Diversification does not cure fat tails. The controlled experiment, the prescription, and five failed hypotheses.
The Evidence Base:
Over 640,000 daily observations across 68 global futures markets, spanning more than four decades, from September 1984 to July 2026.
One consistent method throughout.
After four episodes spent pulling things apart, we owe you an alternative. We built one empirically rather than rhetorically, and the data won most of the arguments. This episode reports what worked, what didn’t, and the five things we believed going in that turned out to be wrong.
The finding, and who it is news to
Start with the belief almost everyone in this industry holds.
Fat tails are a problem for a single market. Spread your money across enough markets and they average out.
This is the standard defence of Modern Portfolio Theory. It rests on a theorem: add together enough independent random things and the result converges to a bell curve. That is not a rule of thumb. It is proven mathematics.
It is also, in real markets, false, and we did not need this research to tell us so. We have spent years building around exactly this failure. A trend follower diversifies to cast the widest net for outliers and to reduce the luck of catching them, not to smooth returns, and never in the belief that breadth buys tail protection. We keep the risk on each stream small and construct ensembles designed to hold correlation down, because we take it as given that a book which looks diversified in the calm can become a single position at the worst moment. Crisis correlation is not a discovery here. It is a design assumption.
What this research adds is the proof. A controlled experiment that isolates the mechanism and measures exactly how completely diversification fails at the tail.
The controlled experiment
We built portfolios of increasing size, from one market to all 68, two ways.
The real portfolio: actual data, markets left free to crash together.
The independent portfolio: the same markets, same returns, the same fat tails in each, but every history shuffled separately, so no two markets crash at the same moment.
Same ingredients. The only difference is whether markets are allowed to fall together.
The blue line is the independent portfolio, and it behaves exactly as the textbook promises. Add markets, the fat tails melt away, and the portfolio converges on a bell curve. Diversification works perfectly.
The red line is the real world. The tails fall briefly, then climb back up. A fully diversified 68-market portfolio ends up almost as tail-heavy as a single market, and worse than a 5-market portfolio.
Why this happens, and why it still pays
Diversification averages away the ordinary noise. The daily wobble in each market partly cancels against the others, and the body of the distribution thins beautifully. The extreme events do not cancel. They arrive together, as Episode Three showed when correlation tripled in a crisis. So what remains is a thinner body attached to the same tails, which is a more extreme distribution, not a gentler one.
Diversification thins the body. It does not thin the tails.
You cannot diversify your way out of fat tails using instruments whose extremes arrive together, and across all 68 markets they do.
But the tail cuts both ways, and this is what saves the story from despair. A fat-tailed market is not fat only on the downside; the big moves come in both directions, and a wider net reaches further into both. So while breadth cannot save you from the crash, it does the thing a trend follower cares about most: it improves the odds of catching the rare, enormous winner that pays for everything.
This quietly inverts a promise at the heart of Modern Portfolio Theory. The textbook sells diversification as risk reduction bought at the cost of return. In a fat-tailed world, held at constant risk, the sign is backwards. Do what a real allocator does, holding portfolio risk at a target and spending the risk that diversification frees on more positions rather than banking it as lower volatility, and the compound return does not stay flat. It climbs.
The engine is the winning tail. Widening from 5 markets to 68, the best trade grows from around 22% to 44% while the worst grows far more slowly, because per-trade sizing caps the losers and lets the winners run. This is what our Fractals of Finance work found: the tail properties fade slowly, but the tail events, both kinds, grow more likely as the sample grows. Diversification is not the return-diluting hedge the textbook describes. Held at constant risk it is powerfully return-accretive, an engine of return and a spreader of luck. It is simply not a shield, because in the crash the correlations converge. That job falls to structure and size.
The obvious fix does not work
Given everything in this series, the natural remedy is to stop squaring: use a ruler that doesn’t magnify the extremes. We tested it properly, same signal, same markets, same period, changing only the ruler used to size positions.
Changing the ruler does almost nothing. Mean absolute deviation, which we spent Episode One praising, came out marginally worse. Expected shortfall looks better only because it quietly takes more risk for more return, a leverage decision wearing the costume of a measurement.
We were fully prepared to recommend “use a better ruler” as the conclusion of this series. The data refused it. And the reason, once you have seen the diversification result, is not mysterious: on a diversified book the ruler is not the binding constraint. The tails are correlated, they survive the diversification, and changing how you measure a fat tail does not make it thinner.
