The Ensemble Approach
No single rule sees the whole market. That is not a problem to solve. It is the reason we use more than one.
Previous: System Anatomy 4: Portfolio Construction | Next: System Anatomy 6: Rollover Mechanics
A radio telescope is an extraordinary instrument.
Point one at the sky and it can detect things no human eye will ever see.
But it still sees only what it is capable of seeing.
Change the instrument and you change the observation.
Change the frequency and another part of the universe appears.
This is why astronomers build arrays.
Multiple instruments observe the same sky from different positions. None possesses the complete picture. None is designated the one true telescope.
Their value comes from their differences.
Trend following has much the same problem.
A trading system is a way of observing price.
Give it a particular definition of directional movement, a particular sensitivity and a particular way of entering and leaving, and it will detect some trends beautifully.
Others it will encounter late.
Some it will leave early.
Some it will miss altogether.
That does not necessarily mean the system is defective.
It means it has a point of view.
The mistake is believing that one point of view can somehow become universal if we optimise it hard enough.
It can’t.
So we do something else.
We build the array.
Your Trend Is Not My Trend
We speak about “the trend” as though it were an objective object sitting inside the price series waiting to be discovered.
It isn’t.
Show the same market to two trend-following systems and they can legitimately disagree.
One may be long.
Another may still be waiting.
A third may already have exited.
Which one is right?
That question assumes there is a single correct trend hidden underneath the disagreement.
There may not be.
A trend exists relative to the mechanism observing it.
A fast system encounters one aspect of the path.
A slower system encounters another.
A mechanism responding to price escaping a range sees something different from one requiring a more developed directional structure.
Another may become interested only after an established move has retraced.
Same market.
Same price history.
Different observations.
This is not noise we need to eliminate from the programme.
It is diversity we can use.
Episode 4 distributed our search across markets because we do not know where the next outlier will occur.
Episode 5 distributes the search again.
This time across ways of seeing.
Why Not Find the Best System?
There is an obvious alternative.
Test a large number of systems.
Find the best one.
Use that.
It is wonderfully appealing.
It is also where trouble begins.
What does “best” mean?
Best over which markets?
Which years?
Which regimes?
Which sequence of trends?
Which assumptions about execution?
Which parameter values?
The moment we identify the historically best system, we have selected something partly because of the particular path history happened to take.
Perhaps its superiority reflects something durable.
Perhaps it reflects a fortunate alignment between its structural sensitivity and the trends contained in our sample.
We do not know.
And the future is under no obligation to reproduce the distribution of opportunities that made yesterday’s winner look exceptional.
This is the same problem we have encountered throughout the series.
We want certainty where none is available.
So rather than asking one system to be best, I prefer a different question:
What does this system contribute that the others do not?
That changes everything.
A system no longer has to win the beauty contest.
It has to earn its place in the architecture.
Difference Is the Point
An ensemble of systems that all behave almost identically is not much of an ensemble.
Give five systems slightly different parameter values and they may generate five slightly different equity curves.
But if they enter at roughly the same time, exit at roughly the same time and respond to roughly the same features of price, we have not created much structural diversity.
We have created variations on a theme.
That may still be useful.
Small variations can matter.
But the deeper opportunity comes from combining mechanisms that encounter trend differently.
Some respond quickly.
Others require more evidence from the developing path.
Some tolerate substantial movement around a trend.
Others react sooner when the path changes.
Some are naturally suited to one temporal expression of trend and others to another.
The objective is not disagreement for its own sake.
It is useful disagreement.
If every instrument in an astronomical array were pointed in precisely the same direction, measuring precisely the same thing in precisely the same way, adding more instruments would give us more observations but little additional perspective.
The same is true here.
“The ensemble becomes useful when its components disagree for structural reasons, not because we deliberately made them different.”
The Systems Do Not Vote
This distinction is crucial.
Suppose several systems are operating on the same market.
One is long.
Another has no position.
A third enters long two weeks later.
What does that tell us?
It does not mean the market has accumulated three units of bullish evidence.
The systems are not voting.
We do not count the number of bullish signals and conclude that the forecast has become stronger.
Each system is doing its own job according to its own rules.
One has observed the condition it was built to detect.
Another has not.
That disagreement requires no resolution.
