In this article
← All articles

Kelly criterion for investors: how much stock exposure maximises long-run growth?


The Kelly criterion picks the stock exposure that maximises expected long-run compounded growth, but its answer rests on inputs known only afterwards: in our calculation on US data from July 1926 to September 2026, the exposure that compounded fastest was 191%, a mix that fell 98.3% in 1929-1932.

That 191% figure is a historical, in-sample result, reset monthly and financed at the Treasury-bill rate plus 1.5 points, one large online broker's first-tier margin spread today, applied to the whole century. Plain 100% exposure compounded at 10.36% a year against 12.15% for 191%. At 5 points over bills, borrowing added nothing: the fastest-growing exposure was 100%, the market itself. The useful lessons are the shape of the growth curve, what each step up in exposure cost in crashes, and how far the best exposure moved between periods.

Key takeaways

  1. Growth after borrowing costs flattened near the peak and then fell fast. From 150% to 250% exposure, compounded growth stayed within 0.9 points of the 12.15% peak; by 344% it was below the 3.29% of Treasury bills.
  2. A crash turned extra exposure into lost decades. At 191% the 1929-1932 fall took 98.3% of wealth and kept it below its previous high for 22.8 years, against 83.7% and 15.3 years at 100%; from 347% the account was wiped out.
  3. Leverage adds exposure to the same return engine, not diversification. Every exposure from 15% to 300% had its worst fall in the same August 1929 to June 1932 decline, so borrowing to buy more of one index cannot cushion that index's crash.
  4. The best exposure moved more than the growth it bought. In four 25-year windows the realised peak was 116%, 273%, 237% and 285%, and the formula's own uncertainty spans 141% to 354%: a full-Kelly figure is an estimate, not a target.
  5. Borrowing paid only when it was cheap. The best exposure fell from 191% at 1.5 points over bills to 127% at 4 points and 100% at 5 points; held out of sample at 4 points, the fitted exposure barely beat the market.
  6. A household can copy the method, not the number. It can measure its total exposure, the fall it could survive without selling and its real borrowing cost; it cannot know the future excess return that sets the optimum.

What is the Kelly criterion for an investor?

It is the rule of choosing the exposure that maximises the expected logarithm of wealth, which over many periods means the fastest expected compounded growth. Busseti, Ryu and Boyd (2016) describe John Kelly's 1956 proposal as a way to allocate wealth across bets "so as to maximize the long term growth rate when the gamble is repeated". Investing a constant share of wealth in a stock index, period after period, is that kind of repeated gamble.

Expected log growth is not the average return. A gain of 50% followed by a loss of 50% averages zero but leaves 75 of every 100. The log of each period's wealth multiple adds up over time, so maximising its expected value maximises the expected rate at which wealth compounds. Our comparison of the Sharpe ratio and CAGR shows why the two measures can rank investments differently.

Maximising expected log growth is one objective, not the only reasonable one. It matches an investor with a specific, fairly high tolerance for swings over a very long horizon. A household that is more risk-averse, has a shorter horizon, needs money soon or sets a limit on how far it can fall can rationally hold less; in the same model, risk aversion γ divides the Kelly exposure, and half-Kelly corresponds to γ = 2.

The same authors warn that "with Kelly optimal bets there is a risk of the wealth dropping substantially from its original value before increasing". The criterion maximises growth; it does not limit losses. That is why the drawdown at each exposure matters as much as the growth rate below.

How does the formula choose an exposure?

For one stock index and cash, the expected log growth of an exposure f is approximately g(f) ≈ r + f(μ − r) − f²σ²/2, and it peaks at f* = (μ − r)/σ². Here r is the cash rate, μ − r the expected excess return of stocks over cash and σ their volatility. The first term rewards exposure in proportion; the second penalises it with the square of exposure.

With the 100-year averages as inputs, the excess return is 8.33% a year and the volatility 18.36%, so f* = 0.0833 ÷ 0.0337 ≈ 247%. If every borrowed dollar costs 1.5 points more than cash, the optimum falls to about 203%. These are historical averages plugged into a model that assumes normally distributed, independent log returns and continuous rebalancing.

The formula has two properties that matter for households, both before borrowing costs. At half of f*, expected growth keeps three quarters of the extra growth over cash. At twice f*, the extra growth is gone: the model's expected log growth falls back to the cash rate, with twice the volatility of the optimum. The leveraged-ETF analysis applies the same formula to the Nasdaq 100 and finds the multiple swinging with the sample.

