Why position sizing in pure trend following is determinate, once the domain of the question is properly drawn.
Carlo Zarattini published a thoughtful piece earlier this week on position sizing in trend following. He compared three approaches against the same signal engine and showed that each wins under a different measurement lens. The conclusion he drew is what he called the observer effect: the apparent best sizing approach depends on which metric you evaluate it through. No universal answer exists. Different practitioners, different constraints, different winners.
The piece is careful, methodologically sound, and worth reading on its own terms. You can find it here.
This response sits alongside his. The observer effect is real within the domain his analysis addresses. The question is whether the domain itself captures the full space of sizing decisions. Our view is that it does not, and that when the domain expands to include the architectural choices that pure trend programmes actually make, the observer effect resolves into determinacy. This piece sketches how. It rests on a broader empirical programme we have published elsewhere, and what follows summarises only the implications for sizing.
Naming the objective
Before proceeding, one point needs to be made explicit.
The scope of this piece is the 100% trend programme, the standalone vehicle whose entire purpose is to compound capital through trend exposure. This is the vehicle Zarattini tested, and the vehicle our research addresses. Trend following is also used in other ways: as a diversifying sleeve within a traditional portfolio, as an explicit crisis alpha allocation seeking tail-hedge characteristics, as one component of a multi-strategy framework. These applications have their own legitimate objectives and their own evaluation criteria. The argument that follows is about the standalone trend programme, where the question of what the vehicle is for has a single answer.
For a pure trend programme considered as a capital-compounding machine, the primary objective is wealth creation. By wealth creation we do not mean reckless maximisation of terminal wealth. We mean durable real compounding over repeated market cycles, with losses kept small enough to preserve participation in the next outlier. Stability, allocator comfort, institutional optics, and smooth monthly statements are better understood as constraints layered on top of that primary objective rather than as alternative objectives in their own right. A programme that fails to compound capital over time has failed regardless of how comfortable its equity curve looked along the way. A programme that creates significant wealth across a full cycle has succeeded regardless of how uncomfortable the ride was.
This framing matters because the sizing debate tends to proceed as if different practitioners running the same kind of programme might be aiming at fundamentally different things. Some of that variation is real. But most of it, on inspection, turns out to be constraint structure rather than objective variation. Different practitioners operate under different agency pressures, and those pressures shape their sizing choices. The pressures are real. But they are constraints on how a shared objective gets pursued, not evidence of a different objective.
With the primary objective named, the observer effect can be re-examined.
What Zarattini showed
He tested three sizing rules applied to the same breakout strategy across forty futures markets from 1978 to 2026.
Volatility Targeting (VT) adjusts position size continuously to keep daily portfolio volatility constant. When markets become more volatile, positions shrink. When markets become calmer, positions grow. The goal is a smooth equity curve with a stable risk profile.
Volatility Parity (VP) sets position size at trade entry based on volatility at that moment and leaves it alone for the life of the trade. No adjustment mid-trade, no scaling up or down. The trade is sized once and allowed to run.
Volatility Parity with Pyramiding (VPP) starts like VP but adds to winning positions as they move into profit. A trade that keeps working gets more capital committed to it. The approach is capped at four times the initial size to prevent runaway leverage.
When evaluated through Sharpe ratio, VT won. When evaluated through Profit Factor, VPP won. Same data, same trades, different rankings. This ranking reversal is the observer effect Zarattini identified.
The error beneath the metric
Every sizing rule expresses an implicit trade-off between two kinds of error.
A Type I error is entering a trade that turns out to be a false signal and taking a small loss. A Type II error is failing to capture a trade that turns out to be a major outlier. Both errors happen in every trend programme. The question is which one you are willing to pay more for avoiding.
VT minimises Type I errors. By scaling down when volatility expands (which is precisely when trends tend to be forming) it accepts smaller intra-trade drawdowns but also captures less of the outliers when they arrive. VPP minimises Type II errors. By adding to winners, it risks giving back more on reversals but captures more of the outlier when the trend persists. VP sits between them.