There is no committee sitting above them asking which system has the better argument.
They simply operate.
This is important because it prevents the ensemble from becoming another forecasting machine in disguise.
We are not combining opinions to estimate what happens next.
We are allowing different mechanisms to respond independently to the same uncertain path.
The aggregate exposure emerges from those independent responses.
That is very different.
Exposure Can Emerge Rather Than Be Decided
Now something interesting happens.
Imagine a trend begins.
One system responds relatively early.
It establishes a position.
The move continues.
Later, another mechanism encounters the same developing trend according to its own definition.
It enters too.
Still later, another may join.
Nobody decided to “size up.”
Nobody announced that confidence in the trend had increased.
There was no master rule saying:
This is becoming a big one. Add exposure.
The exposure emerged because different mechanisms independently encountered the same developing path at different points.
The same thing can happen in reverse.
As the trend eventually deteriorates, one system may leave.
Another remains.
Then another exits.
Aggregate exposure falls progressively because the individual components reach their own conclusions at different times.
Again, there is no master exit.
No one needs to know where the top was.
The ensemble builds and unwinds through the interaction of its parts.
That is an emergent property of the architecture.
And I think it is one of the most beautiful things about systematic trading.
Complex behaviour does not always require a complex rule.
Sometimes it emerges from several simple rules operating together.
The Outlier Reveals the Array
This becomes particularly important when a genuine outlier appears.
At the beginning, it looks like everything else.
Episode 1 dealt with that.
One mechanism detects something and enters.
The trend continues.
Another encounters it.
Then perhaps another.
As the move extends across time and price, it passes through different structural definitions of trend.
The longer and more persistent it becomes, the more opportunities different members of the ensemble may have to participate.
Not necessarily all of them.
And not necessarily at the same time.
That is the point.
The ensemble does not need to identify an outlier as an outlier.
It simply needs multiple ways of encountering the path while it develops.
Eventually the move ends.
Different components may leave at different points.
Some capture more.
Some less.
Some may have missed the trade entirely.
We do not care.
The objective was never for every component to be right.
The objective was for the architecture to have multiple opportunities to be present.
This is the same logic as portfolio breadth, applied in another dimension.
Many places to look. Many ways to see.
A System Should Survive Leaving Home
But diversity alone is not enough.
We can invent endless trading rules that behave differently from one another.
That does not make them robust.
A rule has to demonstrate that there is something underneath its historical result beyond a lucky fit to the environment in which it was created.
This is where I think one of the most important questions in system design appears:
Does the logic travel?
A system developed on one market can look extraordinary.
Fine.
Now take it somewhere else.
Do not redesign it.
Do not lovingly adjust the parameters until the new market agrees with us.
Just expose the logic to a different history.
Then another.
Different markets.
Different sectors.
Different regimes.
Different price paths.
What happens?
I am not asking whether the system performs brilliantly everywhere.
That would be an absurd standard.
Some markets will suit it better than others.
Some periods will be miserable.
That is inevitable.
The question is more fundamental.
Does the behaviour survive contact with environments that had nothing to do with its construction?
That tells me much more than another decimal place of performance on the market where the system was born.
“Robustness is not how beautifully a rule explains its birthplace. It is how well it survives leaving home.”
Universality Before Brilliance
This is why I have always been more interested in broad robustness than local brilliance.
A beautiful backtest can be seductive.
Smooth equity curve.
Wonderful statistics.
Shallow historical drawdown.
Everything seems to fit.
Sometimes too well.
Markets do not reward us for explaining the past beautifully.
They expose us to futures we have never seen.
So the system that interests me is often the less impressive one.
The simple mechanism that works reasonably across many different environments.
The one that occasionally looks clumsy.
The one that has periods where you wonder why it is in the programme at all.
But move it somewhere else and it still behaves recognisably like itself.
That is valuable.
Because it suggests the result may be attached to something more durable than the peculiarities of one historical sequence.
There is an important distinction here.
We are not searching for a universal parameter that is optimal everywhere.
We are searching for logic that remains plausible across difference.
That is a much lower claim.
And a much harder test.
The Seduction of Precision
The alternative is optimisation.
Add another condition.
Tune another parameter.
Remove another historical loss.
Improve the entry.
Refine the exit.
Keep going and eventually the backtest begins to look magnificent.
