Built to Survive Being Wrong
The Most Dangerous Sentence in Finance
I have sat in hundreds of meetings over twenty-five years where someone said the three most dangerous words in finance: “the model says.”
In the summer of 1998, Long-Term Capital Management had assembled an extraordinary team: two Nobel laureates, a former vice chairman of the Federal Reserve, and some of the most sophisticated quantitative practitioners on Wall Street. Its returns had been exceptional. Its models were formidable.
By September, the fund had lost approximately $4.6 billion in less than four months.
The headline is familiar. The mechanism is worth revisiting because it is the mechanism that matters now.
What Actually Broke
LTCM’s strategies depended on historically observed relationships remaining sufficiently stable for convergence trades to work. When the Russian default triggered a global flight to liquidity, correlations shifted, liquidity evaporated, and positions that appeared diversified began moving against the fund at the same time.
The deeper failure was not mathematical. It was classificatory. LTCM applied tools designed for measurable risk to a situation governed by uncertainty.
The distinction matters because it is not academic. Frank Knight drew it in 1921. Risk describes situations where you can identify the possible outcomes, assign meaningful probabilities from historical data, and rely on the underlying process to remain stable. Uncertainty describes everything else: situations where the outcomes cannot be enumerated, probabilities cannot be meaningfully assigned, or the situation is genuinely novel.
LTCM applied risk tools to uncertainty. So did the CDO models in 2008. So did every model that assigned near-zero probability to a national housing collapse based on the assumption that regional housing markets were independent.
You already know this. Here is the part that should concern you about the current cycle.
Where We Are Making the Same Mistake
In the broadly syndicated leveraged-loan market, the shift from maintenance covenants to covenant-lite structures removed an important pacing mechanism from the credit cycle. More than 90 percent of that market is now covenant-lite, according to LCD data. Private credit remains more heterogeneous: many direct-lending agreements still include maintenance covenants, but lender protections have weakened in parts of the market, and larger deals increasingly resemble syndicated documentation.
In a maintenance covenant world, a breach was an early warning system. A company’s financial metrics were tested regularly — typically quarterly — against agreed thresholds, and a breach forced lenders and borrowers to the table while there was still cash, still enterprise value, still optionality. The conversation was painful, but it was functional. In today’s incurrence-heavy market, the covenant test does not trigger on the company’s condition — it triggers only when the company tries to take a specific action: issuing more debt, making an acquisition, paying a dividend. A company can burn through its cash reserves, lose 30 percent of its EBITDA, and remain technically in compliance. Deterioration is observable. It is not actionable.
In a leveraged portfolio company, you can see the deterioration before the documents give you a clean right to intervene. But absent a covenant breach, liquidity trigger, or another contractual right, lenders may have limited ability to force an early intervention.
By the time the contractual tripwire fires, liquidity may be materially depleted and the remaining options may be narrower, more expensive, and more adversarial. The optionality that existed months ago, when the deterioration was first visible, has evaporated. If your process relies on covenant breach as a trigger, your process is structurally late.
This changes what happens when the “actionable” moment finally arrives. The resolution is no longer a clean restructuring negotiation. It is a liability management exercise — a term for the increasingly complex playbook that has replaced traditional restructuring in covenant-lite markets. In an LME, lenders must coordinate across a fragmented creditor group with often competing interests. Sponsors may inject additional capital, negotiate amendments, or support transactions intended to extend liquidity. In more aggressive cases, borrowers may use liability-management techniques such as uptiering, dropdown transactions, or other priming structures that alter recoveries and creditor priority. These are no longer isolated edge cases. They are increasingly prevalent tactics in stressed credits operating under covenant-lite structures, and they can leave unprepared lenders with meaningfully worse recovery outcomes than the headline credit metrics would suggest.
Simultaneously, the extend-and-pretend dynamic in private credit means that a meaningful portion of the “refinancing” between now and 2028 will not appear in traditional metrics at all. PIK elections, maturity extensions, and covenant amendments may avoid a payment default while still signaling economic stress. They may not be visible in headline default statistics, and private-market pricing does not always reveal the deterioration immediately. But the impairment can still be real, deferred, and compounding. Low headline default rates can understate the degree of economic stress if amendments, PIK income, and maturity extensions are not analyzed separately.
Add to this the CLO market dynamics. A CLO, or collateralized loan obligation, is a structured vehicle that pools leveraged loans and issues tranched securities against them. During the reinvestment period, a CLO manager can recycle principal repayments into new loan purchases within the transaction’s criteria. Once that period ends, reinvestment flexibility becomes more limited and principal collections increasingly support liability amortization. The borrower-level issue is not a single market-wide cliff. It is the composition of the lender base in each syndicate. A loan held disproportionately by CLOs with limited remaining reinvestment flexibility may face a weaker marginal bid when refinancing becomes necessary.
As an internal screening heuristic, portfolios may choose to flag borrowers where a material share of CLO-held exposure sits in post-reinvestment vehicles. The appropriate threshold should be calibrated against the syndicate structure and the availability of replacement demand. This dynamic does not appear on most standard risk dashboards, but it can be material for borrower-level refinancing outcomes.
These are not tail risks. They are structural features of the current market that operate below the surface of the standard metrics most credit committees monitor.
The Behavioral Layer You Cannot Firewall
Structural protections are weaker, impairment is harder to observe, and the lender base is more complex. But the problem is not only mechanical. Even when the data is visible, the people responsible for acting on it face incentives to delay.
If the structural blind spots were the whole story, you could fix them with better data. You cannot, because the operators are also compromised, and the compromises look rational from every individual seat at the table.
