Diversification

Diversification is not what you think

The common definition of diversification: "own lots of things so one bad pick can't ruin you." This is close, but it hides the actual mechanism. Diversification works only to the extent that the things you own are uncorrelated. Owning ten tech stocks during a tech rout isn't diversification — it's ten positions in one bet.

Markowitz's 1952 insight that earned him a Nobel Prize: portfolio variance depends on the correlations between holdings, not just the individual variances. Kaufman reproduces the intuition (kaufman.txt:44319, 45346):

Portfolio variance is the weighted sum of individual variances plus the weighted sum of pairwise correlations. The correlation term is what diversification is actually managing.

This lesson covers what to measure, how to measure it, and why every diversification strategy fails in exactly one scenario — the scenario that matters most.

The math in one step

For a two-asset portfolio with weights $w_1$ and $w_2$, volatilities $\sigma_1$ and $\sigma_2$, and correlation $\rho$:

Three cases worth internalizing:

• $\rho = +1$ (perfectly correlated): $\sigma_p$ is the weighted average of $\sigma_1$ and $\sigma_2$. Combining two identical-risk assets gives you the same risk. Zero diversification benefit.
• $\rho = 0$ (uncorrelated): $\sigma_p$ is reduced proportional to $\sqrt{1/n}$ for $n$ equal-weighted assets. Diversification benefit.
• $\rho = -1$ (perfectly anti-correlated): $\sigma_p$ can be driven to zero with the right weights. Maximum diversification. (No such real-world pairs exist, but some come close.)

The key mental model: a 30-stock portfolio of all-tech names during a correlation spike is effectively one bet of higher-volatility tech. A 5-asset portfolio of stocks, bonds, gold, commodities, and cash — genuinely low pairwise correlations — can have substantially lower variance than the best individual asset.

Compute correlation on returns, not prices

This catches everyone once (kaufman.txt:9669). Never correlate price series directly. Two stocks that both trended upward over 5 years will show high correlation of price, but the daily returns — the thing you actually care about for risk — might be completely uncorrelated.

Always compute:

Then correlate the return series. Kaufman verbatim:

If you compute correlation on raw prices, you are measuring whether the assets trend together — not whether they move in sync on any given day.

For daily returns, 60+ observations is the practical minimum for a correlation estimate; 250 (one year) is better; 1000+ for stable rolling estimates.

The correlation matrix — your risk map

For a portfolio of $n$ assets, you compute an $n \times n$ matrix of pairwise correlations. Read it like this:

The diagonal is always 1. The off-diagonal entries are what matter. Read: SPY and QQQ are 0.94 correlated — essentially the same bet. SPY and TLT are −0.30 — meaningful diversification. SPY and GLD are 0.10 — nearly independent.

Heuristic: if a correlation exceeds +0.7, you have redundancy, not diversification. Two positions with correlation 0.94 behave as roughly 1.4 positions in risk terms, not 2.

Risk parity vs equal-dollar weighting

Equal-dollar weighting: $10,000 of each asset in the portfolio. Simple, intuitive, and wrong — because asset volatilities differ enormously. $10k of stocks and $10k of Treasuries doesn't split risk equally; stocks contribute maybe 5× as much variance as Treasuries, so the portfolio is an equity portfolio with a Treasury decoration.

Risk parity: size each position so it contributes equal variance to the portfolio. A high-vol asset gets smaller dollar allocation; a low-vol asset gets larger.

For a two-asset case:

If stocks have 16% vol and bonds have 4% vol, risk parity says weights = bonds 80%, stocks 20% — the opposite of what equal-dollar weighting would do. (Pure risk parity portfolios often leverage the bond side to hit target returns, which is how Bridgewater's All Weather fund operates conceptually.)

Neither framework is universally correct. The key insight: whatever weighting scheme you use, know which assets are driving your portfolio variance and by how much. Assume those are the ones that matter.

The correlation spike — where diversification fails

Here is the one thing every diversification framework misses. Kaufman's direct warning (kaufman.txt:39715):

During price shocks — 2008, 2020, 1987 — correlations across all risk assets spike toward 1.0. The diversification you built in calm markets evaporates precisely when you need it most.