Every imperfection has an explanation.
Every awkward period has been engineered away.
And every improvement quietly adds another dependency on the historical path we already know.
Precision feels like knowledge.
Sometimes it is merely memory.
This is why I prefer simple components inside a diverse architecture.
Each component can be wrong.
Each can endure periods in which its particular way of seeing trend is poorly suited to the environment.
That is acceptable because no individual component is carrying the programme.
We do not need one magnificent rule.
We need a collection of robustly different ones.
The answer to uncertainty is not necessarily a smarter component. Sometimes it is a better arrangement of simple components.
The Ensemble Is Not a Collection of Winners
This changes how we think about selecting systems.
Imagine two candidates.
The first has spectacular historical performance but behaves almost exactly like something already in the ensemble.
The second has less impressive standalone performance but encounters the market differently.
Which is more valuable?
There is no automatic answer.
But the second deserves serious attention precisely because its contribution cannot be judged from its individual return stream alone.
This is portfolio thinking applied to systems.
A component has two identities.
What it does by itself.
And what it does in combination.
The second can matter more.
A system that looks mediocre in isolation may reduce dependence on a particular path, horizon or manifestation of trend.
Another spectacular system may add almost nothing because we already own its behaviour elsewhere.
So the question is not:
Which systems have the highest historical returns?
It is:
What architecture do these systems create together?
That is a much harder question.
It is also the right one.
Out-of-Sample Means Something Very Simple
Eventually the architecture has to meet something it has never seen.
That is the point.
Research can only use history.
There is no alternative.
But we can at least distinguish between the history involved in developing an idea and observations that played no part in its construction.
The latter matters because it removes one particular advantage from the system.
Familiarity.
No adjustment after seeing the answer.
No parameter change because a particular period looks ugly.
No clever explanation for why this market should be treated differently.
Leave it alone.
See what happens.
Out-of-sample testing is not a magical guarantee against overfitting.
Nothing is.
But it asks a clean question:
What happens when the architecture encounters data that had no opportunity to influence its design?
That is precisely the kind of question a system built for an unknowable future should be willing to face.
Two Axes of Diversification
We can now connect Episode 4 and Episode 5.
Episode 4 spread the programme across markets.
This episode spreads it across ways of responding to those markets.
Neither is enough by itself.
A broad portfolio running one system has many places to search but only one way of seeing what happens there.
A diverse ensemble applied to a narrow portfolio has many ways of seeing but very few places to look.
Put them together and something much richer emerges.
Many markets.
Many structural drivers.
Different temporal scales.
Different sensitivities.
Different entries.
Different exits.
Different paths through the same underlying world.
Not independence.
We dealt with that fiction in Episode 4.
Difference.
Enough difference that the programme is not dependent on one particular market, one particular manifestation of trend, or one particular rule being the correct description of the future.
This is diversification at the level of architecture.
No Single Instrument Sees the Whole Sky
Return to the array.
No astronomer is disappointed because one telescope cannot observe the entire universe.
The limitation is understood from the beginning.
So they build around it.
That is what we are doing.
A trading system is an instrument.
It observes price through a particular structural lens.
Its limitation is not something we need to hide with more and more complexity.
We can acknowledge it.
Then put another instrument beside it.
And another.
Each simple enough to understand.
Each different for a reason.
Each capable of operating independently.
Each tested beyond the environment in which it was conceived.
None knows the future.
None sees the whole sky.
That is why we built the array.
And when the array finally detects something extraordinary, the intelligence does not belong to the instrument that happened to see it first.
It belongs to the architecture that made sure we were looking in more than one way.
Next we go beneath the signal itself.
Because every instrument in the array is observing something we have so far treated as given:
the price series.
In futures, even that is constructed.
And if we construct it one way in research and another way in live trading, the integrity of everything above it begins to unravel.
That is rollover mechanics.
READ DEEPER
→ From Hurricanes to Financial Markets: Harnessing Ensemble Forecasting in Complex Systems
→ Enhancing Trend Following Performance Using System Diversification
→ Diversification for Trend Following Models: The Small Variations Matter
→ Your Trend is Different to My Trend
Previous: System Anatomy 4: Portfolio Construction | Next: System Anatomy 6: Rollover Mechanics
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.