The textbook answer is 20 to 30 stocks. It comes from one calculation: how fast the standard deviation of a randomly chosen portfolio falls as you add names. Meir Statman's 1987 paper in the Journal of Financial and Quantitative Analysis put the figure at 30 stocks for a borrowing investor and 40 for a lending one. He was arguing upwards, against a then-common belief that about 10 was plenty.
Here's a second number. Across the CRSP database from 1926 to 2016, 1,092 companies accounted for all of the net wealth the US stock market created. That's 4.31% of the 25,332 firms in the sample. The other 96% collectively matched one-month Treasury bills and no more.
The first number answers how much a portfolio wobbles. The second answers whether you owned the companies that made the money. Those aren't the same question, and they don't have the same answer.
Where the 20-to-30 answer came from
Statman's method was variance reduction. Take a randomly selected portfolio, add stocks one at a time, and watch the standard deviation fall towards the market's. The drop is steep early and flattens fast. Past some point the extra risk reduction isn't worth the cost of another holding. That point is where the classic answer sits.
Statman himself moved it later. In a 2004 Financial Analysts Journal paper he wrote that the optimal level of diversification under mean-variance rules exceeded 300 stocks, while the average investor held three or four. He called the gap a puzzle and answered it with behavioural portfolio theory: people build portfolios in layers, with a protective base and a lottery-ticket top.
So the variance answer was never fixed. It moves with the correlation structure of the market. Campbell, Lettau, Malkiel and Xu documented that in 2001, showing that firm-level volatility rose relative to market volatility between 1962 and 1997. The number of stocks needed for a target risk level rose with it.
Their 2022 follow-up gets quoted far less. Idiosyncratic volatility spiked around 1999 and 2000, then declined. The share of firm-level variance in equal-weighted data fell from roughly 95% at the end of the original sample to around 80% in recent data. The authors say plainly that they never expected the upward trend to continue. They also wrote the update partly to answer two replication studies, by Chiah, Gharghori and Zhong and by Leippold and Svaton, whose samples stop in 2017 and 2016 respectively.
The honest summary is awkward. The variance-based answer sits somewhere between 20 and 300, it depends on the decade, and its own authors have revised it. That's before you ask whether variance is the right thing to be measuring.
What the skewness evidence measures instead
Hendrik Bessembinder's 2018 paper asked a blunter question. Take every common stock in CRSP since 1926, hold it from first appearance to delisting, and compare the result with rolling one-month Treasury bills over the matched window. Only 42.6% of stocks beat the bills. The single most frequent lifetime outcome, with returns rounded to the nearest 5%, is a loss of 100%.
Individual stocks also don't last long. The median time a stock spent in the CRSP database across those 90 years was seven and a half years. Just 36 stocks were present for the full period.
The wealth-creation side is more lopsided still. Net wealth creation across the sample came to $34.82 trillion. The top 90 firms, 0.36% of the total, produced half of it. The top 295 produced three quarters. Exxon Mobil alone accounted for 2.88%, and the top five firms for 10.07%.
None of this contradicts the variance work. It's measuring something the variance work never priced: the chance that a small portfolio simply misses the companies that mattered. That's a different failure mode from a bumpy ride, and it doesn't show up in a standard deviation. The distinction is close to the one between volatility as a statistic and the drawdowns investors actually feel, except here the mismeasured thing is upside rather than downside.
A five-stock example that shows the mechanism
Heaton, Polson and Witte built the cleanest illustration in their 2017 paper "Why Indexing Works". Picture an index of five securities. Four of them return 10% and one returns 50%. The equally weighted index returns 18%.
Now form every possible one-stock or two-stock portfolio from those five. There are 15. Ten of them contain only the 10% names. Five contain the winner. The average across all 15 portfolios is 18%, identical to the index, exactly as the arithmetic requires. The median is 10%. Two thirds of the possible portfolios lose to the index.
Nothing in that example involves volatility, correlation or risk tolerance. It's pure distribution shape. The mean and the median of a skewed distribution sit in different places, and a small portfolio lands near the median.
Variance converges long before outcomes do
Bessembinder separated the two effects with bootstrap simulations, drawing random value-weighted portfolios 20,000 times and linking monthly returns over one-year, decade and 90-year horizons.
The variance story resolves quickly. The skewness coefficient of annual returns falls from 6.99 for single stocks to 1.08 at five stocks and 0.10 at 25. At 50 and 100 stocks it turns slightly negative, at -0.09 and -0.21. By any variance-based test, a 50-stock portfolio is fully diversified.
The outcome story doesn't resolve. Measured against the cap-weighted market over the full 90 years, five-stock portfolios won 22.68% of the time. Twenty-five stocks got you to 36.81%. Fifty stocks reached 40.94%. A hundred stocks reached 43.29%. No fees or trading costs are deducted from any of those figures.
Read that sequence again. Going from 25 holdings to 100 buys about six and a half percentage points, and still leaves you likelier to lose to the index than beat it. The variance gap had closed by 50 names. The outcome gap was still open at 100.
Against Treasury bills the picture is friendlier, which matters when the benchmark is cash rather than an index. At the decade horizon, the share of bootstrapped portfolios beating bills rose from 47.77% for a single stock to 72.29% at five, 86.86% at 25 and 93.08% at 100. Diversification does reliable work against cash. It does much slower work against the market.
