If one program beats Berkshire on drawdown-adjusted compounding, it might be luck. If seventeen do, using different models, in different countries, across different decades, it is evidence of something deeper than skill.
Episode 10 presented the individual compounding machines: dozens of track records ranging from $1.32 to $75.50 per dollar invested. The range was dramatic, but the convergence was more important. Across the long-term programs analyzed, roughly seven in eight produced positive skew, about four in five produced negative equity correlation, and seventeen, more than two in five, beat Berkshire Hathaway on MAR ratio. These are not properties of any individual. They are properties of the process.
This convergence is the argument. A single program’s excellence might be explained by luck, market selection, or a favourable sequence of returns. When dozens of independent implementations converge on the same distributional properties, the explanation must be structural. The process determines the direction. The individual program determines the magnitude.
A word on the universe is necessary before the evidence. The NilssonHedge database tracks well over a hundred trend-following programs, the great majority of which have track records too short to test across multiple market regimes. This series assesses only those with a sufficient record. The TTU TF Index comprises 47 systematic, globally diversified programs, each meeting a minimum 15-year unbroken track record. The per-program statistical exhibits in this episode use the stricter subset of 38 programs whose records span at least 20 years, the same horizon-consistent subset used in Episode 10, so that every program shown has been tested through at least the global financial crisis. Where this episode quotes convergence as proportions, the proportions are stable across all of these roster definitions.
But convergence on favourable properties does not prove the case. The sceptic has two powerful objections. First: survivorship bias. The database retains programs that survived; perhaps the failures had different properties, and the convergence is an artefact of selection. Second: start-date dependence. The analysis begins in January 2000, three months before a major equity crash, so any strategy that was not long equities at that moment would look good. These are legitimate objections. This episode confronts them directly.
The Statistical Fingerprint
Figure 1 shows the distribution of four key properties across the 38 long-term programs.
Figure 1: Statistical properties of the 38 long-term trend followers with 20-year-plus records. Top left: CAGR distribution with the S&P 500 reference line. Top right: MAR ratio distribution with the Berkshire reference line. Bottom left: skewness distribution, overwhelmingly positive. Bottom right: S&P 500 correlation, clustered negative. All returns net of fees.
The median program in this set has a CAGR of 6.4%, a maximum drawdown of 32.3%, a MAR of 0.201, a skewness of +0.26, an S&P 500 correlation of -0.07, and an annualised volatility of 14.5%. All returns are net of management and performance fees.
The typical machine underperforms the S&P 500 on raw CAGR by under two percentage points. This is the fact that critics correctly identify. But the typical machine also produces positive skew where the S&P 500 produces negative skew, contains its drawdown to roughly a third where the S&P 500 suffered its full 50.9% GFC decline, and delivers negative equity correlation where the S&P 500 is, by definition, perfectly correlated with itself. The modest CAGR shortfall is the price of a completely different return geometry.
The distributions themselves are revealing. The CAGR distribution is positively skewed, with a long right tail driven by programs like Mulvaney. The skewness distribution is overwhelmingly positive: only 4 of the 38 programs show negative skew, and even those are only slightly negative. No program has an S&P 500 correlation above +0.27. The database does not contain a single trend follower that moves with equities during crises. Thirty-four of 38 show positive skew, thirty of 38 show negative equity correlation, and seventeen of 38 show a MAR exceeding Berkshire. The convergence is not clustering around a mediocre outcome. It is clustering around a specific geometric fingerprint: positive skew, negative equity correlation, and competitive drawdown-adjusted return.
Survivorship Bias: The First Objection
The survivorship objection is the most serious challenge to any analysis of long-term track records. It deserves a thorough response, not a dismissal. The objection runs as follows: any database of surviving programs excludes those that failed, closed, or merged. The programs with multi-decade records are, by definition, survivors. Perhaps the failures had negative skew, positive equity correlation, and poor MAR ratios, in which case the convergence on favourable properties would be an artefact of survival, not evidence of the process. There are four responses, each addressing a different dimension.
