Diversification Beyond Sector Splits: Factors, Geography, and Correlation Regimes

Owning stocks in different sectors is the beginning of diversification, not the end of it. Sector labels classify what companies sell; risk lives in how their stocks move together. A portfolio can span five sectors and still amount to a single bet, on growth, on US mega-caps, on falling rates, repeated in different wrappers.

Every figure below is computed rather than cited: factor loadings use the fixed regression window from August 10, 2021 to May 29, 2026; co-movement uses daily closes over the five years to August 2026; look-through weights use SEC N-PORT equity constituents, with VXUS as of April 30, 2026, VGT as of May 31, 2026, and the other funds as of June 30, 2026. The fund pairs named all have live comparison pages you can check the inputs against. The September 15 restatement replaces issuer-scraped holdings with filings: VTI/VXUS moves from 1.1% to 0.2%, and VEA/VXUS from 61.3% to 71.5%. Both comparisons use the older April 30 date as their anchor and disclose the June 30 date on the other side.

Why isn't sector diversification enough?

Because sector classification is a labeling system, not a risk model. Apple (tech), JPMorgan (financials), and Pfizer (healthcare) feel diversified, but in a broad risk-off move correlations spike toward 1.0 and all three fall together. Meanwhile two stocks in the same sector can behave completely differently, a high-multiple software firm and a legacy IT dividend payer share a label and almost nothing else.

Three deeper layers matter more than the sector pie chart:

Can a perfect sector chart hide a single bet?

It can, and it is worth seeing the extreme case. Take five equally weighted stocks in five different sectors: Microsoft (technology), Amazon (consumer cyclical), Alphabet (communication services), JPMorgan (financial services) and UnitedHealth (healthcare). Scored the way a sector chart scores things, effective number of sectors, 1 ÷ Σ(share²) over sector weights, this portfolio comes out at 5.00 effective sectors. That is the maximum available from five buckets. There is no sector chart that shows a problem here.

Now measure the same five holdings by behavior. Over the 1,244 shared trading days to August 2026, Microsoft, JPMorgan and UnitedHealth all closed lower on the same day 16% of the time. Restrict to the 125 worst days for the US market, every session a total-market fund fell more than 1.24%, and all three closed lower together on 65% of them.

Drop the two sectors that were doing the diversifying and the effect sharpens. Microsoft, Amazon and Alphabet sit in three different sectors and score 3.00 effective sectors. They fell together on 28% of all days and 92% of the market's worst 125. A three-factor regression on that trio returns a value loading of −0.50, a single, concentrated growth bet with three sector labels on it.

The sector chart is not lying. It is answering a question about labels while you are asking a question about risk.

What does factor diversification look like?

Factors, market, size, value, and friends, explain co-movement across sector lines. A large-cap growth tech position plus a large-cap growth consumer position is not two bets; it's one factor bet, twice. Genuine factor spread means owning things on opposite sides of a driver: value alongside growth, smaller caps alongside mega-caps. Compare the dated holdings reports for VTV vs VUG, a value and growth pair, and VGT vs XLK, two technology-sector funds. Holdings overlap and factor exposure answer different questions; one cannot be inferred from the other.

Regressing each fund's daily excess returns on the Fama-French market, size and value factors puts numbers on where each one actually sits. The loadings below are computed from daily closes over 1,206 trading days, August 2021 to May 2026, the window in which both the fund's prices and the published daily factor series overlap.

Three things fall out of that table that a sector chart cannot show you.

Opposite signs can offset a particular tilt. VUG loads −0.35 on value; VTV loads +0.39. At equal weight, their combined value loading computes to +0.02 for this window. That does not remove their market exposure or establish their holdings overlap.

A different issuer does not establish a different exposure. VGT and XLK are two technology funds from different providers. In this regression, their market loadings are identical to two decimals and their value loadings differ by 0.02. Their dated holdings comparison measures shared company weights separately.

Low R² is information too. The three US factors explain 99.5% of VOO's daily variation and 66.8% of VXUS's. The unexplained variation can include currency, local policy, local-market conditions and other influences. R² does not identify or separate those causes. The Ken French Data Library's factor definitions, accessed September 14, 2026, specify the US market, size and value series used here.

Where does fund overlap fit in?

Look-through is the layer most sector charts skip entirely. Funds are containers; diversification happens (or doesn't) at the level of what's inside them. Holding an S&P 500 fund, a total-market fund, and a handful of mega-cap stocks directly means your largest positions are counted two or three times over, your effective number of holdings is a fraction of your nominal one. The mechanics and the fix are covered in What Is Real Diversification?

A concrete version: 45% VTI, 30% VUG, 25% VGT. Three funds, three different index rules, 3,500 distinct underlying companies. Looked through, NVIDIA comes to 10.9% of the portfolio and Apple to 9.8%, each of them arriving through all three funds, and the ten largest underlying companies hold 44.5% of the money. Effective holdings: 31.2, against 3,500 names owned.

