banks at the end of 1998. The GAO's Table 1 is blunter still:

| institution | balance-sheet leverage | |---|---| | Goldman Sachs Group, L.P. | 34-to-1 | | Merrill Lynch & Co., Inc. | 30-to-1 | | LTCM | 28-to-1 | | Lehman Brothers Holdings, Inc. | 28-to-1 | | Morgan Stanley Dean Witter & Co. | 22-to-1 |

The SEC objected to that comparison in its formal comments, correctly, on the ground that the securities firms' assets were less volatile and the firms were subject to capital requirements; the GAO added language and kept the table. Note also the shape of the objection: our leverage is safe because our assets are different — which is, structurally, the argument LTCM's counterparties made about LTCM's geographic diversification.

Claim 3 — "This agreement was not a government bailout." Alan Greenspan, 1 October 1998. True as stated, and routinely reported as false.

> This agreement was not a government bailout, in that Federal Reserve funds were > neither provided nor ever even suggested. Agreements were not forced upon > unwilling market participants.

Graded: accurate on the facts and incomplete as a defence, and Greenspan said so himself in the next paragraph. No public money was committed; no firm was compelled; the fourteen firms did it because they calculated their claims were worth more under an orderly wind-down than a fire sale, and they were right — they were repaid. The incompleteness is that a central bank convening the meeting is itself an intervention, which Greenspan conceded ("even as obliquely as in this episode — the issue of moral hazard arises") and then argued was worth the cost. A reading that calls this "the LTCM bailout" is repeating a claim the primary record contradicts; a reading that calls it costless is repeating a claim the same primary record contradicts.

Claim 4 — the official diagnosis of 1999. Right, and one step short.

The PWG (28 April 1999) and the GAO (29 October 1999) both concluded that LTCM's diversification was geographic rather than strategic and that its counterparties had shared the mistake.

Graded: correct, and it stops at LTCM's own book. Both reports diagnose one firm's internal lack of independence. Neither says that much of that book was simultaneously held by other people, which is the claim that turns a story about one badly-diversified fund into a story about market structure. That step took four more years and required interviews with the participants, because the composition of the superportfolio is not in any dataset — a limitation section 5 returns to.

Claim 5 — "What is remarkable is not this episode, but the relative absence of such examples over the past five years." Alan Greenspan, 1 October 1998, and the reform agenda that followed it. Both overtaken.

Graded: the base rate was wrong and the reform was not implemented. The PWG made "how to constrain excessive leverage" the central policy question in April 1999. In August 2007 the same shape recurred among quantitative equity funds — Khandani and Lo's What Happened to the Quants in August 2007? attributes it to the forced liquidation of one or more large market-neutral portfolios and the subsequent price impact on "similarly constructed portfolios," a cascade of deleveraging that is the superportfolio mechanism under a different name and in a different asset class. In 2008 the leverage of the very firms in the GAO's table was central to the crisis, and two of them no longer existed by the end of it. John Meriwether's successor fund closed its main portfolio in July 2009 after losing 44 percent between September 2007 and February 2009.

The GAO saw this coming, in a sentence worth reading twice:

> as the President's Working Group noted, market history indicates that even > painful lessons recede from memory with time.

The two grades differ, and the disagreement is about whose account you take

This canon grades a predictor's self-assessment against an independent one when it can. Here the split is unusual and a reading should know about it before leaning on section 2.

The insiders' grade — which is also MacKenzie's — is that the world changed underneath a sound method. Meriwether: "The nature of the world had changed, and we hadn't recognized it." Scholes: the error was "not realizing that the world is becoming more and more global over time." MacKenzie's chapter concludes that LTCM's wide diversification "failed to do so, because of the effects of a global microstructure rooted in one of the most basic of social processes: imitation."

The outsiders' grade is hubris and leverage. Roger Lowenstein's When Genius Failed (2000) and Nicholas Dunbar's Inventing Money (2000) are the accounts most people have absorbed, and both are considerably less forgiving. Philippe Jorion's Risk management lessons from Long-Term Capital Management (European Financial Management 6(3), 2000) argues that LTCM severely underestimated its risk through reliance on short-history data and risk concentration, and that optimising positions against the same covariance matrix used to measure risk biases the measurement — a purely statistical critique that needs no imitators at all.