First: the fingerprint is stable across roster definitions. The convergence does not depend on a single cut of the data. Whether the population is defined as the 38 programs with 20-year records, the 47 in the TTU Index with 15-year records, or the full set of long-record share classes in the database, the proportion showing positive skew stays near seven in eight and the proportion showing negative equity correlation stays near four in five. The fingerprint does not emerge only in the very longest-surviving programs. It is present wherever the record is long enough to measure, which is the opposite of what a pure survival artefact would produce.
Second: survival is a consequence of the process, not its cause. Programs that cut losses, produce positive skew, and contain drawdowns are precisely those most likely to survive decades. A program with negative skew and deep drawdowns is more likely to suffer a terminal loss, a client exodus, or a closure. Survivorship in this context is not a statistical artefact. It is the process revealing itself: the programs that survived did so because they adhered to the structural properties that the geometric framework predicts should produce longevity.
Third: the range of survivors includes near-failures. Episode 10 documented that the bottom of the surviving set includes a program that turned $1 into just $1.32 over the full window, a practical failure by any investment standard. That program survived but delivered essentially zero real return. Even this near-failure exhibits the geometric fingerprint: positive skew, negative equity correlation, and contained drawdowns. The structural properties are not an artefact of success. They persist even among the weakest survivors.
Fourth: the historical record of closures inverts the objection. The survivorship argument assumes the departed programs failed catastrophically, blowing up and destroying capital with the negative skew and deep drawdowns that would undermine the fingerprint. In equity long-short and leveraged macro that assumption is often correct: LTCM, Amaranth, and others suffered permanent capital impairment. But the systematic trend-following landscape tells a different story. Many early CTAs that no longer operate in their original form, programs such as Mint, John W. Henry, and other respected pioneers, did not implode. The dominant exit routes were founder retirement after decades of success, asset-base shrinkage during long trendless stretches, and consolidation into multi-strategy firms. Capital was generally returned to investors, not destroyed, and many of these programs carried 20-to-30-year records before closing.
This is a fundamentally different survivorship narrative than the objection assumes. If the programs that left the database were also trend followers exhibiting the same geometric fingerprint, then survivorship bias does not inflate the evidence. It understates the population that exhibited the fingerprint. The missing programs were not the negative-skew, deep-drawdown failures the objection requires. They were additional positive-skew, crisis-alpha producers that happened to stop operating.
One further point strengthens the response. All returns in the database are net of management and performance fees. The positive skew, the negative equity correlation, and the competitive MAR ratios survive the full fee burden. Survivorship bias cannot explain properties that persist after fees have been deducted, because fees are a real cost that erodes real returns. The geometric fingerprint is net of the most common objection to hedge-fund performance.
Start-Date Dependence: The Second Objection
The analysis in this series begins in January 2000. The dot-com crash began three months later, so any strategy that was not long equities at that moment would look good by comparison. The sceptic asks whether the story would be different if the series started in 2003, or 2009, or any date that did not coincide with a crash. The sceptic is right. The story is different. Here it is.
Figure 2: Sub-period CAGR comparison. The S&P 500 wins decisively during bull markets. The TF Index wins during crises. The mechanism is consistent across all sub-periods regardless of start date. All returns annualised, net of fees.
Table 1: Sub-period annualised CAGR, S&P 500 versus the TF Index, net of fees. The final two rows remove the dot-com crash entirely and start at the exact GFC bottom, the two most hostile framings for trend following.
The pattern is unambiguous. The S&P 500 outperforms during bull markets: +14.5% versus +10.5% in the post-dotcom recovery, +17.1% versus +2.5% in the 2009 to 2019 expansion, +22.8% versus +0.8% in the recent rally. The TF Index outperforms during crises: +18.2% versus -17.3% in the dot-com bust, +28.9% versus -41.4% in the GFC, +11.3% versus +7.0% across COVID and 2022. This is not a flaw in the analysis. It is the mechanism. Trend following underperforms during sustained equity rallies because it is not concentrated in equities; it outperforms during crises because it captures the trends that crises produce. The start date does not change the mechanism. It changes how much of each regime the observation window captures.
The Lost Decade: 2010–2019
The period from 2010 to 2019 is the single most challenging decade for the trend-following argument. During these 120 months the S&P 500 returned 13.6% annually and the TF Index returned 3.3%. The gap of 10.3 percentage points per year, sustained for a full decade, tested the conviction of every allocator who held trend-following exposure.
