Why Every Model Describes a World You Are Not In
"The menu is not the meal" - Alan Watts
There is a moment every systematic trader knows. You have spent months building it. The backtest is clean. The equity curve climbs with the steady inevitability of a well-constructed argument. Drawdowns are contained. Returns compound. The logic is sound, the data is honest, and the results confirm what you suspected: the signal is real.
Then you turn it on.
And something shifts. Not dramatically, not immediately, but with the quiet persistence of a tide changing direction. The fills are a fraction worse. The entries cluster differently. The drawdowns arrive sooner, last longer, bite deeper. Nothing in the code has changed. Nothing in the logic has failed. But the strategy that gleamed under the laboratory light now stumbles in the open air, and you cannot quite say why.
The standard explanations arrive on cue. Slippage. Transaction costs. Overfitting. Curve-fitting. Data-snooping bias. Each of these is real, each is worth addressing, and none of them captures the deeper problem. Because the gap between your backtest and your live performance is not a technical failure. It is a structural property of the system you are trading. And until you understand why, you will keep building maps of a territory that no longer exists the moment you step onto it.
The Photograph of an Empty Room
Consider what a backtest actually is. Strip away the statistical sophistication, the Monte Carlo simulations, the walk-forward optimisations. At its core, a backtest is a description of what would have happened if you had been there. But you were not there. That is the entire point. The data records a market that unfolded without your participation. Every price, every volume bar, every volatility cluster reflects the actions of agents who did not include you.
This is not a subtle distinction. It is the distinction.
A backtest is a photograph of a room you were not in. It captures the furniture, the light, the arrangement of objects with perfect fidelity. But the photograph cannot tell you what the room looks like when you walk into it, because your presence changes the room. You displace air. You cast shadows. You move things. The room with you in it is a different room from the room without you, and no amount of photographic precision bridges that gap.
In markets, this displacement is literal. Your orders consume liquidity that was available in the historical record. Your entries add buying pressure at points where, historically, that pressure did not exist. Your stop losses create clustered exit points that generate selling cascades the backtest never encountered. Your very presence as a capital-deploying agent alters the microstructure of the market at exactly the moments your strategy is most active.
And this is just you alone. Now multiply it by every other participant who has also studied the data, also found patterns, also deployed capital on the basis of historical regularities. The market you enter is not the market your backtest described. It is a market that includes the collective weight of everyone who studied that same history and acted upon it.
The Observer Problem, Writ Large
Physics encountered this puzzle a century ago. At the quantum scale, the act of measurement disturbs the system being measured. You cannot observe a particle’s position without altering its momentum. The instrument and the object are coupled. Observation is participation.
Markets are not quantum systems, but the structural parallel is exact. The act of measuring a market pattern (backtesting it, validating it, deploying capital on the basis of it) changes the pattern. Not because the market is perverse or conspiratorial, but because measurement in this context means participation, and participation means influence.
The efficient market hypothesis understood half of this. It recognised that information gets priced in, that patterns disappear when discovered. But it framed the problem as one of competition: enough smart money chases the anomaly until it vanishes. The reality is stranger and more fundamental. The anomaly does not simply vanish. It transforms. It becomes something different under the weight of collective engagement, something the original measurement could not have captured because the original measurement was taken in the anomaly’s absence.
This is the observer problem writ large. You are not a scientist peering through a microscope at an inert specimen. You are a microbe in the culture, and the culture shifts with every move you make.
The Cartographer’s Paradox
Imagine a cartographer tasked with mapping a river. She sets up her instruments on the bank, records the depth, the current, the contours of the riverbed. The map is exquisite: precise, detailed, faithful to what she observed. She rolls it up, hands it to a navigator, and says, “This will guide you.”
The navigator enters the river. His boat displaces water. The current shifts around the hull. The depth changes where his weight presses down. The eddies form differently because there is now an obstruction that was not there when the map was drawn. The map is not wrong. It was perfectly accurate at the moment it was made. But the moment it was made was a moment before the navigator entered, and the river with the navigator in it is not the river the cartographer measured.
Now scale this up. Not one navigator but thousands, all entering the river at the same points (because they all have the same map), all displacing water at the same locations, all creating eddies that compound and interfere. The river becomes something the cartographer would not recognise. Not because her instruments were poor, but because the act of using the map changed the territory the map described.
This is what happens when a backtested strategy goes live. The map was drawn from the bank. You are now swimming in the current. And the current includes you.
Model Degradation Is Not Decay
The trading industry has a term for this: model degradation. It implies entropy, a slow wearing away, as though strategies are subject to some natural force of erosion. The metaphor is comforting because it suggests the problem is gradual and manageable. Recalibrate. Reoptimise. Add new data. Rebuild.
