Diversification is not a position count

Five semiconductor stocks are not five independent bets.

They are often one bet wearing five ticker symbols.

That is the central mistake behind position-count diversification.

An investor sees ten companies and assumes the portfolio contains ten separate sources of risk.

The broker confirms ten rows.

The portfolio may contain only three economic ideas.

Lesson 23 introduced one portfolio with two engines.

This lesson asks what both engines eventually place inside the same account.

The answer cannot be judged by ticker count.

It must be judged by what makes the positions rise and fall together.

Company count is the easiest number

Position count is attractive because it is visible.

One stock feels concentrated.

Ten stocks feel diversified.

Twenty stocks feel safer still.

But the number of legal entities does not tell us:

Whether they serve the same customers
Whether they depend on the same commodity
Whether they borrow at similar rates
Whether they share a supply chain
Whether they respond to the same valuation factor
Whether one capital-spending cycle funds all their growth
Whether the same regulation can damage them
Whether they will become correlated during stress

A portfolio can own five excellent companies and still have one dominant risk.

Quality does not cancel concentration.

Different company names do not guarantee different economic drivers.

The five-costume problem

Imagine a portfolio containing:

A semiconductor designer
A chip-testing company
An electronics manufacturer
A data-centre contractor
A connector and sensor supplier

The companies occupy different industry classifications.

Their income statements look different.

Their products are different.

Their immediate customers may be different.

But all five may benefit from the same underlying decision:

Hyperscalers and enterprises continue expanding AI and data-centre capital expenditure.

When that spending accelerates, the five positions can rise together.

When investors question the return on that spending, all five can reprice together.

The portfolio owns five businesses.

It may still own one capital-cycle thesis.

That does not make the positions invalid.

It means their aggregate risk must be measured as a cluster.

July provided the live case

My July deployment included APH, FIX, JBL, KEYS and TER.

At first glance, the names appear spread across several industries:

Company	Immediate business	Shared exposure
Amphenol	Connectors and sensors	Electronics and data infrastructure
Comfort Systems	Mechanical and electrical contracting	Data-centre and industrial construction
Jabil	Electronics manufacturing	Hardware production and capital cycles
Keysight	Test and measurement	Electronics, communications and semiconductor development
Teradyne	Semiconductor and automation testing	Chip production and automation investment

This is not one technology stock copied five times.

The businesses have real differences.

They also contain meaningful overlap.

All five can be helped by expanding infrastructure, electronics and automation investment.

All five can be hurt if the market rapidly reduces the value it assigns to that investment cycle.

The July recap described the deployment as diversified across industries.

That was directionally true.

The more precise statement is:

The deployment reduced dependence on mega-cap software and advertising, but it did not eliminate exposure to the AI infrastructure and electronics cycle.

Precision improves the portfolio.

It does not weaken the original argument.

Diversification has layers

A serious review should examine several layers of overlap.

1. Company risk

This is the risk unique to one business.

Examples include:

Management failure
Product defects
Fraud
Customer loss
Litigation
A failed acquisition
An earnings miss
An unexpected capital raise

Holding more than one company can reduce this risk.

This is the part position count measures reasonably well.

2. Industry risk

Companies in the same industry share pricing, competition and regulation.

Five banks may all respond to credit quality and the yield curve.

Five airlines may all respond to fuel, capacity and travel demand.

Five semiconductor companies may all respond to inventory, manufacturing capacity and AI expectations.

Different management teams do not remove the shared industry cycle.

3. Factor risk

Companies from different sectors can still share the same factor.

Common factors include:

Growth versus value
High versus low duration
Large versus small capitalisation
High versus low quality
Momentum
Interest-rate sensitivity
Commodity sensitivity
Cyclicality

A software company and a biotechnology company may have little operating overlap.

Both can behave like long-duration assets when yields rise.

4. Revenue-driver risk

This layer asks where demand ultimately comes from.

A contractor, component supplier, testing company and chip designer may all depend on data-centre construction.

Different invoices can trace back to the same capital budget.

Revenue-driver analysis often reveals overlap that sector labels miss.

5. Customer risk

Several companies may sell into the same small group of customers.

The portfolio may appear diversified while depending on the spending decisions of Microsoft, Amazon, Alphabet and Meta.

If the hyperscalers cut capital expenditure, multiple holdings can feel the same decision through different supply-chain levels.

6. Supply-chain risk

Different businesses can depend on the same manufacturer, material, geography or logistics route.

Examples include:

TSMC capacity
CoWoS packaging
Copper
Rare earths
The Strait of Hormuz
One contract manufacturer
One regulatory jurisdiction

The revenue side may look diversified while the supply side is concentrated.

7. Valuation risk

Two companies can have unrelated businesses but similar valuation sensitivity.

If both require years of distant growth to justify today’s price, higher yields can compress both.

Valuation can create correlation where operations do not.

8. Liquidity risk

In normal markets, positions may behave independently.