What did work
One intervention improved outcomes consistently, across every kind of portfolio we tested.
A hard ceiling on borrowing, bolted on top of the standard machinery, reduced the worst loss every time, and on the concentrated portfolio it removed close to four percentage points of drawdown. It works for a reason that has nothing to do with better measurement: it breaks the automatic link between a quiet market and a large position.
That link is the Japanese bond failure from Episode Three. Volatility of 0.9%, model authorises 16.4 times leverage, market moves six tenths of a percent, nine percent of the capital vanishes. A cap makes that sequence impossible, not by measuring the risk more accurately, but by refusing to act on a measurement that has drifted somewhere it no longer means anything.
The deeper reason: how to grow
The cap is one instance of a larger principle, and it answers a question every successful manager eventually faces. Your capital has grown. How do you deploy it?
There are only two directions. Make your existing bets bigger, through leverage or size. Or add more bets, each kept small, spread into streams as uncorrelated as you can find. The two are not equivalent, and the difference is arithmetic.
Scaling up pays less for each unit of risk you add. The first turn of leverage on the diversified book adds 4.5 percentage points of compound growth; the fourth adds only 4.2, while the drawdown swells from 16% to 52%, faster than the leverage that caused it. The culprit is volatility drag, and it grows faster than the leverage itself, so doubling your exposure more than doubles the toll. Widening does the reverse: as we saw, holding risk constant and adding markets lifts the compound return rather than eroding it, because the new risk is fresh ground instead of more weight on the positions where drag compounds.
Think of it as the difference between shouting louder and hiring more voices. Scaling up strains one throat. Widening adds singers. So when the capital grows, spread wider at the smallest position per stream you can run. Do not make the bet bigger. Make more bets. We built the compounding case for this in our Geometry of Wealth work, and the leverage cap is its defensive edge.
One large position that is allowed: the one the market built for you
There is an apparent contradiction here, and resolving it is the difference between the leverage that ruins you and the leverage that pays you. We have just said: do not make your bets bigger. Yet Episode Two rested entirely on a handful of trades that grew enormous. How can both be true?
The answer is which capital is at risk. We enter small, sized from volatility against realised, closed-balance equity, and defended by a stop. That is the position we refuse to inflate with borrowed money, because that would put our principal in front of the next shock. But once a trend runs, we let the position grow on its own, swollen by the winning trade’s own open profit. A small bet becomes, over months, a large one, funded entirely by unrealised gains the market handed us. We are pyramiding on the market’s money, not ours.
That asymmetry is the whole point. We do it only with beneficial volatility, never the adverse kind: losers are cut small by the stop and never added to. So the risk we carry on a large winner is a risk to open profit, not to the realised capital that keeps us solvent. A winner giving some of itself back is a different event, in kind, from a borrowed bet going wrong. The first spends the market’s money; the second spends yours. That is how an outlier hunter runs a huge position without the ruin that sinks a leveraged one: the size is earned, not borrowed, and it is large precisely where the distribution rewards size. Nor is it unique to us. In our Geometry of Wealth work, on a separate set of real track records net of fees, the same shape appears from the outside, with the best 10% of months carrying roughly 90% of terminal wealth. Two datasets, one shape. The wealth lives in the right tail, and a process is either built to hold on through it or it is not.
The prescription
- Cap your leverage in absolute terms, not just relative to volatility. Any model that permits sixteen times exposure because a market has been quiet has confused the absence of recent evidence with the absence of risk.
- Diversify for the net, not for the tail. Breadth is an engine of return: a wider net reaches further into the winning tail, catching more of the rare, large moves that pay for everything, and it dilutes the luck of finding them. Risk-adjusted return rises the whole way. Just do not mistake it for crisis protection, because the correlations that matter converge exactly when you need them not to.
- Measure the path, not the distribution. Drawdown, time spent underwater, the sequence of losses. These are what an investor experiences. Episode Four put a number on the gap. Across the 68 markets, the median worst drawdown took 1,949 trading days, close to eight years, to travel from its peak to its low, and fewer than one in ten of those declines reached bottom inside a single year. A distribution has no duration. The thing that does the damage lasts the better part of a decade, and no summary statistic computed on the whole sample contains that fact. Path measures are harder to optimise because the mathematics isn’t smooth, which is exactly why the industry avoids them.
- Size for survival, not for score. The floor at zero has no counterpart on the upside, and any rule that treats them symmetrically is mispricing the only asymmetry that matters.