But model degradation is not decay. It is displacement. The model is not slowly becoming less accurate about the same world. The world itself has moved. The model continues to describe, with perfect accuracy, a market that no longer exists. The map has not faded. The territory has shifted under it.
This distinction matters because it changes what you do about it. If degradation is decay, the solution is maintenance: tune, update, refresh. If degradation is displacement, the solution is something more profound. You must accept that your model will always describe a world that includes your absence. Every optimisation captures a state that will be altered by the act of deploying the optimised strategy. You are not chasing a moving target. You are chasing a target that moves because you are chasing it.
The implications cascade. Your walk-forward analysis, designed to capture changing conditions, still measures a market that did not include the strategy being walked forward. Your out-of-sample test validates on data where your capital was not present. Your stress tests simulate scenarios that would unfold differently with your participation factored in. Every layer of rigour you add still photographs the room from outside. The room with you in it remains unseen.
The Absent Observer
There is a deeper philosophical point here, one that extends beyond trading but finds its sharpest expression in markets.
Western science was built on the premise of the detached observer. The scientist stands apart from the experiment, records data, formulates laws. The observer is a transparent window through which reality is viewed without distortion. This premise has been extraordinarily productive. It gave us physics, chemistry, biology, engineering. It works beautifully when the observer is genuinely separable from the observed.
Markets are not such a system. In a complex adaptive system composed of interacting agents, there is no position from which to observe without participating. Every vantage point is also a position within the system. Every act of observation is also an act of influence. The detached observer is not merely impractical in markets. The role does not exist. It is a structural impossibility.
This is what distinguishes the embeddedness problem from ordinary model error. Model error says: your map is inaccurate. Embeddedness says: no map can be accurate, because the act of making the map (and using it) changes the territory. The error is not in the cartography. It is in the assumption that cartography and navigation can be separated.
Backtests assume the transparent observer. They assume you can measure a system without being part of it, that the measurement captures something stable enough to act upon. In a mechanical system, this assumption holds. In a complex adaptive system, it does not. The measurement is always of a state that your measurement will alter.
What the Map Cannot Show
So what does the embedded agent do? If no model can describe the world as it will be once the model is deployed, does modelling become pointless? Is the backtest worthless?
No. But its value is different from what most practitioners assume.
A backtest does not tell you what will happen. It tells you what kind of thing you are building. It reveals the character of a strategy: its relationship to momentum or mean-reversion, its sensitivity to volatility regimes, its behaviour under stress, its structural dependence on certain market conditions. These characteristics are real and durable, even if the specific returns are not. The map may not show you the river as you will find it, but it tells you something about rivers. It tells you about current, about depth, about the kinds of obstacles that form and dissolve.
The embedded agent reads a backtest the way a sailor reads a weather chart: not as a prediction of what will happen, but as an orientation to what might happen and what the vessel is built to handle. The specific forecast will be wrong. The general preparation will not be wasted.
This requires a fundamental shift in how we hold our models. Not as descriptions of the future, but as descriptions of a relationship. The backtest shows the strategy’s relationship to certain market conditions, conditions that existed before the strategy was present. The live performance reveals a different relationship: the strategy’s interaction with a market that now includes it. Both relationships are real. Neither is the whole truth.
The practitioner who understands this holds models lightly. Not dismissively, not cynically, but with the kind of careful provisionality that comes from knowing the map was drawn of a place you have not yet entered. You use the map. You do not worship it. You update it as you walk, knowing the territory will keep shifting because you are in it.
The First Step
This is the foundational insight of the entire series that follows. You are embedded. Your models are not. The world they describe is a world that included your absence, and the world you inhabit is one that includes your presence. These are not the same world, and no amount of technical sophistication makes them the same.
Every consequence explored in the essays ahead flows from this single structural fact. Your strategy is not separate from the ecology it trades within; it is a species in that ecology, reshaping it by existing. The market computes at a level no participant can access, because the computation includes the participant. Your beliefs about the system become forces within the system, building structure from expectation. The boundary between you and the market is a useful fiction that dissolves under examination. Understanding the system does not grant you power over it, because understanding is itself a form of participation.
All of this begins here, with the map that walks.
You built a model of the market. It was good work, honest work, careful work. But the model described a world that did not include you. Now you have entered that world, and it is not the world the model described. Not because the model was wrong, but because the model was complete in a way that excluded the one variable it could never capture: you.
The map walks on the territory it describes, and the territory shifts with every step.
This is not a problem to be solved. It is a condition to be understood, respected, and practised within. The embedded agent does not seek a better map. The embedded agent learns to walk.
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 and Complex Adaptive Markets. The forthcoming Carved by Impossibility completes the trilogy.
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.