They differ, and the reason is a methodological problem a reading must carry. MacKenzie's account rests on interviews, and his central interviewees include the partners of the firm whose competence the account partly exonerates. He is transparent about this and his evidence is not only interviews — the price data in his tables is independent of anything the partners told him, and the 2001 comparison is independent of both. But "the world changed under us" is exactly what a defeated manager says, and the fact that a distinguished sociologist found supporting evidence for it does not make it the only reading. The honest position is that the superportfolio hypothesis is the best-evidenced account available, that its strongest evidence is quantitative and does not depend on the interviews, and that it has never been tested against a dataset of who actually held what — because no such dataset exists.

---

Commonly misused as

Not required for moment. Included because this episode is invoked in AI argument constantly and almost always for something the record does not support.

1. "Nobel laureates trusted a model and it blew up, so distrust models."

The most popular use and the least supported. MacKenzie: "Models were much less critical to LTCM's trading than commonly thought … LTCM's traders had often themselves developed the models they used: no one was more aware than they of the models' likely deficiencies." The risk model was adjusted toward caution at every parameter this file could check — 15 basis points against an observed 12, correlations of 0.3 against an observed 0.1, a predicted volatility of 14.5 percent against a realised 11. The failure was not credulity about a model. It was that the quantity the model most depended on was a property of other people's portfolios, and no amount of care with one's own data can observe that.

That is the version worth transferring: an assumption can be conservative relative to all available evidence and still be catastrophically wrong, if the thing it describes is determined by what everybody else is doing.

2. "It was a taxpayer bailout."

Not on the record. See section 3, claim 3. The opposite over-correction — "nothing happened, the market handled it" — is also wrong: a central bank convened it, the Fed cut rates three times in seven weeks, and Greenspan's own testimony argues the intervention was necessary to prevent damage to "the economies of many nations, including our own." Both simplifications are available and this entry supports neither.

3. "LTCM was uniquely, recklessly leveraged."

28-to-1, against Goldman Sachs at 34-to-1 and Merrill Lynch at 30-to-1 in the same GAO table. The leverage explains why LTCM could not survive the price moves; it does not explain the price moves, and it does not distinguish LTCM from the firms that rescued it.

4. "Correlations go to one in a crisis."

They went to about 0.7, and the number is the least interesting part. The point is why: not because the assets turned out to share a fundamental — swap spreads, Royal Dutch/Shell and Italian bond yield bulges genuinely have almost nothing in common — but because the same owners were selling all of them at once. A reading that treats rising correlation as a fact about assets rather than about ownership has taken the wrong lesson, and it is the lesson that transfers to AI: if many systems fail together it may be because the world is hard, or because they are the same system wearing different labels, and those call for different responses.

5. Confused with index concentration.

"AI-related firms are driving a large share of index returns" is a real observation and it is not this. Concentration of market value in a few issuers is a different mechanism from correlated positions across many independent decision-makers. The Bank of England's recent work names both, in the same documents, and they should be cited separately.

6. Confused with Goodhart's law, or used as its illustration without the caveat.

See the header note. goodharts-law-1975 is a regularity collapsing under pressure applied for control purposes; this is one collapsing under pressure applied for exploitation. Related, not identical, and the difference is why LTCM's collapse was invisible in advance: nobody had announced a target, so there was nothing to notice being gamed.

7. Confused with model-laundering-2008.

The proposal's own distinction, corrected: 2008 was a number manufactured to preserve a result and then given institutional standing it could not bear; 1998 was a number arrived at honestly, adjusted conservatively, and falsified by the behaviour of people the modellers could not see. The shared 0.3 is a coincidence and is flagged in section 1 so that nobody builds on it.

8. The one this room has to watch in itself.

The tell bitter-lesson-2019 states and credits to moravecs-paradox-1988: a citation available to both sides of a reading decides nothing. From this material one can obtain, at zero evidentiary cost:

• AImageddon — correlated automated decisions have already produced a systemic event; it took six weeks, the participants could not see it coming, and it required a central bank to stop.
• AItopia — the system absorbed it. No public money, no counterparty failed, the consortium was repaid with a profit, the fund wound down in an orderly fashion, and when the same strategy met a comparable shock in 2001 with less correlated capital behind it, nothing happened at all.