This decade was not a normal market environment. It was the product of an unprecedented policy experiment. Coordinated central-bank intervention, through quantitative easing, near-zero interest rates, and forward guidance, deliberately suppressed the volatility and cross-asset dispersion that trend following requires. The equity rally was remarkably steady, with a maximum drawdown of only about 16%, never deep enough to trigger the crisis-alpha mechanism. Cross-asset trends were muted, with commodities, currencies, and bonds all exhibiting lower directional persistence than historical norms. All three features were direct consequences of a policy regime with no precedent in financial history.
This distinction matters. The standard framing treats the lost decade as the normal operating cost of holding trend following, the price of insurance during calm years. That framing is incomplete. The 2010 to 2019 decade was not merely calm. It was a regime in which central banks explicitly engineered the suppression of the very market dynamics that trend following captures. Independent quantitative research covering the same period identifies three structural forces that combined to impair returns: a suppression of directional persistence across futures markets to its weakest readings in decades; a collapse in the standalone profitability of the short side, as central-bank backstops and passive inflows supported falling markets; and a redistribution of the available edge from shorter lookbacks, where systematic crowding reduced signal quality, toward longer horizons. The process was not broken. It was suppressed by concurrent forces, all traceable to the same policy regime.
The honest conclusion is that this decade tested the patience of every investor who held a trend-following allocation. Many reduced or eliminated their positions. The geometric argument was always prospective: trend following’s value would be demonstrated in the next crisis. During the 2010 to 2019 decade, the next crisis never came. Then it did. When the policy experiment ended and inflation returned in 2022, the mechanism reasserted itself immediately. Stock-bond correlations surged as bonds and equities fell together, breaking the traditional 60/40 framework. From January 2020 through December 2022 the TF Index returned 11.3% annually while the S&P 500 returned 7.0% and suffered a drawdown approaching 25%. The investor who held through the lost decade was rewarded with exactly the crisis alpha the geometric framework predicts. The cycle then repeated: from 2023 to January 2026 the S&P 500 surged at 22.8% while the TF Index returned 0.8%.
The 2010 to 2019 decade is included in every calculation in this series. No data has been excluded. The gap was real, and for ten years it felt permanent. But the extended record provides perspective the 26-year window cannot. DUNN WMA absorbed the full weight of that hostile decade and still compounded at 13.30% over 39 years, turning $1 into $130. EMC Classic absorbed the same decade and compounded at 12.87%, turning $1 into $112. If the QE decade was the worst regime the process has ever faced, and it was, both programs compounded through it and still outpaced the S&P 500 over the full cycle.
The forward-looking implication is that the historical dataset may understate the process’s long-term potential, because it includes a decade of conditions engineered to suppress the very dynamics the process captures. If structurally higher inflation, more volatile monetary policy, and positive stock-bond correlations persist, the environment ahead is more favourable than the historical average, not less. This is not a prediction. It is a regime observation. Episode 12 will quantify the full patience cost. For now, the relevant point is that the sub-period pattern held even through the most hostile regime in modern history: underperformance during the bull, outperformance during the crisis, exactly as the mechanism predicts.
The Honest Concession
If an investor started in March 2009, at the exact bottom of the GFC, and held through January 2026, the S&P 500 returned 16.2% annually and the TF Index returned 3.7%. Buy-and-hold won this path decisively. The geometric advantage requires crises to occur. In a world of permanent bull markets with shallow drawdowns, trend following would permanently underperform.
The argument is not that trend following always wins. It is that crises always come. The question is not whether the S&P 500 will outperform during the next bull market; it will. The question is when, not whether, the next 40%-plus drawdown occurs, and what happens to the investor’s compounding base when it does. The sub-period data also addresses the start-date criticism directly. Excluding the dot-com period entirely, from January 2003 to January 2026, the S&P 500 returned 11.4% annually and the TF Index returned 6.4%. The TF Index still underperforms on CAGR, as expected, but the mechanism is unchanged: it captured the GFC, the 2020 crisis, and the 2022 drawdown while the S&P 500 captured the bull markets between them.