During a forced selloff, investors sell what they can.

Correlation rises because liquidity becomes the shared driver.

This is why historical average correlation can understate the relationship that matters most.

Correlation is not fixed

Investors often speak about correlation as though every pair of assets carries a permanent number.

It does not.

Correlation depends on:

Time period
Return frequency
Market regime
Currency
Interest rates
Volatility
Liquidity
The specific shock being measured

Gold may diversify equity risk during one inflation or geopolitical episode and move with risk assets during another.

An airline and an industrial contractor may behave independently until oil and yields shock both.

Technology suppliers may show modest day-to-day correlation until one capex headline changes the entire AI complex.

The relevant question is not only:

What was the correlation?

It is:

What common event could make these positions behave as one?

That is a stress question, not merely a statistical one.

Why a correlation matrix is not enough

A correlation matrix is useful.

It is not a complete diversification test.

First, it is backward-looking.

It measures the period selected, not the next regime.

Second, it can be unstable.

Changing the window from one year to three months may produce a different answer.

Third, low historical correlation can hide a common future catalyst.

Two suppliers may not have traded together until a new customer concentration develops.

Fourth, correlation measures co-movement, not the size of the loss.

Two small positions can be highly correlated without threatening the account.

Two modestly correlated large positions can create more aggregate damage.

The portfolio needs both:

Relationship
Exposure size

Correlation without weight is incomplete.

Weight without correlation is incomplete.

Count risk in clusters

A practical portfolio can assign every holding to one or more risk clusters.

For example:

AI infrastructure spending
Semiconductor cycle
Consumer spending
Travel
Financial conditions
Commodity prices
Healthcare policy
Interest-rate duration
Geopolitical protection
Broad US equity beta

A position can belong to several clusters.

That is not double-counting.

It reflects reality.

Amphenol may sit inside electronics demand, data-centre infrastructure and industrial activity.

Freeport-McMoRan may sit inside copper, China/global growth and inflation sensitivity.

Delta may sit inside travel demand, fuel prices and consumer cyclicality.

GLD may sit inside real yields, currency and geopolitical demand.

SPY may touch almost every broad US equity factor.

The purpose is not to force one label onto each holding.

It is to make shared drivers visible before they become shared losses.

Position risk and cluster risk

Suppose five positions each carry a planned 0.5% loss to invalidation.

Viewed separately, every trade looks small.

If all five depend on the same capital-spending cycle, the portfolio may carry 2.5% of planned risk to one thesis.

That is before:

Gaps through stops
Correlation rising
Liquidity deteriorating
A single headline affecting all positions
Existing holdings carrying the same exposure
Index holdings adding indirect weight

The per-position limit worked.

The portfolio limit may still have failed.

Lesson 23 established the principle:

One account needs one risk budget.

This lesson adds the unit:

Shared economic drivers need a cluster budget.

The exact limit depends on the system.

The requirement does not.

SPY can hide concentration

A broad index is diversified across companies.

Adding SPY does not automatically diversify an existing US equity portfolio.

SPY may add exposure to many sectors.

It also adds more weight to companies already held directly.

If a portfolio owns Microsoft and then buys SPY, it has not created an independent Microsoft hedge.

It has increased Microsoft exposure indirectly while also adding other companies.

The same applies to QQQ more strongly.

An investor holding several mega-cap technology names plus QQQ may count the ETF as one new diversified position.

Economically, the ETF can increase the existing cluster.

Funds must be looked through to their underlying exposures.

The wrapper is not the risk.

The assets inside it are.

Cash is not another stock

Cash reduces invested exposure.

It does not need to be uncorrelated with equities to perform that job.

It carries different costs and risks:

Inflation
Currency
Reinvestment
Opportunity cost

But when equity clusters fall together, cash does not require a buyer, a hedge payoff or a precise inverse relationship.

It simply preserves capacity.

That is why cash can reduce aggregate portfolio risk even when it produces no visible return.

Lesson 25 will separate the jobs of cash, gold and insurance.

They are not interchangeable forms of “defence.”

Gold is not guaranteed insurance

Gold may provide a different macro exposure from operating companies.

It can benefit from falling real yields, currency concerns, inflation anxiety or geopolitical demand.

It can also decline.

It can become correlated with equities during liquidity shocks.

It does not guarantee a gain when stocks fall.

Owning GLD beside technology companies may improve the portfolio’s driver mix.

It does not create a mechanical hedge.

The distinction matters because diversification reduces dependence.

Insurance defines a contractual or structural payoff.

Gold is an asset.

Cash is capacity.

A hedge is protection purchased at a cost.

Those are different jobs.

The incremental-position test

Before adding a position, do not ask only whether the company qualifies.

Ask what it adds to the portfolio.

