Skewness and Tail Risk: Why a Profitable Average Can Hide Large Losses

Skewness and Tail Risk: Why a Profitable Average Can Hide Large Losses

A strategy can show a positive average return, a high win rate and a calm equity curve—then lose months of gains in one move. Mean and volatility describe part of the record. Skewness and the left tail show whether rare losses dominate it.

বাংলায় পড়ুন: Skewness ও Tail Risk: Profitable Average কীভাবে বড় Loss লুকায়

Learning objective: distinguish positive and negative skew, inspect the loss tail instead of relying on an average, calculate a simple Expected Shortfall, and stress-test strategies whose frequent small gains depend on avoiding a rare large loss.

1. Start with the whole return distribution

A return series has more information than its average:

• Mean: the arithmetic average outcome;
• Median: the middle observation;
• Volatility: how widely outcomes vary around the mean;
• Skewness: whether one side of the distribution stretches further;
• Tail: the rare, extreme observations at either end.

CFA Institute's statistical measures of asset returns treats central tendency, dispersion, skewness and kurtosis as separate parts of return analysis. One number cannot replace the others.

2. Positive and negative skew in plain English

Positive skew has a longer right tail. Many outcomes may be small losses or modest gains, with occasional very large winners. A trend-following system that cuts losses and lets a few trends run can have this shape.

Negative skew has a longer left tail. It may produce frequent small gains and occasional severe losses. Strategies that earn a premium for absorbing crash, gap, liquidity or short-option risk can resemble this shape.

Skewness does not label a strategy good or bad. It tells the trader what kind of surprise matters. A positively skewed strategy can still be unprofitable; a negatively skewed strategy can earn a return, but its rare loss deserves explicit capital planning.

3. Worked example: the same average, a different experience

Compare two simplified 20-period records before costs:

• Strategy A: nineteen returns of +1% and one return of −15%;
• Strategy B: twenty returns of +0.2%.

Both have the same arithmetic mean:

[(19 × 1%) − 15%] ÷ 20 = +0.2%

Yet their compounded results differ:

• Strategy A: 1.01¹⁹ × 0.85 − 1 ≈ +2.68%;
• Strategy B: 1.002²⁰ − 1 ≈ +4.08%.

Strategy A also has a 95% win rate, but the single loss erases fifteen ordinary winners and creates a negatively skewed distribution. The example shows why mean return and win rate can look attractive while path risk remains uncomfortable.

4. Why volatility and Sharpe can miss the shape

Standard deviation treats upside and downside deviations symmetrically. A Sharpe ratio then divides excess average return by that volatility. This is useful for comparison, but a smooth stream of small gains followed by one crash is not equivalent to a symmetric series with the same mean and standard deviation.

CFA Institute's review of higher moments and disaster risk notes that traditional measures can understate skewness and rare-disaster exposure. Use the Sharpe and Sortino guide alongside tail measures, not as the final verdict.

5. Inspect the worst observations directly

Sort returns from worst to best and report:

1. worst single-period return;
2. worst five and ten observations;
3. maximum drawdown and recovery time;
4. largest gap beyond the planned stop;
5. losses during market stress;
6. results before and after fees, spread and financing.

Then examine whether those losses share a cause: earnings gaps, volatility spikes, crowded exits, exchange closures or correlated positions. A tail statistic without a mechanism may describe the sample but fail to predict the next regime.

6. VaR is a threshold; Expected Shortfall looks beyond it

Historical Value at Risk (VaR) estimates a loss threshold at a chosen percentile. With 100 equally weighted daily observations, a simple 95% historical VaR is around the boundary separating the worst five observations from the other 95.

VaR does not say how severe losses are after that boundary. Expected Shortfall (ES) averages the tail losses beyond it.

Suppose the five worst daily returns are −4%, −5%, −6%, −8% and −12%:

95% historical ES = (−4% − 5% − 6% − 8% − 12%) ÷ 5 = −7%

The estimate says the average of this sample's worst 5% was a 7% loss. It does not promise that the next tail loss is capped at 7%.

7. Small samples make tail estimates fragile

In 100 observations, 95% ES depends on only five data points. In 250 daily observations, a 99% tail contains only a few observations. One new shock can change the estimate sharply.

Improve the review by:

• using enough data to include different regimes;
• preserving crisis periods rather than removing “outliers” automatically;
• block-bootstrap sampling so clustered returns remain together;
• stressing losses beyond the historical worst;
• testing wider spreads, gaps and unavailable liquidity;
• reporting confidence ranges instead of false precision.

Historical skewness is evidence about the sample, not a stable law of the strategy.

8. Translate tail risk into position rules

If a strategy can lose several times its normal stop during a gap, position size should use the stressed loss, not only the intended stop. Correlated positions should be shocked together.

A practical review might ask:

• What happens if the worst historical loss occurs twice?
• What if volatility doubles before the position can be reduced?
• What if all positions in the same theme gap together?
• What if the exit fills one or two price levels beyond the stop?
• Does the resulting drawdown remain below the account's ruin threshold?

Tail analysis is useful only when it changes size, diversification, hedging, execution or the decision to avoid the trade.

Common mistakes

1. Choosing a strategy from average return alone.
2. Treating a high win rate as evidence of low risk.
3. Assuming returns follow a normal distribution without checking.
4. Deleting extreme losses as data errors without investigation.
5. Reporting VaR as a maximum possible loss.
6. Estimating a 99% tail from too little data.
7. Ignoring clustered losses and changing market regimes.
8. Measuring each position separately while missing a common portfolio tail.

Practical checklist

• Mean and median both reported?
• Histogram or empirical distribution inspected?
• Skewness direction understood?
• Worst observations reviewed individually?
• Maximum drawdown and recovery time measured?
• VaR paired with Expected Shortfall?
• Sample has enough tail observations?
• Crisis and gap scenarios retained?
• Correlated positions stressed together?
• Tail findings reflected in position size?

The Money Man View

An average is a summary, not a survival plan. Traders should be most suspicious when returns look unusually smooth, the win rate is very high and the loss mechanism is described as “unlikely”. The question is not only how often a strategy loses, but how much it can lose when the familiar pattern breaks.

Educational risk disclosure

This tutorial uses simplified hypothetical examples and was prepared with AI-assisted research and writing under The Money Man editorial workflow. It is educational information, not personalised investment advice. Tail estimates depend on limited historical data, extreme events can exceed the sample, and trading can cause substantial loss.