Most summaries compress a trading system into win rate and average return. That compression can hide the defining fact of an asymmetric strategy: the result may depend on a tiny fraction of observations that look nothing like the typical trade.

Skew describes which tail extends farther

A symmetric distribution has similarly shaped left and right sides around its center. A positively skewed distribution has a longer or heavier right tail. Most observations can sit near small losses while a few large positive observations extend far to the right.

The standardized third central moment is one formal measure:

Skewness = E[(R − μ)3] / σ3

Cubing deviations preserves their sign and gives distant observations enormous influence. Positive values generally indicate a longer right tail. In small trading samples, however, estimated skewness itself is unstable because it depends so heavily on rare observations.

The median trade can lose while the mean trade wins

Consider a deliberately simple 100-trade distribution:

Outcome bucketCountR eachContribution
Controlled losses70−1R−70R
Ordinary winners25+2R+50R
Tail winners5+8R+40R
Whole sample100+0.20R mean+20R

The median is −1R because the 50th and 51st sorted observations are both losses. The mode is also −1R. Yet the mean is +0.20R. A trader experiences a loss as the most common and middle outcome while relying on the average pulled right by a few large wins.

The behavioral mismatch: the process feels like its median one trade at a time, but pays like its mean only across a sufficiently complete sample.

Measure how much the best trades contribute

Remove the five +8R tail winners from the example. The remaining 95 trades total −20R. The top 5% contributes +40R, or 200% of the final +20R net result. Values above 100% are possible because the rest of the sample loses money.

Tail contribution = sum(top x% outcomes) / total net outcome

Useful robustness views include:

  • Top-trade removal: recompute results without the best 1%, 5%, and 10%.
  • Trimmed mean: remove equal percentages from both tails to describe the central process.
  • Winsorized mean: cap extreme observations at chosen percentiles instead of deleting them.
  • Contribution concentration: rank outcomes and graph cumulative gross profit by percentile.
  • Time concentration: identify whether one market or short regime produced the entire tail.

A system collapsing when its best outcomes are removed does not automatically prove it is invalid. Tail dependence may be the intended mechanism. It does prove that missed signals, execution limits, and sample uncertainty matter much more than the headline average suggests.

A fat tail is not the same as positive skew

Skew describes asymmetry. Tail thickness describes how quickly extreme-event probability declines. A distribution can be symmetric and fat-tailed, skewed with relatively thin tails, or both skewed and fat-tailed.

Financial returns often produce more extreme observations than a normal model predicts. That empirical statement does not establish one universal power law, and finite samples make precise tail-shape estimation difficult. For practical risk work, the important point is simpler: a normal curve can assign too little probability to both extreme gains and extreme losses.

Do not use "fat tail" as a synonym for "good upside." The left tail can be fat too. Leverage, gaps, crowded exits, short options, and liquidity failure can create rare losses that overwhelm years of small gains.

The strategy transforms the market distribution

A market's return distribution and a strategy's trade distribution are not identical. Rules transform one into the other. A stop can truncate many ordinary losses without placing a hard bound on gap risk. A trailing exit can keep exposure during a persistent move and create a long right tail. A fixed profit target can do the opposite by clipping large wins.

A stylized trend-following transformation looks like this:

  1. Enter on evidence that a move may be developing.
  2. Exit failed attempts near a predefined invalidation point.
  3. Do not impose a small fixed cap on a trade that continues to move.
  4. Repeat across time or instruments so rare trends can appear in the sample.

None of those rules guarantees positive expectancy. They explain a mechanism that could produce positive skew. Evidence must still show that the right tail pays for the false starts after costs.

The biggest operational risk is missing the outlier

If five trades carry the sample, skipping one removes 20% of the intended tail bucket. Skipping signals after a losing streak is not a small implementation variation. It selectively removes opportunities at exactly the point when confidence is lowest.

The same problem appears when positions are closed early to "lock in" a normal gain. Turning the five +8R trades into +2R trades changes their contribution from +40R to +10R. The sample moves from +20R net to −10R net even though the win rate improves not at all.

A practical tail audit

For any claimed asymmetric edge, calculate and inspect:

  • Mean, median, mode, standard deviation, and downside deviation in R.
  • Win rate, average win, average loss, profit factor, and maximum realized loss.
  • The percentage of net return created by the best 1%, 5%, and 10% of outcomes.
  • Results with those observations removed and with plausible execution degradation.
  • Tail concentration by instrument, year, volatility regime, and direction.
  • How often valid signals were missed and what those missed outcomes became.
  • Whether large winners survive out-of-sample and under nearby parameter choices.

The purpose is not to make the tail disappear. It is to identify exactly how much the process depends on it, then decide whether execution and evidence are strong enough to support that dependence.