Is Permanent Portfolio really diversified? Where its risk comes from

The portfolio

25% VTI, 25% TLT, 25% BIL, 25% GLD; rebalanced yearly. See Permanent Portfolio in the portfolio app, the Permanent guide, Permanent Portfolio with a trend filter.

Same figures, over the same window (Dec 2004–Sep 2026), as the Diversification section of the live portfolio page. The independent-bets count is the app's own measure; it counts the groups, so it inherits how the grouping handles chains of holdings (see Methodology). The sentences and bars on this page use the money and risk shares, which do not depend on the grouping.

What Permanent Portfolio is really betting on

• GLD (Gold) is 25% of the money and 45% of the risk.
• Each of the 4 holdings moves largely on its own: no pair correlates at 0.5 or more.
• BIL is 25% of the money but 0% of the risk.
• Over the last 10 years (Oct 2016–Sep 2026), the largest group carries 36% of the risk (45% over the full window).
• In the 36 months to Feb 2009 GLD (Gold) carried 45% of the risk; in the 36 months to Dec 2022, 29%.

Verdict: several groups of comparable size; the largest carries 45% of the risk.

Money versus risk, holding by holding

Full window Dec 2004–Sep 2026; last 10 years Oct 2016–Sep 2026. Shares use the target weights and sum to 100% in each window.

The full window versus the last 10 years

Over the last 10 years (Oct 2016–Sep 2026), the largest group carries 36% of the risk (45% over the full window).

Holdings are grouped when their monthly returns correlate at 0.5 or more, directly or through a chain of other holdings (single linkage on the clustering tree the app uses). For each multi-holding group the table gives the average correlation among its members and the least correlated pair; a lowest pair under 0.5 marks a chained group.

How the risk split moved

36 months to Feb 2009 is Mar 2006–Feb 2009, which holds the 2008 decline; 36 months to Dec 2022 is Jan 2020–Dec 2022, when stocks and bonds fell together. The full-window and last-10-year columns use the same fixed groups.

Correlation heatmap

How the holdings cluster

An equal-risk version of the same holdingsHypothetical, chosen with hindsight

An equal-risk version would hold BIL at 88%, with volatility 1.3% versus 7.3%, a worst drawdown of -2.1% versus -15.9%, and 2.4% a year versus 7.2%. It reaches equal risk mostly by holding cash-like BIL, so it also earned less.

Each holding contributes 25% of the risk in the equal-risk version (the "equal risk contribution", or risk-parity, weights). The weights come from the covariance of the whole window, which nobody knew at the start, so this is chosen with hindsight. It uses no leverage (leveraged risk-parity funds are a different thing), keeps the same yearly rebalancing, and is shown without a Sharpe ratio: at a risk-free rate of 0% a book that is mostly cash would look better than it earned. The running risk-parity strategy is on the risk-parity strategy page.

Check it in the builder

Frequently asked questions

Is Permanent Portfolio really diversified?

GLD (Gold) is 25% of the money and 45% of the risk. Each of the 4 holdings moves largely on its own: no pair correlates at 0.5 or more. Verdict: several groups of comparable size; the largest carries 45% of the risk.

Where does the risk in Permanent Portfolio come from?

GLD (Gold) is 25% of the money and 45% of the risk. BIL is 25% of the money but 0% of the risk. Over the last 10 years (Oct 2016–Sep 2026), the largest group carries 36% of the risk (45% over the full window). In the 36 months to Feb 2009 GLD (Gold) carried 45% of the risk; in the 36 months to Dec 2022, 29%.

What would an equal-risk version of Permanent Portfolio look like?

An equal-risk version would hold BIL at 88%, with volatility 1.3% versus 7.3%, a worst drawdown of -2.1% versus -15.9%, and 2.4% a year versus 7.2%. It reaches equal risk mostly by holding cash-like BIL, so it also earned less. It is hypothetical and chosen with hindsight: the weights come from the whole window's covariance.

