Ray Dalio's Holy Grail of Investing, Tested: Four Asset Classes Beat Nine Funds

TL;DR — the answer box

• Four beats nine. Same twenty-year window, same method: four asset classes (stocks, bonds, gold, oil) bought 3.72 independent bets out of 4. Nine asset-class funds bought 3.37 out of 9. Per holding, 0.930 against 0.375.
• The formula is right; the round number is not. Fifteen streams at zero correlation remove 74.18% of risk, not 80%. Twenty remove 77.64%.
• In his own units: fifteen uncorrelated streams take the chance of a losing year from 40.1% to 16.6%. The 1% at the bottom of his axis needs 100 streams, not fifteen.
• And real assets do not sit on that curve. The nine funds tested land at 32.4%, and no number of them ever gets below 29.4% — they behave like his 20%-correlation line, not his zero.
• Correlation, not count, sets a ceiling. At the nine funds' measured 0.2120, no number of them can ever remove more than 53.95% of the risk. Above about 0.25, halving your risk is unavailable at any count.
• Diversification did not fail in 2008. Across four asset classes it went negative in five of six major declines since 1983 — including the financial crisis at −0.0312. Only 2022 turned positive, at +0.1194.
• The 60/40 hedge is about 25 years old. Stock/bond correlation was +0.2465 in the 1970s, +0.3056 in the 1980s and +0.2685 in the 1990s. It only went negative after 2000.
• And the decade everyone measures is the misleading one. Over 2016–2026 the nine funds averaged 0.2816 correlation; over the full twenty years, 0.2120.

How we tested

Four things, because the claim has four separable parts.

The formula. For n equally weighted return streams that share a volatility and a pairwise correlation, portfolio volatility is the single-stream volatility multiplied by the square root of 1/n + (1 − 1/n) × correlation. That one line is Dalio's chart — the one he draws in his own explanation of the idea, and the one Bridgewater Associates built a firm on. I evaluated it for 1 to 20 streams at his slide's own five correlation levels. Nothing here is estimated; you can check any cell with a calculator.

Nine asset-class funds, twenty years. SPY, EFA, EEM, IEF, TLT, AGG, GLD, VNQ and DBC — US, international and emerging equity, two Treasury maturities, aggregate bonds, gold, property and commodities. Daily adjusted closes from 2006-02-07 to 2026-07-27, which is 5,146 daily returns each and covers the entire global financial crisis. Adjusted closes, not raw prices: for the bond and property funds the dividend is most of the return.

I report the 2016–2026 decade alongside it throughout, because that is the window I used the first time and the two disagree. The disagreement is a finding, not a footnote.

Four asset classes, forty-three years. The S&P 500 index, the 30-year Treasury future, gold and crude oil, from 1983-06-27, which is 10,776 days. One trap here and it is worth understanding: continuous futures are back-adjusted, so the historical level carries the accumulated roll gaps. The 30-year Treasury series is negative on 3,725 of its days. Percentage returns on a series like that are not slightly wrong, they are meaningless — so every futures leg is a daily point change divided by that market's own trailing 60-day volatility. Checked against two independent bond series, the normalised Treasury future correlates 0.8718 with a proxy built from Fed yield data and 0.8879 with the TLT fund.

Stocks against bonds, since 1970. Daily S&P 500 returns against a constant-maturity 10-year Treasury return built from FRED's published yields. That proxy correlates 0.9617 with the real 7–10 year Treasury fund over their 5,989 shared days, so it carries a correlation honestly. Its return level is optimistic, so no return figure is ever quoted from it.

This is measurement, not a trading system. Nothing is timed, nothing is signalled, and every basket is frictionless — no commissions, no spread, no taxes, no rebalancing cost.

What does Dalio's chart actually say?

Start with his own picture, because the headline gets quoted endlessly and the two columns that make it mean something almost never do. He walks through it himself here, and it is worth four minutes: Ray Dalio, "The Holy Grail of Investing".

The slide plots annual portfolio volatility as you add assets, at five correlation levels. Down the right it carries two more scales: a return-to-risk ratio and the probability of losing money in a given year. Reproductions of this chart almost always stop at the curves — and the right-hand columns are where the argument actually lands, because nobody feels "8% standard deviation" but everybody feels "a 38% chance of a losing year".

