7. Black swans: when the improbable dominates the outcome

Level 2: Distributions and why markets are not "normal"

A fund run by Nobel laureates in economics. Sophisticated mathematical models. And a collapse in 1998, when debt markets moved in a way those models considered practically impossible.

Long-Term Capital Management did not blow up for lack of mathematics. It blew up by trusting a distribution that was not the right one in the tails.

The pieces of this level, together

Throughout this level you have been accumulating pieces that now fit into a single picture:

• Real returns have fatter tails and more negative skew than a comparable normal (module 3).
• Prices compound multiplicatively, generating a log-normal distribution with unlimited room on the upside (module 4).
• VaR computed under normality assumptions tends to underestimate the risk of severe losses (module 5).
• The central limit theorem smooths aggregate long-run behavior, but offers no protection at all to the individual day against its fat tails (module 6).

The concept of a black swan is, in a sense, the practical synthesis of all of the above. It is the name given to the events that suddenly materialize all that tail risk which normal theory had been underestimating.

The three features of a black swan

A black swan, as the concept is used in finance, is defined by three features that must occur together:

• It is a rare event. It lies outside the range of what normal experience, or standard statistical models, consider reasonably expectable.
• It has an extreme impact. When it happens, its consequences are disproportionately large compared with 'normal' events.
• It is rationalized after the fact. Once it occurs, narratives are constructed that make it look predictable or even obvious.

The third feature is the subtlest one, and it is worth pausing on. In the moment beforehand, almost nobody anticipated it at the magnitude it ultimately reached.

That combination of genuine surprise at the time and false retrospective obviousness is what separates a black swan from a plain 'bad day in the market'.

Why the normal rules them out

The underlying statistical reason why normal-based models systematically underestimate these events has to do with how quickly probability decays in the tails.

In a normal distribution, the probability of observing a value falls exponentially the further you move from the mean. Each additional standard deviation cuts the probability drastically.

So an 8 or 10 standard deviation move becomes, under that model, something that should never be observed in practice.

Empirical evidence from real markets, by contrast, suggests otherwise. The tail decays far more slowly than that, in a way closer to what statisticians call power-law behavior. Extreme events are still rare, yes, but vastly less rare than the normal predicts.

That difference in decay speed may look like a minor technicality. It is what separates a model that says 'this cannot happen' from the historical record in which it does happen, and not just once.

Episodes with the same signature

Market history offers recurring examples of this discrepancy:

• The already mentioned Black Monday of 1987, with a fall of more than 20% in a single S&P 500 session.
• The collapse of the Long-Term Capital Management fund in 1998, the one that opened this module.
• The 2008 financial crisis, with credit and liquidity moves that many banks and insurers had classified in their own risk models as near-zero probability events.
• The 2010 flash crash and the March 2020 pandemic crash, as more recent examples.

The LTCM case is a particularly emblematic one. A fund run by Nobel laureates in economics, with sophisticated mathematical models resting largely on normality assumptions, blew up when debt markets moved in a way its models considered practically impossible. It is the demonstration of how excessive confidence in normality can lead to real, not merely theoretical, disaster.

Each episode had its own trigger, and all of them share the same underlying statistical signature: moves that, under a normal model calibrated on 'usual' volatility, carried a probability so low they were treated, for practical purposes, as dismissible.

When diversification fails just when it counts

There is a further dynamic that aggravates these episodes and is worth understanding.

During calm periods, assets typically show moderate correlations with one another. That gives investors the sense of being reasonably diversified.

But precisely during tail events, correlations between assets tend to spike upward. Everything falls at once, including assets that in normal times moved fairly independently.

The reason is mechanical. Forced selling, margin calls and generalized panic push people to sell whatever can be sold, not just what it would in theory make sense to sell.

This means that diversification, the most cited risk management tool of all, tends to fail precisely when it is needed most. Another reason why risk models calibrated in quiet periods give a false sense of security.

What to do about it

What does all this imply for practical risk management, beyond simply 'knowing that black swans exist'? Three things.

First, complement VaR with stress testing. The question is not only what the 5th percentile loss is according to recent historical data. You also explicitly simulate extreme scenarios already observed in history: a repeat of 1987, of 2008, of 2020. And you look at how the current portfolio would respond. That way you do not depend on those scenarios being well represented in the recent data window.

Second, size positions and leverage levels on the assumption that the worst plausible scenario is far worse than three standard deviations. The historical evidence shows moves of far greater magnitude occurring with a frequency that is anything but negligible.

Third, remember that correlation between assets is not a fixed constant. It can change drastically at exactly the moment you are relying on it most, which limits how much real protection diversification offers in the worst scenarios.

Statistical honesty

The closing message of this level is, fundamentally, a matter of statistical honesty.

The normal distribution is a useful starting point, mathematically tractable and reasonable as a first approximation. It is not a law of nature that markets are obliged to obey.

Fat tails, negative skew, the multiplicative structure of prices and the limits of the central limit theorem all point in the same direction.

Extreme events in financial markets are not a marginal mathematical curiosity. They are a structural and recurring feature that any serious quantitative analysis has to build in from the start, rather than discover far too late.