What did each exposure actually compound to since 1926?

Realised growth rose from 3.29% a year in bills to 10.36% at 100% and peaked at 12.15% at 191%, then dropped faster than it had risen. We held each exposure constant from July 1926 to September 2026, resetting it at every month-end, and charged the Treasury-bill rate plus 1.5 points on borrowed money. The 1.5 points match the spread over its benchmark rate that one large online broker, Interactive Brokers, listed for its first tier of US-dollar and euro margin loans when we checked on 18 November 2026.

Compounded growth a year of constant exposures to the US stock market, July 1926 to September 2026, reset monthly, borrowing at the bill rate plus 1.5 points. 0% all cash: 3.29%. 50%: 7.22%. 95%, half the peak: 10.08%. 100%, the market: 10.36%. 150%: 11.77%. 191%, in-sample peak: 12.15%. 250%: 11.27%. 300%: 8.83%. 340%: 4.41%. From 347% the September 1931 fall wiped the account out.

July 1926 to September 2026, compounded growth a year: 0% 3.29%, 50% 7.22%, 95% 10.08%, 100% 10.36%, 150% 11.77%, 191% 12.15%, 250% 11.27%, 300% 8.83%, 340% 4.41%; from 347%, wiped out in September 1931. Source: Kenneth R. French Data Library, Fama/French 3 Factors (Mkt-RF and RF, file built from the September 2026 CRSP database); chart values (CSV), every exposure from 0% to 400% (CSV).

Near the top the curve was flat. Anything from 150% to 250% compounded within 0.9 points of the peak, so the exact optimum mattered little for growth. Beyond it each step cost more growth than the last: from 274% growth fell below that of 100%, from 344% below bills, and from 347% one month ended the account.

In September 1931 the market fell 28.74% more than bills. An account reset to 347% at the end of August 1931 lost all of its value in that month. A real broker would have closed the position earlier through a margin call, which also locks in the loss. The model is gentler than reality in one respect: it ignores falls inside each month.

Over the full period, $1 grew to about $97,932 at 191%, $19,567 at 100% and $15,228 at 95%, before taxes and trading costs. Over a century, small differences in yearly growth compound into large differences in these endpoints, and the windows below show how much they depend on the period. A historical CAGR is an observed outcome; it does not show what exposure maximises expected future growth.

What did the growth-maximising exposure cost in crashes?

Almost everything, once: at 191% the August 1929 to June 1932 fall took 98.3% of month-end wealth, which then needed a 5,773% gain to recover. At 100% the same fall was 83.7%, needing 512%. In the July 2001 to September 2026 window the worst fall was October 2007 to February 2009 at most exposures: 50.3% at 100% and 76.9% at 191%.

Worst fall of month-end wealth from its previous high, constant exposures reset monthly. July 1926 to September 2026 and July 2001 to September 2026: 0% all cash, minus 0.09% and 0%; 50%, minus 55.81% and minus 27.94%; 95%, minus 81.83% and minus 48.37%; 100%, minus 83.65% and minus 50.31%; 150%, minus 94.90% and minus 66.95%; 191%, minus 98.30% and minus 76.86%; 250%, minus 99.75% and minus 86.71%; 300%, minus 99.97% and minus 92.07%.

Worst month-end fall, 1926-2026 and 2001-2026: 50% −55.81% and −27.94%; 100% −83.65% and −50.31%; 191% −98.30% and −76.86%; 300% −99.97% and −92.07%. Source: Kenneth R. French Data Library; chart values with recovery gains and dates (CSV).

Time matters as much as depth for compounding. The longest stretch below a previous high was 22.8 years at 191%, against 15.3 years at 100% and 3.3 years in bills. A household that must sell during that stretch, to meet spending or a margin call, does not get the recovery the backtest assumes. Our guide to market crash recovery time explains why a deeper loss needs a disproportionately larger gain.

Margin rules turn this into a hard limit. For a US broker's margin account, Regulation T requires an initial margin of 50% on stocks, which caps exposure at 200% when a position is opened, and FINRA Rule 4210 requires 25% equity to be maintained. Under today's rules, an account reset to 200% at the end of September 1929 would have reached that 25% floor after a 33.3% fall; the market fell 31.8% by the close of 29 October, within 1.5 points of it.