A useful checklist:

Which existing cluster will this position increase?
Which new driver, if any, will it introduce?
Which holdings share its customers?
Which holdings share its supply chain?
Which macro shock would hit them together?
How much total capital will sit in the shared cluster?
How much loss could occur if the cluster invalidates at once?
Does the position improve expected return enough to justify the overlap?
Could a smaller size preserve the opportunity without dominating the cluster?
Is another qualified company a cleaner source of the same expected return?

A good company can fail the incremental-position test.

That does not mean it is a bad investment.

It means the portfolio may already own enough of its underlying idea.

Diversification can reduce return

Diversification is not free.

If the best-performing theme continues rising, limits on that theme will reduce upside relative to full concentration.

Cash can drag.

Gold can lag equities.

A second engine can underperform the current leader.

That is not evidence that diversification failed.

It is the cost paid to avoid requiring one forecast to remain correct.

The investor must decide what the portfolio is designed to survive.

Maximum upside and controlled dependence are different objectives.

A portfolio cannot promise both without trade-offs.

Over-diversification is also real

More positions can eventually dilute the best ideas without adding meaningful new drivers.

A portfolio may become:

Too similar to an index
Hard to monitor
Expensive to rebalance
Full of second-choice ideas
Unable to express conviction
Dependent on broad beta while claiming active selection

The solution is not to own everything.

It is to own enough distinct, qualified exposures to meet the portfolio objective.

Diversification should be judged by marginal function.

If position 18 adds no new driver, no superior expected return and no useful protection, its ticker increases complexity without improving architecture.

Common failure modes

Sector labels as proof

The portfolio owns technology, industrials and materials.

All three depend on the same AI construction cycle.

Classification hides the common driver.

Equal weights as safety

Every position receives 5%.

The weights are equal.

The risks are not.

Volatility, invalidation distance and cluster overlap differ.

One-percent thinking

Each trade risks less than 1%.

Ten correlated trades can still create a large portfolio event.

ETF blindness

The investor counts SPY or QQQ as separate from the underlying stocks already owned.

Look-through exposure is ignored.

Historical-correlation confidence

The last year shows low correlation.

The next shock is assumed to behave like the last year.

Narrative diversification

Each position has a different story.

The cash flows depend on the same customer budget.

Defensive-label confidence

Gold, cash and hedges are grouped together as “defence” despite performing different jobs.

Green-screen blindness

All positions are rising.

Shared risk feels like skill until the common driver reverses.

Build a portfolio exposure map

A useful exposure map does not need institutional software.

Create rows for holdings and columns for the main drivers.

Mark each relationship:

Primary
Secondary
Indirect
None

Then add:

Position weight
Planned loss or review threshold
Sleeve
Sector
Geographic exposure
Major customers
Key input or supply-chain dependency
Event risk

The map should answer:

What is the largest economic cluster?
Which positions would react to the same headline?
Where is exposure duplicated through ETFs?
Which position adds the least new information?
Which cluster consumes the most risk?
What happens if correlation moves toward one?

The objective is not a perfect model.

It is to expose the assumption that every ticker is independent.

The three-stage diversification review

Stage 1: Before selection

Define the portfolio’s existing clusters.

A candidate is evaluated against what the account already owns.

Stage 2: Before execution

Calculate the new total weight and planned risk of every affected cluster.

Reduce size, substitute or reject if the overlap exceeds the limit.

Stage 3: After market change

Reassess clusters when the economic relationship changes.

A company may acquire a new customer, enter AI infrastructure or become dependent on a commodity that changes its portfolio role.

Diversification is maintained.

It is not completed once.

The structural fix

Do not write:

Maximum 20 positions.

Write:

Company limit

Maximum exposure to one issuer.

Cluster limit

Maximum capital and planned risk tied to one economic driver.

Sector limit

A secondary control for obvious industry concentration.

ETF look-through rule

Direct and indirect exposure are combined.

Stress rule

Estimate the loss if every position in one cluster moves adversely together.

Addition rule

A new holding must improve expected return, driver diversity or portfolio function.

Review trigger

Reclassify exposures after material business, customer or macro changes.

Now position count becomes one control among several.

It stops pretending to be the definition of diversification.

The real lesson

Diversification is not owning many things.

It is avoiding dependence on one thing.

One customer.

One capital cycle.

One commodity.

One valuation factor.

One interest-rate path.

One geopolitical outcome.

One market regime.

Different tickers can still depend on the same answer.

The July infrastructure cluster was not a mistake simply because it contained overlap.

The mistake would be counting five names as five independent sources of risk.

Portfolio architecture requires a more honest description:

Five companies.

Several business models.

One meaningful shared driver.

Once the driver is visible, it can be sized.

Once it is sized, the portfolio can decide whether the expected return is worth the concentration.

Count positions for administration. Count economic drivers for risk.

Next in Series 5: What cash, gold and insurance actually do.

Educational only—my own process and opinions, not investment advice. Copy trading involves risk, including loss of capital. Past performance is not an indication of future results.