This is empirically established over every major crisis:

• 2008: stocks, corporate bonds, REITs, commodities, foreign equities all fell together. Correlation of returns between S&P and emerging markets rose from ~0.6 in 2007 to ~0.95 in Q4 2008.
• March 2020: everything sold off simultaneously for 2-3 weeks. Even Treasury bonds briefly fell alongside stocks as liquidity demands forced indiscriminate selling.
• 1987 Black Monday: correlation among all equity sectors went to nearly 1.0 for a day.

Toggle between the two regimes. In the normal market, the three assets (stocks as candlesticks, bonds in blue, gold in amber) move independently — that's diversification working. In the crisis regime, stocks and gold collapse together while only bonds hold. The chart makes the correlation matrix real: what was 0.10 between stocks and gold in normal times becomes 0.95 in a panic.

Two things survive a correlation spike:

1. Cash (always)
2. Long-duration Treasuries (usually, but not always — see March 2020's bond sell-off)

Everything else — sector diversification, international diversification, commodities diversification — breaks down under panic selling. The reason: during liquidity crises, holders sell what they can, not what they want to. The correlation pattern reflects forced selling, not fundamentals.

Implication: diversification manages ordinary risk but not crisis risk. Crisis risk requires separate tools: hedges, tail-risk overlays, or simply lower total exposure.

The "how many positions" question

The academic literature (Evans & Archer 1968 and many replications) says: past about 20 uncorrelated positions, the marginal diversification benefit drops off sharply. A 20-stock portfolio captures ~95% of the diversification available; a 30-stock portfolio captures ~97%.

But "uncorrelated" is the key word. 20 tech stocks is not 20-position diversification; it's closer to 3 effective positions because the pairwise correlations among them are 0.8+.

Practical sizing rule: compute the effective number of positions using the Herfindahl-like index:

If your 30-position portfolio has $N_{eff} = 5$, you have a 5-position portfolio with 30 names in it. That's often the true state of retail "diversified" portfolios.

Diversification across dimensions

Beyond just "lots of stocks," real diversification spans multiple axes:

• Asset class: equities, bonds, commodities, currencies, cash
• Geography: US, developed ex-US, emerging markets
• Sector: within equities, don't concentrate in one of 11 GICS sectors
• Factor exposure: value, momentum, quality, size — these are often negatively correlated
• Strategy: trend-following AND mean-reversion AND carry — three negatively-correlated strategies will produce a smoother equity curve than any one of them alone
• Time horizon: day-trading strategy + swing strategy + position strategy, run simultaneously, don't share the same risk

A fund running 5 strategies across 4 asset classes in 3 regions is 60 cells on the diversification grid. That's what serious multi-strategy shops look like. Most retail portfolios are 1-2 cells — "long US stocks, maybe some bonds."

Quick check

What you now know

• Portfolio variance depends on correlations, not just on individual volatilities — Markowitz's foundational insight
• Always correlate returns, not prices — raw price correlation conflates co-movement with co-trending
• A high pairwise correlation (~0.9+) means your "diversified" portfolio has very few effective positions
• Risk parity sizes by variance contribution, not dollar amount; equal-dollar weighting over-weights whichever assets are most volatile
• Correlations spike to 1 in crises (2008, 2020, 1987) — diversification manages ordinary risk but not crisis risk
• True diversification spans asset class × geography × sector × factor × strategy × timeframe — not just "more tickers"
• Past ~20 genuinely uncorrelated positions, the marginal diversification benefit is small; concentrating in fewer, truly independent bets is often more efficient than many correlated ones

That's the final lesson in the Risk & Portfolio unit — you've now completed the full loop from does-the-system-work? (expectancy) through how-much-to-bet? (sizing) to how-do-bets-interact? (diversification). Next up: advanced patterns, market context, and forensic accounting.

Next: Gap Trading — breakaway, continuation, exhaustion, and area gaps with real statistics on fill rates and failure.