Both columns come back from two constants, and both check out against every value he prints.

• Every curve starts at 10% annual volatility. That is one return stream, held alone.
• Expected return is held constant at 2.5% a year. His return-to-risk column is just 2.5 ÷ volatility. Multiply each printed ratio back by its own volatility and you get 2.48% to 2.52% — one number, read back ten times through rounding. That constant return is the "without reducing your expected returns" half of the claim, made concrete.
• The probability column is the normal-distribution tail — the chance an annual return drawn from a bell curve centred on 2.5% comes out below zero.

Worst disagreement across all ten printed probabilities: 0.60 of a percentage point. Eight of the ten land inside 0.30. The chart is fully decoded.

So what do fifteen uncorrelated streams actually buy?

Fifteen uncorrelated streams take you from a 40.1% chance of a losing year to 16.6%. That is a genuinely large improvement — roughly one losing year in six instead of two in five.

And where do real assets land on his chart?

The dashed line on the figure above is the one that matters, because it is not an illustration. It is the nine asset-class funds this study measured, at their real twenty-year correlation of 0.2120, drawn on his axes and read across to his columns.

Put the two tables side by side and the whole argument is in three numbers. Fifteen genuinely uncorrelated streams: a 16.6% chance of a losing year. Nine real asset-class funds: 32.4%. And no number of funds like those ever gets below 29.4% — against the 40.1% you start with holding one thing.

Diversifying across nine funds moved the odds of a losing year by about eight points. Dalio's fifteen uncorrelated streams move them by twenty-three.

Notice also where that dashed line sits: at twenty holdings it is at 5.01% volatility against 4.90% for his 20%-correlation illustration. Nine real asset-class funds behave like his 20% case, not his 0% one. When people picture themselves on the green curve at the bottom of that chart, they are standing on the black one in the middle.

It is also not the bottom of his chart, and this is the part I had never seen anyone point out. The lowest row he prints is 1% volatility, 2.50 return-to-risk, a 1% chance of a losing year. Reaching it takes 100 uncorrelated streams, not fifteen — volatility falls with the square root of the count, so 10% ÷ √100 = 1%. The rows in between cost 11 streams for 3%, 25 for 2%, and 100 for 1%.

And at 60% correlation, twenty streams still leave 7.87% volatility and a 37.5% chance of a losing year. Barely moved from the 40.1% you started with. Correlation is the whole game.

One assumption is his, and it flatters the answer. That probability column assumes annual returns are normally distributed. Real annual returns have fatter tails, so every number in it is the optimistic reading.

Does the formula behind it hold up?

Yes. It is exact, and it reproduces perfectly.

Read the top-right corner against the bottom-left and the whole argument is there. Twenty streams at 60% correlation buy a 21.26% risk cut. Five streams at zero correlation buy 55.28%. Correlation matters far more than count.

I published an earlier version of this article in December 2024, so let me correct my own numbers rather than quietly restate them.

One correct, one rounded up, one overstated by nearly six points. "80% with fifteen" is the line that travels, and the honest version is 74.18% with fifteen and 77.64% with twenty — both needing a correlation of exactly zero.

What is the most risk diversification can ever remove?

Push the number of streams toward infinity and the formula does not go to zero. It flattens at the square root of the correlation. So correlation alone caps how much risk diversification can ever take out, however long your holdings list gets.

Above about 0.25 average correlation, halving your risk is not expensive. It is impossible. That number is the one to hold onto, because everything below is about which side of it real portfolios land on.

How correlated are the main asset classes, really?

Here is where the theory meets the account statement.

Nine funds, 5,146 daily returns, average pairwise correlation 0.2120 — which is 3.34 independent bets. The arithmetic is the same formula read backwards: n ÷ (1 + (n−1) × correlation), the number of genuinely uncorrelated streams that would carry the same risk as what you actually hold.

The individual pairs say why. The two Treasury funds correlate at 0.9111 — two lengths of one bet. US and international equity at 0.88. Emerging and international equity at 0.87. Seven of the 36 pairs sit above 0.70. Only 13 sit below 0.20.