Both are true and both are in the primary record above. Neither is a needle movement, because this is 1998 and the needle measures now. The defence is that this entry is cited to say what would have to be true for a correlated-failure claim to hold — and specifically that it supplies the 2001 comparison, which is the discipline: the same mechanism with less correlated capital behind it produced no event, so the claim is about how much, not about whether.

---

Sources

Verification status, stated once rather than hedged throughout. Three primary documents were read directly as rendered page images: the GAO report, chapter 8 of MacKenzie's book, and Greenspan's testimony. Every quotation attributed to those three is transcribed from that reading and is verified, including all the numbers in section 1's chronology, the leverage table, the correlation figures, the implied-volatility and swap-spread tables, and every Greenspan sentence. Everything else arrived through a search or fetch layer that returns a summarising model's rendering, and is marked as such. bankofengland.co.uk, imf.org, committees.parliament.uk and home.treasury.gov all refused retrieval from this machine (403 or timeout), which limits section 2's first occasion specifically; the underlying PDFs exist and a later session with different network access should read them.

Read directly, as rendered page images.

• **U.S. General Accounting Office, Long-Term Capital Management: Regulators Need to Focus Greater Attention on Systemic Risk, GAO/GGD-00-3, 29 October 1999 (released 19 November 1999), signed by Thomas J. McCool — gao.gov/assets/ggd-00-3.pdf, read as page images covering report pages 1–14, 28–35 and Appendix I (pages 36–45). The source for: "lost almost 90 percent of its capital"; Table 1's leverage comparison; the $1.4 trillion notional and its breakdown; 20,000 transactions and 75+ counterparties; the BIS "single most active user of interest rate swaps" quotation; the "halo" effect passage in full; the "replicated similar strategies in each market" passage; the "geographic rather than … different strategies" passage; the PWG quotation on firms having no incentive to limit risk-taking for the system's sake; the counterparty loss estimates ($300–500m each, $3–5bn for the top 17); the full chronology of Table I.1; the four-way classification of LTCM's trading strategies; the leverage series (17:1 at end-1994 to 28:1 at end-1997, 21:1 at end-1998); the Goldman/Berkshire/AIG offer and its 12:30 pm expiry; the fourteen consortium members by name; Oversight Partner I and its fee terms; the November 1998 partner-stake figure; the subsequent-events paragraph (+22% to end-April 1999, +14.1% by end-June, $1bn and ~$300m returned by 30 June 1999); and the SEC's formal objection to the leverage comparison. Limits:** I read roughly 32 of the report's 55 pages. Appendices II–VI (agency comments other than the CFTC letter, and staff acknowledgements) and report pages 15–27 on regulatory coordination were not read, and nothing above is attributed to them. Note that the report's own footnote 14 misdates the President's Working Group report as "Apr. 28, 1998"; its Scope and Methodology section gives the correct 1999, and 1999 is used here.
• **Donald MacKenzie, An Engine, Not a Camera: How Financial Models Shape Markets (MIT Press, 2006), chapter 8, "Arbitrage," pp. 211–242** — the copy hosted for teaching at personal.lse.ac.uk/ROBERT49/teaching/ph232/pdf/MacKenzie-Ch8.pdf, read in full as page images. This chapter incorporates and supersedes MacKenzie's earlier article "Long-Term Capital Management and the sociology of arbitrage," Economy and Society 32(3) (2003): 349–380, and the related chapter "How a Super-Portfolio Emerges" in Knorr Cetina and Preda (eds), The Sociology of Financial Markets (Oxford, 2004); I did not read either of those, and the 1998 material here is cited to the 2006 chapter. Source for: the February 1994 start and the Greenwich/London/Tokyo/Cayman structure; the partner and employee counts; the August 1997 $126bn/$6.7bn figures; the under-4-percent individual capital; the gross and net return series; the $1.25 trillion swap book; the five-explanation framework and the disposal of the first three; the "plane crash"/"contact with the ground" line; the 27:1 early-1998 figure and the PWG p. 29 five-investment-bank average; the risk-model parameters (12 vs 15 basis points, 0.1 vs 0.3 correlations, 14.5% vs 11% vs 20% volatility); the stress-testing description; Richard Roll on "mimicry"; the Convergence Asset Management $700m; all Kaplanis, Rosenfeld, Meriwether, Scholes, Wenman, Leahy and Shaw interview quotations; the correlations leaping to ~0.7; the 44 percent August and the $550m day of 21 August; the Royal Dutch premium from 8% to over 17%; the hurricane-bond 20 percent fall; the marking-against mechanism and the ~$1bn index-option cost; the McDonough and Greenspan quotations as he gives them; Tables 8.1, 8.2 and 8.3 and the thirteen-position verdict; the 2001 comparison; and the closing "global microstructure" passage. His own sourcing for many figures is André F. Perold's Harvard Business School case study of LTCM (1999), which I did not obtain.
• Testimony of Chairman Alan Greenspan, "Private-sector refinancing of the large hedge fund, Long-Term Capital Management," before the Committee on Banking and Financial Services, U.S. House of Representatives, 1 October 1998 — the copy in the FCIC document archive at fcic-static.law.stanford.edu.pdf), read in full as page images (7 pages), and identical to the Board's own copy at federalreserve.gov/boarddocs/testimony/1998/19981001.htm. Source for every Greenspan quotation above, including the emulation passage, the "destined to fail" judgment, the "not a government bailout" and moral-hazard passages, the "impressive qualifications of LTCM's principals" passage, the "relative absence of such examples" line, the one-tenth equity sliver, and the J.P. Morgan 1907 analogy.