- Grow by widening, not by scaling. When your capital increases, add uncorrelated return streams at the smallest workable size rather than enlarging the bets you already hold. Leverage and bigger positions buy return at a worsening exchange rate, because the drag they add compounds against you. A wider book does not.
- Enter small, let winners run. Size entries small against realised equity, defended by a stop. Then let a winning position grow on its own open profit rather than by adding borrowed money. The large position an outlier hunter wants is the one the market built out of unrealised gains, not the one leverage built out of principal. Harvest beneficial volatility; never harvest the adverse kind.
- Remember where the money comes from. Episode Two found that 4.6% of trades produced 100% of the profit. A framework designed to suppress variation is designed to suppress the thing that pays you.
We should be honest about the limits here. Points one, two, five and seven are tested and they replicate: the leverage cap, the constant-risk widening result which carries both points two and five, and the concentration of profit in a handful of trades all come straight out of the data. Points three, four and six describe how we build and follow from the evidence and from principle, but we have not run them to ground in this series the way we ran the leverage cap. That work is ahead of us, and we would rather say so than dress an argument up as a result. Since publishing, the addendum to Episode Four has made a start on points three and four. On a single index, a symmetric volatility target improved every number on the risk report and left the investor with forty percent less money, while an asymmetric rule that estimates nothing more than halved the worst loss and raised the compound return. One index is not 68 markets, and we are not going to treat it as though it were.
What we got wrong
Five hypotheses went into this research with our confidence behind them. In several places the evidence forced us to change our minds. In one place it confirmed something we already believed, and then surprised us anyway.
- We expected that deleting the winning trades would improve the Sharpe ratio. It does not. Deleting the best trade lowers the score in 68 of 68 markets. The popular claim that great trades hurt your risk-adjusted returns is false, and we were preparing to publish it. What is true is subtler and worse: winners help your score far less than they help your wealth. Two cents on the dollar.
- We expected volatility to miss real risk entirely. It doesn’t. It explains about three quarters of the variation in worst-case loss between markets. We tried to break it three ways in Episode Four and could not. The absolutist version would have been dismantled within a day. The narrower claim survives: volatility tells you the order of magnitude of the danger and cannot locate you within a range fifty points of capital wide.
- We expected the long-run volatility estimate to be visibly unstable. It isn’t. Removing the October 1987 crash from four decades of S&P data shifts the estimate by less than 3%. Given forty years, the number settles. The instability lives at the three-year horizon, which turned out to be far more damaging, because that is where every real capital decision gets made.
- We expected the leverage trap to depend on calm being uninformative. It does not, and that is worse. Here the surprise ran the other way. Our earlier Fractals of Finance work had already shown that volatility clusters, with a Hurst exponent on absolute returns averaging 0.87 across the 68 markets, so we knew going in that calm does predict calm, and the data confirmed it: a shock is roughly three times less likely after a quiet stretch. We assumed that persistence was the risk manager’s friend. It is not. Calm shrinks the eventual shock to about a third of its size, but it more than triples the position, and the second effect is larger. The safety is real and the models still blow up, because they respond to the calm faster than the calm protects them.
- We expected a better ruler to be the fix. It isn’t. This was the conclusion we intended to write.
We also abandoned, before testing, the claim that market variance is mathematically infinite. Defending it would have meant spending the series arguing about a parameter instead of making an argument. The two-ruler test in Episode One gives us what we needed and assumes nothing.
Data hygiene, published in full
Two markets contained errors in the source data that would have corrupted our results.
Euro FX carried pre-1999 contract data on a different price scale, producing fictitious daily returns as large as 96.8%. Truncated to 28 December 1998 onward.
Natural Gas had a similar scaling error. Truncated to 10 August 2005 onward.
The other 66 passed unchanged, and every headline result survived the correction.
We raise this because an earlier draft reported that volatility explains only 41% of drawdown variation. That figure was an artifact of the corrupted markets. The correct figure is roughly 76%. It is considerably less flattering to our argument, and it is what we have published.
Building for the world that exists
Step back from the individual results and a single design philosophy connects them, and it is the mirror image of the one the Gaussian framework assumes.
A mean-variance investor holds the market to capture its average, exposed to the whole distribution, the dull middle included, because the model says the middle holds the information and the tails are a rounding error. Episode One showed this is backwards: the tails carry the risk and, for a trend follower, all the reward. So the design should be inverted, and ours is. Four choices, each drawn from how fat-tailed markets behave rather than how a bell curve is assumed to.