And the decade I measured the first time was the more correlated half: 0.2816 over 2016–2026 against 0.2120 over the full twenty years, which is 2.77 bets against 3.34. Measuring the convenient window made the picture look worse than it is.

Does holding more things buy more diversification?

No. Holding more different things does, and the gap between those two sentences is the whole article.

Same twenty-year window. Same method. Four holdings beat nine.

Two things could have faked that, and both are controlled. The four could have won on the longer window — they did not, because over their full 43 years they score 3.79 and over the matched twenty they score 3.72, essentially the same. Or they could have won on the different return definition — they did not, because running the nine funds through the identical normalisation moves them from 3.34 to 3.37.

Even choosing the best four funds out of the nine with perfect hindsight only reaches 3.50. You cannot fix a shopping list of similar things by picking more carefully from it.

You hired nine analysts and six of them read the same newspaper. Firing five and hiring an oil trader would have told you more.

Does diversification hold up in a crisis?

Mostly it does, and the crisis everyone cites as proof that it fails is the one where it worked best.

Five of six went negative. The four asset classes pulled apart under stress, which is exactly what you want and the opposite of the received wisdom.

The nine-fund panel tells the same story in its own units. The 2008 bear measured 0.1316 average correlation while the S&P 500 fell 54.77% — the calmest episode in the sample. The Lehman quarter measured 0.1499. And 2022 measured 0.3353, the worst of any episode in twenty years.

This also kills a claim I made in the earlier version of this article, and it is worth being blunt about it. Over the 2016–2026 decade, stressed days looked more correlated than calm ones — 0.3191 against 0.2583 — and I wrote that correlations rise when markets fall. Over the full twenty years it reverses: 0.2000 on days the S&P 500 sat more than 10% below its high, against 0.2233 otherwise. The decade said one thing. Twenty years says the opposite.

The single number the whole chart rests on is not a number either. The rolling 60-day average pairwise correlation ran from −0.0163 to 0.5841, against a median of 0.1987.

Is the stock/bond hedge a law or a phase?

A phase. This is the finding I did not expect and the one I would keep if I could keep only one.

For the first thirty years of the sample, stocks and bonds moved together. The negative correlation that makes a 60/40 portfolio work — and that made bonds the hero of 2008 (−0.4673) and of COVID (−0.4960) — only shows up after 2000.

Which reframes 2022 completely. At +0.1147 it is not an anomaly that broke the rules. It looks like the pre-2000 normal coming back. Over the whole 1970–2026 sample the correlation is −0.0264 — essentially nothing, because the two eras cancel out.

If you built a portfolio on the assumption that bonds hedge equities, you built it on twenty-five years of a fifty-six-year record.

What did all that diversification actually buy?

MAR is annual return divided by the worst percentage drawdown — what you earned for the deepest hole you had to sit in. Volatility alone settles nothing; a portfolio can halve its volatility by halving its return and has achieved precisely nothing.

Against the average single asset, diversification won clearly. Volatility 39.03% lower, worst drawdown 34.21% smaller, MAR 55.63% higher. That is the honest comparison, because the alternative to a diversified portfolio was never "hold whichever asset turned out best" — it was "hold one thing", chosen in 2006 without knowing which.

Against the S&P 500, it lost narrowly. MAR 0.19 against 0.20. And note the row that beat everything: a plain 60/40 scored 0.23 from two holdings. Nine funds did not beat two.

There is a gap worth naming too. The formula assumes every stream carries the same volatility; these run from 5.29% to 29.02%. At the measured correlation the formula predicted the nine-fund basket would run at 0.5474 of a single stream's volatility. It actually ran at 0.6255 — more volatile than promised, because in an equal-weight portfolio the riskiest holdings dominate the variance.

The verdict — and the honest limits

Where the hype is right. The maths is a proof, not a marketing claim, and it reproduces to the decimal. And when you apply it to genuinely different things it delivers: four asset classes bought 3.79 of 4 possible independent bets across 43 years, and pulled apart rather than together in five of the six worst declines in that span. Against the realistic alternative of holding one asset, the diversified basket cut drawdown by a third and improved return per unit of drawdown by more than half.