The same pattern outside the US
Bessembinder, Chen, Choi and Wei extended the study to more than 64,000 global stocks over 1990 to 2020. In compound terms, 55.2% of US stocks and 57.4% of non-US stocks underperformed one-month US Treasury bills. The top-performing 2.4% of firms accounted for all $75.7 trillion of net global wealth creation. Outside the US, 1.41% of firms accounted for $30.7 trillion.
The UK figure deserves pulling out. Of the 4,192 UK stocks in that sample, 26.2% beat the value-weighted market over their lifetimes. Germany came in at 25.5%, Hong Kong at 21.1% and Japan at 17.7%. In 41 of the 43 markets studied, fewer than half of listed stocks beat their own market index.
The portfolio-size result repeats abroad as well. Among non-US stocks over the full 31 years, the share of bootstrapped portfolios beating US Treasury bills ran 26.8% at one stock, 53.3% at five, 75.9% at 25, 83.2% at 50 and 89.0% at 100. Against the value-weighted market, 100-stock portfolios won 39.6% of the time in US stocks and 45.4% in non-US stocks. Since a UK or European sleeve is often where a portfolio's concentration hides, this bears on how much of an equity allocation ends up outside the home market.
Three caveats belong next to those numbers. The aggregate isn't a universal law: in twelve of the markets more than half of individual stocks did beat US Treasury bills, and more than 60% did in Saudi Arabia, Israel, Switzerland and Finland. Where accurate delisting returns were unavailable for non-US firms, the authors imputed a return of -30%, which is an assumption rather than a measurement. And Bessembinder discloses financial support from Baillie Gifford, a manager whose investment case rests on concentrated exposure to extreme winners.
The case against: index funds already settle this
The strongest objection is that none of this matters to most people. A cap-weighted total-market fund owns the 4% by construction, so the skewness evidence is an argument for indexing rather than a live puzzle about position counts. Heaton, Polson and Witte put that in their title. Bessembinder's own abstract argues that his results help explain why poorly diversified active strategies most often underperform market averages. On the mechanism, the critics are right.
The objection is weaker on what it implies about the number. If the answer is to own the index, then the answer to "how many stocks?" isn't 30. It's roughly all of them, weighted the way the market weights them. Indexing doesn't dissolve the question. It answers it with a figure two orders of magnitude larger than the textbook one, while the textbook number keeps circulating as guidance for people who pick stocks.
And people do pick stocks. Goetzmann and Kumar examined more than 40,000 accounts at a large US discount brokerage over 1991 to 1996. More than 25% held a single stock. More than half held fewer than three. Statman's 2004 figure of three or four names for the average investor points the same way. That brokerage sample is old and self-selected, and the authors acknowledge they have probably picked up speculators, but no plausible correction turns three stocks into thirty.
There's also a middle case the objection skips past. Plenty of portfolios sit between one stock and the whole market: a concentrated active fund, a thematic tracker, or several broad funds that quietly hold the same giants. Our look at five popular funds sharing 39% of assets across ten companies is one version of that problem, and two decades of SPIVA data on active fund underperformance is the aggregate scoreboard the skewness argument predicts.
What would change the conclusion
Several things, and the first is already published.
Oh and Wachter's 2018 NBER working paper is the sharpest response to Bessembinder. They argue that the headline result, most stocks losing to Treasury bills, is exactly what a standard lognormal return model predicts, and so poses no challenge to that model. They find 48% of monthly stock returns exceed the bill return in the data, and 48% in their simulated economy too. Their criticism runs the other way: the lognormal model fails on the magnitude of monthly cross-sectional skewness, which far exceeds what the model implies, and it overstates skewness in long-run returns. This is a working paper rather than a peer-reviewed article, and it doesn't dispute the arithmetic. It disputes how surprising the arithmetic ought to be.
That distinction is worth sitting with. If the skew is mostly mechanical, the practical consequence for a small portfolio is unchanged, because it still lands near the median. But the finding becomes evidence about compounding rather than evidence about markets.
Sample composition is the second lever. Bessembinder is explicit that the failure to beat Treasury bills is concentrated in stocks below median market capitalisation and in stocks that entered the database after the mid-1960s. An investor whose universe is large, established companies faces a milder version of this distribution.
Third is the counting method. Every firm counts once, whether it was a micro-cap that survived four years or a mega-cap that ran for nine decades. "Most stocks" and "most of the money" use different denominators, and a capitalisation-weighted count would flatter the result.
Fourth is the holding period. These are lifetime buy-and-hold returns to delisting, on a median life of seven and a half years. Real portfolios rebalance, sell and replace, and this evidence doesn't settle whether that helps or hurts.
Fifth is the trend itself. If firm-level volatility keeps falling relative to market volatility, as the 2022 Campbell, Lettau, Malkiel and Xu update indicates it has since 2000, then the variance answer and the skewness answer both shrink together.
What the two numbers actually say
Variance-based diversification is close to finished somewhere between 25 and 50 stocks. That part of the textbook survives. What it never covered is the probability of missing the winners, and that probability stays material at 100 holdings: 43.29% of 100-stock portfolios beat the cap-weighted market over 90 years, and 39.6% did over the 31-year global sample.
The two figures answer different questions, and the older one got treated as though it answered both. If the aim is a smoother ride, 30 names does most of the work. If the aim is to reliably capture a distribution in which about 4% of companies produced all the gains, no realistic number of hand-picked holdings gets there.