Fat tails: why rare losses cannot be removed by an average

Risk cannot be fully expressed by the usual size of daily movement. Research on financial distributions considers heavy tails: large deviations occur more often than a convenient thin-tailed model of comparable scale predicts. Gabaix’s review discusses power-law regularities in returns and volumes. But not every unusual observation confirms a power law, and not every market follows one distribution at every horizon.[1]

Rare does not mean impossible

For a reader, the practical task is more important than choosing one universal curve. Ask what losses the model permits, what its design excludes and whether liquidity is sufficient in an adverse outcome. A short calm history may contain no stress regime. Its absence in the table says nothing about future impossibility. An honest scenario without a probability is often more useful than a precise probability from an unsuitable model.

The same ordinary day, different dangers

Suppose a fictional portfolio’s ordinary daily standard deviation is estimated at 1%. This is a historical characteristic, not a maximum loss. If an analyst automatically equates possible loss with three such units, an unstated distribution assumption has been added. A price jump and execution problems can take the actual result beyond that range. The word “sigma” does not turn a range into a protective barrier.

Separate four quantities: price change, portfolio revaluation, available cash and collateral requirements. An owner with no obligation to sell immediately and a leveraged participant may experience the same price change very differently. Stress testing should therefore describe not only ending value but also cash payments along the way. The starting framework is [portfolio risk](/guides/portfolio-risk/).

An author-created stress scenario

Take 100 fictional units of capital: 60 in asset A and 40 in asset B. In a stated, non-forecast scenario, A loses 30% and B loses 10%. Portfolio loss is 22 units and the remainder is 78. Returning to the original 100 requires growth of about 28.21%, not 22%. Different percentage bases make recovery asymmetric.

Now add a separate obligations payment of 10 units. If no free cash exists, selling part of the assets may become necessary during the adverse revaluation. Do not simply add the payment to the percentage loss: it may repay a previously recognised debt rather than create a new economic loss. Build a separate cash bridge and a separate net-capital bridge.

Which dependencies change in a bad scenario

Do not model only one instrument falling while all else stays unchanged. Set joint changes: more expensive financing, lower liquidity, weaker collateral and more expensive currency conversion. This does not assert that the events must coincide; it tests the ability to withstand their combination. In an educational model, mark which dependencies are author assumptions and which were actually estimated from data.

The momentum-crashes study shows why a strategy’s average historical return can hide rare series of significant losses. That result concerns the strategies and samples studied; it cannot identify the date of the next crash. It is used here as an example of the difference between an average characteristic and stress-state behaviour.[2]

What to ask the model’s author

Ask for raw observations, missing-data rules, distribution assumptions and sensitivity to removing the worst episode. If the result improves sharply after an “outlier” is removed, that observation deserves separate analysis. Ask whether losses are before costs and whether the assumed trades could be executed at the stated volume.

A stress-test conclusion should be conditional: “Under these changes, these losses and payments arise.” It does not state scenario probability or exhaustively assess all dangers. Comparing several conditions and reserves can reveal plan fragility without a numerical forecast of rare catastrophe. Tail shape is discussed separately in [power-law scaling](/guides/power-laws-scaling/).

Reverse stress testing

You can begin with a critical condition rather than a historical fall: what combined cash payment would break the ability to meet obligations? Then choose several different combinations that could reach it. Moderate price decline together with delayed cash receipts may be more dangerous than a larger decline with ample liquidity. This remains scenario analysis, not probability estimation.

Reverse framing shows dependence on the source of vulnerability. If the problem appears after a small change in one parameter, clarify that parameter before complicating the rest of the model. If resilience depends on immediate sale at the last quote, test that assumption specifically. The term [heavy tail](/glossary/tyazhelyy-hvost/) does not replace this work: the loss distribution and ability to survive a particular cash shortage are connected but not identical.

Sources and verification boundaries

[1] [Power Laws in Economics and Finance](https://www.nber.org/papers/w14299). Review of empirical power-law distributions and possible mechanisms.

[2] [Momentum Crashes](https://www.nber.org/papers/w20439). Historical study of rare momentum losses, not a signal.

General financial literacy; foreign research and disclosures are not Russian rules.

Source references retained: **rm-power-laws**, **rm-momentum-crashes**; NBER links, source numbering, formulas, figures, scenario caveats and related links are preserved.