Where the hype is wrong. "Fifteen uncorrelated return streams" gets quoted as though the streams are lying around waiting to be collected. Nine of the most distinct funds an ordinary investor can buy delivered 3.37 bets between them. The binding number is not fifteen, it is 0.2120 — and above roughly 0.25, halving your risk is off the table no matter how long the list gets.

And the word everyone drops. Dalio says "fifteen or more good, uncorrelated return streams." The internet quotes it without good. Over twenty years broad commodities returned 2.05% a year with a 76.36% drawdown; 20-year Treasuries returned 2.82% with a 48.35% drawdown. Both diversified the portfolio beautifully and dragged its return down while they did it.

I ran one system at a time for years, and each time one stopped working I assumed I had picked the wrong system. I had not. I had picked one bet and called it a portfolio. What took me longer to learn is the thing this study measures: adding a tenth version of the same bet is not diversification, it is paperwork.

Limits, because a study that hides them is not worth trusting:

• Twenty years for the funds, forty-three for the asset classes, fifty-six for stocks and bonds. None of it is a century. The fund panel cannot see the 1970s at all.
• The bond series before 2002 is a construction, not an index. Built from published constant-maturity yields; it correlates 0.9617 with the real fund but has no fees, spread or financing cost, so it is used for correlation and never for a return figure.
• The S&P series in the long block is price-only, without dividends. Correct for correlation, wrong for return, and never used for return.
• Back-adjusted futures needed a workaround. Point changes normalised by trailing volatility, validated at 0.87–0.89 against two independent bond series. It is a defensible choice, not the only one.
• Asset classes, not trading strategies. Dalio says "return streams", and for a systematic trader that means strategies. That is the better test and it is not this one.
• Frictionless. Rebalancing nine funds monthly is not free, and the real basket would be a little worse.
• This is measurement, and it is not investment advice. Nothing here is timed or recommended, and past patterns do not guarantee future ones.

What this means for you

1. Count your bets, not your holdings. Take the daily returns of what you own, compute the average correlation between them, and run n ÷ (1 + (n−1) × correlation). If nine positions come back as three, you now know what you actually own. It is the most useful ten minutes in this article.
2. Add a different thing, not another thing. Four asset classes beat nine funds and beat the best four funds pickable with hindsight. When you consider a new holding, the question is not "is it good" but "what does it do on the days everything else is falling".
3. Check which side of 0.25 you are on. Above roughly that average correlation, halving your risk cannot be bought at any holdings count. If you are there, adding positions is not the fix — replacing them is.
4. Do not count cash as a return stream. It is uncorrelated because it does not move. It belongs in your risk plan, not in your diversification count.
5. Stop assuming bonds will hedge your equities. That relationship was positive for the thirty years before 2000 and has been drifting back toward zero since 2020. Build for both regimes rather than the one you grew up in.
6. Judge every stream on both tests. Uncorrelated and worth owning. A stream that fails the second lowers your risk and your return together, and MAR will tell you it was a bad trade.

For a portfolio of trading strategies rather than asset classes the same arithmetic applies, and the measurement is harder because you need each strategy's own equity curve before you can correlate anything. Watching live strategies side by side and seeing how their returns move together is the job AlgoChef is built for. If you want the drawdown side of this in depth, the S&P 500 drawdown study measures every decline of 10% or more since 1871, and the case against running a single system covers the same lesson from the strategy side.

Sources. Ray Dalio, Principles (Simon & Schuster, 2017) — the origin of the "fifteen or more good, uncorrelated return streams" formulation and of the chart reproduced here. Ray Dalio, "The Holy Grail of Investing" — Dalio explaining the concept in his own words. Harry Markowitz, "Portfolio Selection," The Journal of Finance 7, no. 1 (1952): 77–91 — the mean-variance result the "only free lunch" line descends from. Federal Reserve Bank of St. Louis, series DGS10 — the 10-year Treasury constant-maturity yield used to build the long bond series. Fund and index prices from the Yahoo Finance chart API; continuous futures from TradingView. Full provenance in the study's own data notes.

For the developer — paste-ready: the block below is valid JSON-LD (one @graph). Drop it as-is into a single <script type="application/ld+json"> tag in the page <head> — no edits needed to make it valid. Before publishing, confirm the URL / @id paths resolve and update datePublished / dateModified if the live date differs.