Many delayed decisions in credit are not simply failures of judgment. They are rational responses to misaligned incentives. GPs may face incentives to preserve IRR optics, which can mean extending hold periods and avoiding write-downs that mark investments to zero. Lenders optimize for recovery, which may mean forcing a sale before things deteriorate further, but only if they can coordinate a fragmented creditor group. Management will often ask for one more quarter, and they face compensation risk if they resist. LP evaluation frameworks can unintentionally penalize early recognition of failure, even when early action preserves more capital than delay. The result looks like cognitive bias. It behaves like coordination failure.
The disposition effect compounds this. In Odean’s brokerage-account data, a stock showing a gain was more than 50 percent more likely to be sold than one showing a loss. In credit, where upside is capped at par plus coupon and downside is recovery value, the cost of holding losers is asymmetric in a way that equity investors never face. Recovery rates in the current covenant-lite environment may be materially lower than historical averages because the deterioration ran longer before anyone had standing to intervene. Holding losers in credit is not the same as holding losers in equity. The math is worse.
Framing effects are equally corrosive. “We are extending the maturity and adjusting the coupon” frames the same economic event as “the borrower cannot refinance at market rates and we are accepting below-market terms to avoid recognizing a loss.” Same transaction. Different frame. Different committee vote. If your investment committee cannot describe the same transaction in both frames and check whether the decision changes, the process has a vulnerability.
The overconfidence calibration problem is measurable. In my experience, credit teams often discover that their 80 percent confidence calls resolve successfully closer to 60 to 65 percent of the time. The exact gap will vary by team. The operating requirement is to measure it. Track forecast confidence against realized outcomes by decision type, sector, and vintage. That gap, when quantified, should inform position sizing and risk appetite directly.
What Architecture Survives This
This post is about what to build. The harder question — why organizations that have built the right architecture still fail to execute it when the moment arrives — is what Part 3 addresses. But you cannot execute what you have not built, so the sequence matters.
If the structural metrics are lagging and the operator is biased, the only thing that works is pre-built architecture. Systems designed during calm that execute during chaos.
Nassim Taleb’s antifragility concept is the right frame, but the implementation is what matters. A portfolio that holds concentrated, leveraged positions with incurrence-only covenants and CLO-dependent demand is fragile in a way that the standard risk metrics will not show until it is too late. A portfolio that has pre-mapped its covenant exposure, stress-tested its CLO demand assumptions, and pre-committed its exit triggers is at least robust.
Benjamin Graham’s margin of safety is the oldest and most violated principle in credit. If your analysis says a credit is money-good at 4x leverage, your underwrite should assume 5x. If your refinancing model assumes spreads at 400 basis points, budget for 600. The margin is not pessimism. It is recognition that every assumption in your model has a confidence interval, and the tails of those intervals are fatter than your model suggests.
The Kelly Criterion offers a useful reminder: position size should decline when confidence in the inputs declines. In practice, that principle should be reinforced with hard portfolio guardrails. For example: no single name above 5 percent of portfolio NAV, no single sector above 25 percent, and no single vintage above 30 percent. The specific limits should be calibrated to the mandate, liquidity profile, and liability structure. More fundamentally, the discipline is to ask at each allocation decision whether a single exposure, sector, or vintage could impair the fund’s ability to meet its obligations, manage liquidity, or preserve flexibility under stress. If the answer is yes, the concentration is too high regardless of how attractive the credit looks.
The LTCM partners had every advantage: intelligence, credentials, data, models, capital, and track record. They lacked one thing: a system built to survive being wrong.
Several of them started a new fund afterward. They used similar strategies, similar models, similar people. The most spectacular failure in the history of quantitative finance did not change the behavior of the people who lived through it. That should tell you everything about the depth of these patterns and the inadequacy of awareness as a defense.
Pre-approved playbooks, kill criteria, and decision journals are the execution layer. In Part 3, I examine the problem that trips up even well-designed systems: the gap between having the architecture and getting a room full of people with misaligned incentives to execute it under stress. That gap is where substantial capital is often lost.
The full white paper, “Built to Survive Being Wrong: A Practitioner’s Guide to Risk, Bias, and Antifragility,” covers these frameworks in depth. If you want to discuss how they apply to your specific portfolio or governance structure, reach me at tamika@tamikatyson.com.
Sources
Bernstein, Peter L., Against the Gods: The Remarkable Story of Risk, John Wiley & Sons, 1996.
Knight, Frank H., Risk, Uncertainty, and Profit, Houghton Mifflin, 1921.
Lowenstein, Roger, When Genius Failed: The Rise and Fall of Long-Term Capital Management, Random House, 2000.
Kahneman, Daniel, and Amos Tversky, “Prospect Theory: An Analysis of Decision Under Risk,” Econometrica, 1979.
Odean, Terrance, “Are Investors Reluctant to Realize Their Losses?” The Journal of Finance, 1998.
Shefrin, Hersh, and Meir Statman, “The Disposition to Sell Winners Too Early and Ride Losers Too Long,” The Journal of Finance, 1985.
Taleb, Nassim Nicholas, Antifragile: Things That Gain from Disorder, Random House, 2012.
Graham, Benjamin, The Intelligent Investor, Harper & Brothers, 1949.
Kelly, J.L., “A New Interpretation of Information Rate,” Bell System Technical Journal, 1956.
LCD/PitchBook, “Covenant Trends in Leveraged Lending,” Q4 2025. Documentation of the shift from maintenance to incurrence covenants in the syndicated-loan market.
LSTA, U.S. Leveraged Loan Market Statistics, Q4 2025. CLO outstanding volume and reinvestment period data.