Established through a fetch or search layer, not read end to end.

• **President's Working Group on Financial Markets, Hedge Funds, Leverage, and the Lessons of Long-Term Capital Management, 28 April 1999 — the HTML copy at cftc.gov, read through the fetch layer in two passes; the Treasury PDF at home.treasury.gov/system/files/236/hedgfund.pdf timed out twice. Source for: the $2.7bn return of capital and the reduction "by about 36 percent to $4.8 billion"; the "over $125 billion in assets … a balance-sheet leverage ratio of more than 25-to-1" as of 31 August 1998; the gross notional breakdown; the 31 July 1998 $4.1bn capital figure; the August $1.8bn loss and "over fifty percent" year-to-date; the fourteen firms and $3.6 billion; "The Federal Reserve Bank of New York provided the facilities for these discussions"; the thirty-year low in the long-bond yield on Friday 18 September; and "the central public policy issue raised by the LTCM episode is how to constrain excessive leverage more effectively." The fetch layer twice reported the document as truncated**, and the p. 29 leverage comparison MacKenzie cites was not among what it returned; that figure is therefore carried on MacKenzie's citation, not on my own reading of the PWG.
• Nobel Prize in Economic Sciences 1997, announced 14 October 1997 to Robert C. Merton and Myron S. Scholes "for a new method to determine the value of derivatives" — nobelprize.org, from search summary of the press release; I did not fetch it directly.
• LTCM's launch capital — roughly $1.25 billion from about eighty investors at a $10 million minimum, reported as the largest initial hedge-fund raise to that date. From search summaries of secondary accounts only. The GAO and the PWG do not give a launch figure and MacKenzie does not state one in this chapter, so this is the weakest number in the file and a reading leaning on it should find a primary.
• Federal Reserve rate cuts of 29 September, 15 October and 17 November 1998, taking the federal funds target from 5.5 to 4.75 percent — the 15 October intermeeting action read through the fetch layer from federalreserve.gov ("a reduction in the discount rate by 25 basis points from 5 percent to 4-3/4 percent"; funds rate "expected to fall … to around 5 percent"); the other two from search summaries of the FOMC minutes.
• The wind-down and the consortium's return — MacKenzie states the orderly liquidation "was achieved by December 1999"; the claims that the fund was fully wound up in early 2000 and that the participating banks earned roughly 10 percent are from search summaries of secondary accounts and are not verified against a primary.
• JWM Partners' closure, July 2009, after a 44 percent loss in the Relative Value Opportunity II fund between September 2007 and February 2009 — from search summaries of contemporaneous reporting.
• **Amir E. Khandani and Andrew W. Lo, "What Happened to the Quants in August 2007?: Evidence from Factors and Transactions Data," NBER working paper 14465 (2008), published in Journal of Financial Markets 14(1) (2011) — nber.org/papers/w14465. Abstract and summary only, through search. The unwind hypothesis is quoted as they state it; the paper was not read**, and the identification of August 2007 as the same mechanism in a different asset class is mine, not theirs.
• **Philippe Jorion, "Risk management lessons from Long-Term Capital Management," European Financial Management 6(3) (2000): 277–300 — publication details and the substance of the argument from search summary and from MacKenzie's citation of Jorion (2002) on the value-at-risk channel. The paper was not read**; the author-hosted PDF at merage.uci.edu returned 404. It is used above only to establish that a purely statistical critique of LTCM exists and differs from MacKenzie's, which is a claim about the shape of the literature and not about Jorion's numbers.
• **Bank of England, Financial Stability in Focus: Artificial intelligence in the financial system, April 2025, and the Bank's subsequent work on AI trading agents including Deputy Governor Sarah Breeden's "kill switch" proposal and the April and July 2026 Financial Policy Committee records. bankofengland.co.uk returned HTTP 403 to every direct fetch from this machine, including the report PDF; the characterisations above — correlated positions, common reliance on a limited number of vendor or open-source models, herding, the roughly-half-of-firms adoption figure — are from search summaries of the Bank's own pages and of secondary reporting, and no wording above should be treated as a verbatim quotation of the Bank.** A reading that wants to quote the Bank must get the PDF.
• Financial Stability Board (November 2024, October 2025) and Gary Gensler (January 2024) on AI monoculture and herding — not researched here; the citations are in algorithmic-monoculture-2022, which carries their verification status, and are referenced rather than restated.
• Shuchen Meng and Xupeng Chen, "Artificial Intelligence and Systemic Risk: A Unified Model of Performative Prediction, Algorithmic Herding, and Cognitive Dependency in Financial Markets," arXiv:2604.03272, 23 March 2026 — abstract only, retrieved directly. Noted here rather than used: it models exactly this entry's mechanism for AI adoption, reports empirical work on 13F filings, and does not mention LTCM or 1998, which is itself a small finding about how disconnected the AI-herding literature is from its own best historical case.