The metrics that rank these approaches differently are sensitive to different error types. Sharpe punishes volatility symmetrically, which means it rewards Type I error avoidance and penalises the volatility that accompanies Type II error avoidance equally. Profit Factor rewards capturing the outsized winners that only concentrated exposure produces, which means it rewards Type II error avoidance. The ranking reversal between Sharpe and Profit Factor is the same trade-off viewed through lenses that score the two errors differently.
Zarattini’s framing treats this as preference. The retail investor who cares about absolute returns evaluates at the trade level and chooses VPP. The institutional manager who needs to report stable volatility evaluates through Sharpe and chooses VT. Both are described as making rational choices under their respective utility functions.
But if wealth creation is the primary objective across both practitioners, the error trade-off is not a matter of preference. It has a correct direction, given the structural properties of the market the programme is attempting to compound through.
What the market actually is
The research programme we have published over the past two years examined the statistical structure of futures markets. The short version for readers who have not followed it:
Across four decades and nearly seventy futures markets, the most cited statistic in quantitative finance (the autocorrelation of daily returns) averages to zero. This zero underwrites the random walk, the efficient market hypothesis, and most of modern portfolio theory.
But the zero conceals what produces it. When the data is separated by market state rather than pooled, the zero turns out to be the weighted average of two opposing forces. In some regimes, markets trend and autocorrelation is positive. In other regimes, markets oscillate and autocorrelation is negative. Both forces are always present. The zero appears only when you average across the regime shifts.
Three properties of returns follow from this structure, all documented empirically across every market class we tested. Distributions are fat-tailed at every time scale. Path matters in a way that breaks the assumption of ergodicity, the idea that average outcomes across many imaginary traders describe what any single trader will experience. Different asset classes trend for different durations relative to how long they mean-revert, with the ratio varying by a factor of four across the universe.
These are not theoretical claims. They are measured properties. And they determine how wealth actually compounds through this environment.
Why Type II error dominates wealth creation
In fat-tailed non-ergodic markets, long-run CAGR is driven by a small number of exceptional trades. A typical trend programme generates its lifetime compound return from perhaps ten to fifteen genuinely significant outlier trades across a career. Hundreds of small losses are absorbed along the way. The distribution is not driven by the central tendency. It is driven by the tail.
This means Type II errors are more damaging to long-run wealth creation than Type I errors, in ways that standard risk metrics are poorly equipped to capture. Missing a single career-defining outlier can compromise lifetime wealth creation in ways that ordinary sequences of small, controlled losses usually do not. The caveat is important: Type I losses must remain survivable. A sequence of small losses that breaches the capital base, forces leverage reduction, or triggers behavioural abandonment is not survivable, and no Type II error reasoning applies in that regime. But once survivability is protected, the dominant error in a fat-tailed trend programme is failing to be there when the rare outlier arrives. The programme pays hundreds of small Type I costs so it can remain present for the rare Type II avoidance that justifies the entire method.
VT, in this environment, makes the trade-off in the wrong direction for wealth creation. It reduces exposure at precisely the moments when outlier moves are developing, because outlier moves are accompanied by expanding volatility. It pays for smoother intermediate performance with worse long-run compounding. Sharpe rewards it for this trade because Sharpe measures smoothness, not compounding. The programme that wins on Sharpe may still lose on the primary objective of long-run compounding.
The institutional practitioner who chooses VT is not optimising wealth creation under a different utility function. They are optimising their allocator retention problem under a capital-stability constraint, which is a different objective entirely. The constraint exists because their allocators evaluate through volatility-sensitive metrics, and the manager must deliver stable-looking performance to keep the programme funded. This is a rational response to a real business problem. But it is a constraint on how wealth creation gets pursued, not an alternative to wealth creation.
The retail investor who chooses VPP is closer to optimising wealth creation directly because they do not face the allocator relationship. Their preferences align with the underlying capital’s interest because they are the underlying capital. Their rougher intermediate performance is the cost of serving the primary objective rather than a proxy for it.
The observer effect, viewed this way, is not a diversity of valid objectives. It is a diversity of constraint structures producing different sizing choices in practitioners who are all nominally pursuing the same compounding problem. The choices are not equally good at serving wealth creation. They are equally rational responses to the constraint structures each practitioner operates under.
Where the domain ends
This is not a rejection of Zarattini’s result. It is a boundary condition around it.