Caveats

• Monthly data: risk shares and correlations come from month-end returns. Co-movement inside a month (a fast crash and rebound) is not seen, and daily correlations in a crisis are usually higher.
• A risk share is a share of variance (volatility), not of drawdown and not of the loss in any one year. A holding with a small share can still lose money in a bad year.
• The sample covariance moves with the window: the full window and the last 10 years can differ a lot (this page shows both), and no shrinkage is applied.
• Shares use the target weights, not the drifted weights between rebalances.
• The grouping is single linkage at a correlation of 0.5: a chain of pairs at or above it forms a group, so two holdings in one group can be uncorrelated with each other. Where that happens the page says so.
• Holdings before their ETF's launch use older funds or indexes as proxies (VTI, TLT and BIL); see their fund-history pages for how closely each proxy tracked.
• Gold history starts with GLD in Nov 2004, so this portfolio starts Dec 2004.
• The equal-risk version is in-sample (its weights use the whole window's covariance, which nobody knew at the start) and unlevered. Equal risk without leverage usually lowers volatility and return together.
• One historical path; past results do not predict future results. Not investment advice.

Methodology

Source

Yahoo Finance via MarketHeist, monthly adjusted close (dividends reinvested), to the last complete calendar month. Holdings before their ETF's launch use proxy funds, spliced on returns; see each holding's fund-history page.

Estimator

Sample covariance of monthly total returns (population scaling, which cancels in every share) and the Pearson correlation matrix. No shrinkage. This is the estimator the live portfolio page uses, so the figures can be checked there.

Risk share

Holding i's share of portfolio variance is w_i (Σw)_i / w'Σw, with the target weights. The shares sum to 100% and can be slightly negative for a diversifier.

Windows

Headline: the full common history (Dec 2004–Sep 2026, 21.8 years), the same window as the live page; it contains both the 2008 decline, when Treasuries hedged stocks, and 2022, when they fell together. Second column: the last 120 complete months (Oct 2016–Sep 2026), because the stock-bond correlation turned positive after 2021 and a full-window number averages that away; 120 months is the shortest window that keeps at least 10 observations per holding for the widest portfolio in the set. Regimes: the app's rolling 36-month risk shares. Grouping, the correlation heatmap and the equal-risk weights use the full window only, because 36 months is too few observations for 4 holdings. A page is not published if its window is under 15 years, lacks 2008 or 2022, has fewer than 10 months per holding, or has a singular covariance matrix.

Grouping

Holdings are clustered by single linkage on the correlation distance √(½(1 − ρ)) and cut at a correlation of 0.5, the same rule the live page uses. Single linkage chains: a holding joins a group if it correlates at 0.5 or more with any member, so two members can be uncorrelated with each other. Each multi-holding group's average and lowest internal correlation are shown, and the page says so where it happens.

Independent bets

The stat row shows the live app's "Independent Bets" figure: the exponential entropy of the groups' risk shares. It is shown for comparison with the app, not used in any sentence, because it inherits the chaining above. A different measure (the minimum-torsion effective number of bets) also exists; it can count many more bets for the same holdings, and we do not mix two bets figures on one page.

Equal-risk version

Long-only equal-risk-contribution weights solved by cyclical coordinate descent on the full-window covariance, checked to give every holding a 25% share to within 0.01 points; simulated over the same window with the same rebalancing. In-sample, unlevered, no costs or taxes.

Generated

2026-10-11; data through Sep 2026. Every figure is computed by the portfolio engine and every sentence is rule-based over those figures; no language model is involved. Reproduce hash of the Permanent Portfolio portfolio over its full window (Dec 2004–Sep 2026): 781a14207de07fb8 (the same hash as on its portfolio page).

Related

All risk breakdowns · Permanent Portfolio in the portfolio app · Permanent Portfolio with a trend filter · the Permanent guide

All risk breakdowns →

Historical simulation, not investment advice.