The fact that roughly four in five trend followers underperform the S&P 500 on raw CAGR is the headline a critic will write. The fact that fire insurance underperforms the stock market in years without fires is the same observation, stated differently. The correct metric is portfolio-level geometric efficiency, not standalone CAGR, and Episode 9 demonstrated that the 60/40 blend produces more terminal wealth than either component alone.
The Strongest Counter-Argument
Before presenting the extended record, this section pauses to give the strongest counter-argument the space it deserves. A series that presents only the evidence in favour of its thesis is advocacy. An investigation must steelman the opposition.
The simplest portfolio in the world, a low-cost S&P 500 index fund held for 26 years, produced $7.48 per dollar invested. It required zero manager selection, zero due diligence, zero fee negotiation, zero rebalancing, and zero behavioural agony during lost decades of trend-following underperformance. Its only behavioural demand was not selling during the GFC. An investor who chose that path in January 2000 and did nothing would have outperformed roughly four in five of the trend-following programs in this database on terminal wealth. That is the real benchmark. Not because it is geometrically optimal, it is not, but because it is achievable by anyone with the discipline to hold through a drawdown.
Berkshire Hathaway produced $14.11 per dollar. This series has highlighted that seventeen of the analyzed programs beat Berkshire on MAR ratio. That comparison is accurate, but the reader should understand its construction. MAR is the metric this series argues is most relevant for geometric wealth, and it is the metric on which trend following looks strongest relative to Berkshire. On terminal wealth, only Mulvaney materially outcompounded Berkshire. On raw CAGR, only a handful came close. The geometric lens is the right lens, for the reasons these episodes have argued, but intellectual honesty requires acknowledging that the comparison metric was selected because it illuminates the story the evidence supports. A different lens would tell a different story.
The response to this counter-argument is not to dismiss it. It is to reframe the question. The relevant comparison is not trend following versus equities over any particular window. It is the geometric properties of the combined portfolio versus either component alone. The 60/40 S&P/TF blend produced $8.21 per dollar, more than the S&P 500’s $7.48, with a maximum drawdown less than half as deep. The blend beat the benchmark on the benchmark’s own terms, terminal wealth, while dramatically reducing the drawdown that makes holding through crises so difficult. That synergy is mathematical, not narrative.
But synergy on paper is not synergy in practice. The investor must select programs, pay fees, endure years of underperformance during bull markets, explain the allocation to family or committees, and resist the urge to abandon the strategy when it appears to be costing money. The geometric case is clear. Whether the investor can execute it is a question the mathematics cannot answer. The evidence says the geometric portfolio is superior. The experience of holding it is another matter entirely, and this series respects the distance between the two.
The Extended Record: 39 Years, Six Crises
The analysis so far has operated within the window of January 2000 to January 2026. This is a robust dataset, but the sceptic can still object that 26 years may not be enough, and that the geometric properties might be specific to a particular era. To address this, we extend the analysis backward using two trend followers whose records reach the mid-1980s, alongside the S&P 500 Total Return over the same period.
DUNN Capital Management’s World Monetary and Agriculture (WMA) program and EMC Capital Advisors’ Classic program are both systematic trend followers operating across diversified futures markets, and their implementations differ markedly: DUNN runs at roughly 31% annualised volatility with a focused market selection, EMC at higher volatility with a different construction and higher concentration. They are independent firms, independent systems, and independent track records. Over the 39-year common window with the S&P data, January 1987 to December 2025, both outcompounded the S&P 500.
Table 2: Extended record, January 1987 to December 2025 (468 months), net of fees. Both trend followers outcompound the S&P 500 over nearly four decades, and both 60/40 blends exceed the S&P on CAGR, MAR, and terminal wealth while roughly halving the drawdown. DUNN’s maximum drawdown of 60.3% is the same figure carried in Episode 10.
EMC produced 1.8 times more terminal wealth than the S&P 500; DUNN produced 2.1 times more, both net of fees. The geometric fingerprint is present in both. DUNN’s skew of +0.40 and EMC’s extraordinary skew of +3.99 confirm the positive asymmetry documented across the database, and DUNN’s S&P correlation of -0.08 confirms the negative equity correlation. These are not properties that appeared after January 2000. They have persisted across four decades.