Zarattini’s comparison isolates the signal-layer sizing decision while holding portfolio architecture constant. That is the right methodology for a clean test of a single variable. Our claim is not that his result is wrong within that test. It is that real trend programmes do not live at the signal layer alone. Once ensemble breadth, realised-capital anchoring, and breadth-based scaling are allowed to vary, the concentration penalty attached to single-signal pyramiding changes materially. The domain has changed, and with it the ranking logic.
Sizing decisions exist at three layers. The signal level (which entry and exit rule, sized how), the ensemble level (how many structurally different rules engage each market), and the scaling level (how the programme compounds as capital grows). Zarattini’s framework addresses the first. The other two are where the most consequential sizing choices actually live.
At the ensemble level, the question is how many structurally different rules are engaging each market. A breakout rule, a retracement rule, a channel rule, and a volatility-expansion rule all look at the same price series but engage it through different mechanical pathways. Because their entry conditions are mutually exclusive by construction, they cannot all fire simultaneously on the same bar in the same direction. When a trend forms, whichever rule most closely matches the shape of that specific trend captures it. When the market is choppy, none are engaged. The ensemble does not just diversify noise. It diversifies the failure modes of the individual signals. This reduces Type II error without proportionally increasing Type I error, because each rule engages only the trends whose shape it recognises.
At the scaling level, the question is how the programme compounds as capital grows. The conventional view is that position sizes scale proportionally with equity. This increases both error types symmetrically: wins scale up, and so do losses. The alternative is to hold position sizes relatively constant and deploy new capital into new markets and new systems. Each addition brings an independent error profile rather than amplifying the existing ones. Across decades, compounding through breadth preserves the favourable error asymmetry in a way that compounding through size cannot.
When the sizing decision is analysed across all three layers rather than at the signal layer alone, the ranking reversal that produces the observer effect begins to dissolve. Single-signal pyramiding into a single instrument produces portfolio-level concentration that ensemble construction distributes across staggered entries by different rules. The effective convexity is similar. The portfolio-level risk profile is meaningfully different. The Sharpe penalty that VPP incurs in the single-signal case is partly a concentration artefact. In the expanded domain, the metrics that rank the approaches differently within the narrow domain begin to converge, because the choice that serves wealth creation under one metric starts to serve it under the others as concentration effects dissipate.
The observer effect is a feature of the narrow domain. In the expanded domain, determinacy returns.
The revised reading
We wrote about VPP in late 2024 and concluded that among Zarattini’s three approaches, its philosophy was closest to ours. We stand by the spirit of that response. VPP is the sizing rule, within his framework, that most directly serves wealth creation.
What has clarified since is that the debate between VT, VP, and VPP is a debate within a domain that the research has now expanded beyond. The sizing choices that matter most for wealth creation are not which single-signal rule to apply. They are how to construct ensembles that diversify failure modes rather than just assets. How to anchor position sizing to realised capital rather than to paper gains. How to scale through breadth rather than through size. How to remain deployed during the trending phase of the market without compromising the capital base during the transitions.
These are determined by the structure of the market combined with the objective of wealth creation. Architectures that aim at that target succeed. Architectures that aim elsewhere accumulate a drag that compounds across decades.
What actually matters
The question is not which sizing rule each practitioner will rationally choose under their own constraints within the narrow domain. The question is which sizing architecture best serves wealth creation when the full domain of portfolio construction is allowed to vary.
Given the structural properties of the markets within which pure trend programmes operate, that question has an answer. The practitioner whose constraints force them toward the narrow-domain answer is not choosing a different objective. They are paying a cost measured in wealth not created, in exchange for stability that serves their constraint structure rather than the capital they manage. This is a rational choice given those constraints. It is also a cost. Naming the cost is the first step toward deciding whether it is worth paying.
The two-phase research programme referenced here, The Fractals of Finance research series, documents the structural properties summarised in this piece. The complete architecture is developed in The Fractals of Finance: Determinism, Adaptation and the Geometry of Markets, and its practical implementation in The Aussie Turtles Trend Following Guide: A Field Manual for Hunting Outliers, co-authored with Adam Havryliv.
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.