Swap the two style funds for exposures the core fund lacks, 34% VOO, 33% VTV, 33% VXUS, and the same three-line-item shape gives 210.8 effective holdings with 16.1% in the top ten. The difference is not the number of funds. It is whether each one holds something the others don't.

Does geographic diversification still matter?

Yes, precisely because it hasn't paid for a decade. US outperformance made all-US portfolios feel like prudence rather than concentration, but a 100%-US book is a single currency, policy, and valuation bet. International allocation adds a currency hedge against dollar weakness, exposure to cheaper markets, and different monetary cycles. The uncomfortable rule: diversification you add after it starts working isn't diversification, it's chasing.

What it does and does not buy is measurable. By domicile, VTI targets the US stock market. VXUS targets developed and emerging markets outside the US, and the two funds overlap by 0.2% of weight, about as close to disjoint as two equity funds get.

Their daily returns still correlate at 0.81 over the five years to August 2026, and on the market's 125 worst days both funds closed lower on 98% of them. Geography changes which economies, currencies and policy regimes you are exposed to. It does not turn an equity portfolio into something that rises when equities fall, and treating it as a drawdown hedge is how the layer gets discredited when it fails to behave like one.

Note also that "international" is itself a bundle. VEA and VXUS overlap by 71.5%, the developed-markets fund is most of the ex-US fund. Owning both is the same stacking problem in a different aisle.

What are correlation regimes, and why should I care?

Correlations aren't constants, they move with the macro environment:

The portfolio that matters is the one you hold in the bad regime. Stress-test with crisis correlations, assume everything equity-like moves together, and see whether your "diversifiers" survive the assumption.

One caution on measuring this yourself. The obvious method, compute the correlation across only the bad days, is biased, and biased in the direction that makes the problem look smaller. Conditioning on large market moves truncates the sample's variance, and a correlation computed inside that truncated sample can come out lower than the full-window figure even when the assets are visibly falling in lockstep. Counting co-declines, as above, avoids the trap: it asks how often both fell, not how tightly they scattered around a conditional mean.

How do I tell a deliberate tilt from an accidental one?

The factor table above is a description, not a verdict. A −0.35 value loading is a mistake only if nobody chose it. Three questions separate the two cases:

1. Could you state the tilt before you measured it? An investor who bought a growth fund on purpose can. An investor who bought a core fund, a growth fund and a tech fund because each looked sensible on its own usually cannot, the 45/30/25 portfolio above carries a −0.19 value loading that nobody selected.
2. Did the tilt arrive by drift? A winning factor grows its own weight. Yesterday's balanced allocation becomes today's concentrated one without a single trade, which is why the tilt you chose and the tilt you hold are different measurements.
3. Is the exposure you wanted available elsewhere? If a growth fund was bought for growth exposure a core index already delivers, VOO's value loading is +0.01, essentially neutral, but its largest holdings are the growth names, then the gap it was meant to fill may be somewhere else: smaller companies, other geographies, other factors.

A practical checklist

1. Look through your funds first, duplicate exposure is the cheapest problem to fix.
2. Audit factor tilts, deliberate tilts are fine; accidental ones aren't.
3. Add international exposure on purpose, even 20–30% changes the currency math.
4. Assume crisis correlations when sizing risk, if the portfolio only works when correlations stay low, it doesn't work.
5. Rebalance with intent, drift concentrates every portfolio eventually.

What this analysis can't tell you

The layers above describe structure, and structure is not the whole of risk.

• Every figure here is realized, not forecast. Factor loadings use the fixed August 10, 2021 to May 29, 2026 regression window; correlations and co-decline rates use the five years to August 2026. Look-through weights reflect holdings as of 30 April 2026 for VXUS, 31 May 2026 for VGT, and 30 June 2026 for the other funds. Different windows or holdings dates give different numbers, and none of them is a prediction.
• Factor models are incomplete by construction. The three-factor model left a third of an international fund's movement unexplained. What sits in the residual is not zero risk, it is risk the model doesn't name.
• Look-through is bounded by published holdings. A fund that doesn't publish constituents can't be exploded, and the uncovered portion silently drops out of the arithmetic.
• Sector labels still have uses. They are a poor risk model and a perfectly good inventory. The argument here is against treating the pie chart as the answer, not against drawing it.

A free PortLens scan shows company exposure, overlap and portfolio beta, computed as described in our methodology. Factor loadings require a separate regression, available through PortLens Pro's Fama-French analysis. The goal isn't eliminating risk; it's making sure every risk you carry is one you chose.

This article is for information and education only and is not investment advice. See our methodology and disclosures.