The blended portfolios reinforce the Episode 9 finding. The 60/40 S&P/DUNN blend produced $149.20, more than either component, with a maximum drawdown of just 26.1% and a MAR of 0.524. The 60/40 S&P/EMC blend produced $143.19 with a MAR of 0.556. In both cases the blend exceeds the S&P 500 on CAGR, MAR, and terminal wealth while substantially reducing the maximum drawdown. The geometric synergy that Episode 9 documented over 26 years holds over 39. But the most powerful evidence in the extended data is the crisis-alpha record.
Table 3: Cumulative returns through six equity crises across 39 years, S&P 500 versus DUNN WMA. Six crises, six positive outcomes for trend following. All returns net of fees.
Six crises. Six for six. The 1987 crash, the 1990 recession, the LTCM and Russia crisis, the dot-com bust, the global financial crisis, and the 2022 inflation shock. Across 39 years, every significant equity drawdown produced positive returns for DUNN WMA. The crisis-alpha mechanism did not begin in January 2000. It has been operating for as long as systematic trend following has existed in its modern form.
Figure 3: 39 years, six crises, two independent trend followers, January 1987 to December 2025. EMC Classic ($112) and DUNN WMA ($130) both outcompound the S&P 500 ($63). Shaded regions mark crisis periods; dashed lines show the 60/40 blends. The crisis-alpha box records DUNN’s performance during every significant equity drawdown: six for six. All returns net of fees.
Two observations strengthen the extended evidence. First, these are very different implementations. DUNN runs at high volatility with a 60.3% maximum drawdown; EMC runs at even higher volatility with extraordinary positive skew. Yet both produce the same structural properties: positive skew, crisis alpha, and long-term compounding that exceeds the equity benchmark. The convergence documented across the NilssonHedge database now extends across time as well as across programs. Second, the 39-year window spans six different types of crisis: the 1987 flash crash (an instantaneous shock), the 1990 recession (a slow deterioration), the 1998 LTCM crisis (a leverage and liquidity event), the dot-com bust (a valuation correction), the GFC (a systemic financial crisis), and the 2022 inflation shock (a monetary-regime change). Six different causes, the same result. The mechanism is not contingent on the type of crisis. It responds to the trends that crises produce, regardless of their cause.
What Convergence Proves
The ensemble evidence, taken together, proves something no individual track record can. It proves that the geometric properties of trend following, the positive skew, the crisis alpha, the negative equity correlation, and the competitive MAR ratios, are properties of the process rather than of any individual program. Mulvaney’s excellence could be explained by luck, by favourable market selection, or by a single brilliant design decision. When roughly seven in eight programs show positive skew, about four in five show negative equity correlation, seventeen beat Berkshire on MAR, and two independent programs demonstrate the same properties over 39 years and six crises, the explanation must be structural. Process determines direction; individual implementation determines magnitude.
This distinction matters for the investor. Selecting an individual trend follower introduces idiosyncratic risk: the risk that this particular implementation produces a below-median outcome. But the structural properties are present across the ensemble. An investor who diversifies across multiple programs, or who accesses trend following through a diversified composite like the TF Index, captures the process properties while diversifying away the program-specific risk. This is why the sub-period and extended evidence are ultimately encouraging rather than damaging. The pattern across sub-periods, across the analyzed programs in the 26-year window, and across two programs in the 39-year window is always the same: underperformance during equity bulls, outperformance during equity crises. Different amplitudes, same direction.
The Capacity Objection
There is a third objection the survivorship and start-date arguments do not cover: capacity. As more capital deploys systematic trend-following strategies, the argument runs, the signal erodes, and each successive decade should produce weaker returns as the trade becomes more crowded. If this objection holds, the historical record is the best the process will ever look, and the future is decay.