Named above and not obtained, listed so a later session knows what is owed: André F. Perold's HBS case study of LTCM (1999), which is MacKenzie's source for most of the position-level detail; Roger Lowenstein, When Genius Failed (2000); Nicholas Dunbar, Inventing Money (2000); William J. McDonough's testimony of 1 October 1998, quoted here only as MacKenzie quotes it; the Counterparty Risk Management Policy Group's Improving Counterparty Risk Management Practices (21 June 1999); and the Basel Committee's two January 1999 papers on banks' interactions with highly leveraged institutions.

In this canon. This file was written after reading model-laundering-2008 and algorithmic-monoculture-2022 in full — the first for the kind test and the proposals' own distinction between the two entries, the second for the mechanism and for its regulatory citations — and goodharts-law-1975 in part, specifically its statement of the 1975 sentence and its section on why Goodhart's wording and Strathern's differ. bitter-lesson-2019 supplies the both-sides tell in section 4 and credits moravecs-paradox-1988 for it. expert-systems-collapse-1987, twitter-acquisition-2022, deep-blue-1997, card-counting-1962 and thorp-shannon-roulette-1961 were checked by header only, to see how moment has been used; the claims made about them here are about what kind of thing they are. LENSES.md was read for the concentration lens and the discipline on recording magnitude rather than consequence.

On the boundary. This file names goodharts-law-1975, algorithmic-monoculture-2022, model-laundering-2008, bitter-lesson-2019, moravecs-paradox-1988 and expert-systems-collapse-1987 as entries that exist in canon/. It names Black–Scholes–Merton (1973) and Salomon Brothers' bond arbitrage group as real ancestors with no id, and invents no id for either. It notes that algorithmic-monoculture-2022 might reasonably list ltcm-1998 in its own descends_from and does not touch that file. It refers to LENSES.md solely to establish a citation occasion. It deposits nothing in history/, writes no digest, touches no other entry, and places no needle, no score and no landmark.