We refuse the middle. Entering only on a qualifying trend and standing flat otherwise, we decline to sample the bulk of the distribution. We want exposure only to the tails, where the payoff lives. Our Fractals of Finance work showed this is a real division in the data: breakout periods are persistent while flat periods mean-revert hard. Entering on the breakout selects the trending tail; standing flat declines the oscillating middle. Variance cannot see this choice, because it assumes you are always invested.
We manufacture the skew. Most futures markets are roughly symmetric in their extremes, with no natural tilt to harvest. Strict stops amputate the left side of our realised distribution and leave the right to run, turning a raw skew near zero into a realised skew of around +0.7. That is not skew we found. It is skew we built, and a symmetric measure like variance is blind to it. Nor is it a quirk of our sample: in our Geometry of Wealth work, 37 of 41 long-run managers produced positive skew, in a world where the buy-and-hold equity investor lives with negative. The cut manufactures skew wherever it is applied.
We pyramid through structure, not leverage. Conventional pyramiding stacks correlated risk on one signal that can fail all at once. Instead we run an ensemble of different structures. An infant trend trips only the fastest system; a major outlier has every system active and earning. The book grows large on exactly the moves that deserve size, through breadth of structure rather than borrowed money.
We size from realised capital. Every entry is small against closed-balance equity, so all of the above runs without ever putting principal in front of a shock. The machine reaches hard for the right tail while the capital that keeps us solvent is never the capital at risk.
None of the four is visible inside a Gaussian frame. Selective entry is invisible to a model that assumes full exposure, engineered skew to a symmetric measure, ensemble structure to a correlation that only holds in the calm. And the outlier the whole design reaches for is the very event the Sharpe ratio under-credits and Black-Scholes calls impossible. These are not refinements of the standard framework. They are what you build once you stop believing it.
What the whole thing amounts to
The square is not a mistake. It is a theorem with a domain of validity, and Fisher proved it optimal inside that domain. The domain is the bell curve, and the neighbourhood around it where the theorem still holds is roughly one contaminated observation in three hundred. Real markets sit at one in 243, on the most extreme days alone.
We are outside the domain, and so is every model in Episode Three. The profession has spent seventy years building an ever more sophisticated apparatus on a foundation whose assumptions gradually became accepted as reality. The fix is not a cleverer ruler; we tested that. It is to cap the leverage, respect the path, accept that diversification buys a thinner body and not a thinner tail, and stop optimising against a number that cannot see the events which decide whether you are solvent in ten years.
Finance has spent decades trying to compress uncertainty into a single number, and the markets have spent those same decades refusing to be compressed. A market is a path, not a point. Capital survives or vanishes one sequence at a time, and the sequence is precisely what our instruments were built not to see.
We began with a teacher’s ruler laid flat on the floor, measuring the distance to a colleague across the office. It was the right tool for that room. The mistake was carrying it out into the world and trusting it to measure the storm.
Risk is not variance. Risk is the size of the loss you cannot come back from, arriving in an order no average can predict, on the one path you get to walk. The question was never how volatile the journey looked from the outside. It was whether you were still standing when the move that mattered finally came.
Richard Brennan writes on systematic trading, complex adaptive markets, and the philosophical foundations of trend following at atstradingsolutions.com. His books include The Fractals of Finance, Complex Adaptive Markets, Carved by Impossibility and The Aussie Turtles Trend Following Guide.
Want to explore why structure exists at all?
Carved by Impossibility: What Remains When Everything Else Is Eliminated
The book explores the architecture of constraint, emergence, and reality itself, and what it means for how we understand markets, life, and the universe.
Available now on Amazon in paperback, hardcover, and Kindle.
Want the theoretical foundation for why trend following works?
The Fractals of Finance: Determinism, Adaptation and the Geometry of Markets
The book explores the full architecture of feedback, fat tails, and fractal structure in financial markets, and what it means for how we trade, invest, and understand risk.
Available now on Amazon in paperback, hardcover, and Kindle.
Want a practical field manual for trading trends and capturing outliers?
The Aussie Turtles Trend Following Guide: A Field Manual for Hunting Outliers adapts the timeless principles of the original Turtle traders into a systematic, rules-based approach for modern markets. Co-authored with Adam Havryliv.
Available now on Amazon in paperback, hardcover, and Kindle.