This is the most sophisticated version of the efficiency argument, and it deserves the most direct empirical answer available. A separate research program tested it across 68 futures markets over four decades by dividing the sample into four ten-year periods, measuring the oscillation amplitude of the rolling autocorrelation in each, and testing whether amplitude declines monotonically as efficiency theory predicts. The result did not support decay: the amplitude in the most recent decade was statistically indistinguishable from the first, with the feedback structure retaining the large majority of its original magnitude after four decades of exponential growth in systematic capital. More striking, several asset classes showed higher oscillation amplitude in the most recent decade than the first, including the markets where the most systematic capital has been deployed. The most parsimonious explanation is that the feedback structure is not an exploitable anomaly but a structural feature of how a permanently coupled system of agents with heterogeneous horizons processes information. Capital chasing the signal does not erase it; it becomes part of the signal.
Episode 10 identified two archetypes: the concentrated right-tail exploiter, represented by Mulvaney, and the geometric optimizer, represented by broader-universe programs like Salus Alpha and the Winton programs. The ensemble evidence resolves which archetype is more appropriate as a portfolio component. Mulvaney’s $75.50 is extraordinary, but it is produced by concentrated pyramiding, high leverage, and a 60.9% maximum drawdown that is not replicable at the portfolio level for most investors or institutions. The geometric optimizers, by contrast, produce the structural properties the ensemble confirms across all the analyzed programs, positive skew, crisis alpha, and negative equity correlation, while operating at drawdowns manageable enough to hold through the patience cost of a lost decade. The archetype that serves the investor constructing a diversified geometric portfolio is not the one with the highest terminal value. It is the one whose geometric properties are both structural and endurable.
The Running Ledger
The running comparison now incorporates the ensemble evidence. The median profile represents the typical trend follower, not the best or the worst.
The median-program row anchors the ensemble finding. Even the typical trend follower produces positive skew and negative equity correlation, with a MAR of 0.201 against the S&P’s 0.157. The extended record adds a further dimension: these properties have persisted for 39 years and across every type of financial crisis the modern era has produced.
The Bridge
The archetype question from Episode 10 is also resolved. Mulvaney’s concentrated right-tail exploitation produces the highest terminal value in the database. The geometric optimizers, operating with broader universes and shallower drawdowns, produce the structural properties confirmed across the entire ensemble in forms that are endurable through a lost decade. For portfolio-construction purposes, the relevant archetype is the one whose properties can be held. The ensemble confirms both archetypes work; the evidence identifies which is more suitable for the long-term geometric portfolio the series has been building.
But costs have not yet been fully confronted. The 2010 to 2019 decade is the most dramatic cost, but it is not the only one. The low win rate, the months of treading water, the time spent below the high-water mark: these are the daily texture of the patience cost that the annual numbers obscure. Episode 12 will quantify these costs. It will show what geometric wealth demands from the investor who seeks to earn it.
Data and Sources
All performance data is drawn from the NilssonHedge CTA database as captured in the TTU Trend Following performance file. The database tracks well over a hundred trend-following programs; this series assesses only those with a sufficient track record. The primary analysis universe is the TTU TF Index, comprising 47 active, systematic, globally diversified programs that each meet a minimum 15-year unbroken track record to the current reporting month. The per-program statistical exhibits (Figure 1 and the median-program row) use the stricter subset of 38 programs with records of at least 20 years, the same horizon-consistent subset used in Episode 10. Convergence proportions are stable across these roster definitions. All returns are net of management and performance fees.
The 26-year analysis window runs from January 2000 to January 2026 (313 months). Sub-period figures (Figure 2, Table 1) are annualised CAGRs computed over the stated month ranges from the S&P 500 Total Return and the TTU TF Index series on the Data sheet. The extended record (Figure 3, Tables 2 and 3) uses the DUNN WMA and EMC Classic monthly series from the CrossTab Dump sheet over the common 39-year window with the S&P data, January 1987 to December 2025 (468 months). Crisis figures are cumulative returns over the stated peak-to-trough month ranges.
CAGR is computed as (terminal)^(12/months) − 1. Volatility is the annualised standard deviation of monthly returns. Maximum drawdown is the largest peak-to-trough decline in the cumulative return series. MAR = CAGR / |MaxDD|. Skewness is the skewness of monthly returns. Blends are rebalanced monthly at the stated weights. All headline figures in this episode were verified directly against the source monthly returns. The capacity-objection statistics derive from a separate study of 68 futures markets and are not reproducible from the performance database; that study should be cited directly when this episode is published.
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.