Why does half-Kelly give up so little growth?

Because the growth curve is flat near its peak while losses keep deepening, so the first steps back from the optimum cost little growth. Half the in-sample peak, 95% exposure, compounded at 10.08% a year. That kept 6.8 of the 8.9 points of extra growth over bills that 191% earned, about three quarters, close to the formula's prediction.

In exchange, the worst fall was 81.8% instead of 98.3%, and the recovery gain needed was 450% instead of 5,773%. Those are very different outcomes for a household even though both look catastrophic. Half-Kelly also leaves room for error. In the formula, before borrowing costs, if the estimated excess return is twice the true one, full Kelly sits at twice the true optimum and expects only the cash rate, while half-Kelly lands on the true optimum.

Half-Kelly is not crash protection. Here it is close to simply owning the market, and it still lost 81.8% from 1929 to 1932 and 48.4% from 2007 to 2009. Busseti, Ryu and Boyd call fractional Kelly the best known of "several ad hoc methods" for limiting drawdown risk; their method instead adds an explicit limit on the probability of a drawdown. Our resilient allocation guide discusses that approach alongside other ways of limiting losses.

How stable is the growth-maximising exposure?

Not stable: across four 25-year windows the realised peak ranged from 116% to 285%, and the 100-year formula estimate carries an uncertainty band from about 141% to 354% before borrowing costs. The excess return is the weak input. Its 100-year average of 8.33% a year has a standard error of 1.83 points, so a rough 95% interval runs from 4.74% to 11.92%; that interval treats volatility as known and returns as independent.

Growth-maximising stock exposure by period, realised peak and formula estimate. July 1926 to June 1951: 116% and 116%. July 1951 to June 1976: 273% and 280%. July 1976 to June 2001: 237% and 271%. July 2001 to September 2026: 285% and 313%. Whole period 1926 to 2026: 191% and 203%.

Realised peak and formula estimate: 1926-1951 116% and 116%; 1951-1976 273% and 280%; 1976-2001 237% and 271%; 2001-2026 285% and 313%; 1926-2026 191% and 203%. Source: Kenneth R. French Data Library; peaks and inputs by period (CSV).

The first window contains the Depression, with 26.0% volatility; the later three had 14.1% to 15.5%.

The later peaks are model results that could not have been held in a US margin account. All three exceed the 200% cap of Regulation T, and each would have hit the 25% maintenance level inside a month at least once: 237% in October 1987, 273% in May 1970 and 285% in seven months, including September 2001 and October and November 2008.

Even at month-end, 273% and 285% fell 91.4% and 90.7% at their worst. Hindsight matters too: holding the 100-year peak of 191% through 1926-1951 would have compounded at 5.03% a year, below the 7.20% of plain 100%. Each window's peak is known only once the window has ended.

A fairer test fits the exposure on all earlier data and then holds it in the next window.

Window heldExposure fitted on earlier dataGrowth a year: fitted, half, 100%Worst fall: fitted, half, 100%
July 1951 to June 1976116% (fitted 1926-1951)10.81%, 7.70%, 10.21%−53.1%, −26.0%, −46.5%
July 1976 to June 2001151% (fitted 1926-1976)16.31%, 12.72%, 14.34%−44.0%, −22.7%, −29.9%
July 2001 to September 2026171% (fitted 1926-2001)13.03%, 8.76%, 9.81%−72.4%, −44.5%, −50.3%

Source: out-of-sample test (CSV).

In all three windows the earlier-fitted exposure grew faster than 100%, with deeper falls, and half of it grew more slowly with shallower falls. That result favours leverage in this sample only when borrowing is cheap.

Repeating the test at 2.5 points over bills, the fitted exposures beat 100% by 0.06, 0.91 and 1.79 points a year; at 4 points they were 100%, 100% and 111%, and beat the market by 0, 0 and 0.28 points. It is also three observations from a single national market, with overlapping fitting samples, so it cannot show how the same rule fared where stocks did worse; each fitted figure also sat far below the window's own peak.

What do fat tails, borrowing costs, taxes and rebalancing do to the answer?

Each one we could measure moved the optimum, and most moved it down. The realised peak sat about equal to or below the formula estimate in every window. The monthly excess returns had an excess kurtosis of 7.36, and 11 months fell more than three standard deviations below the mean where a normal distribution would predict about 1.6. That is consistent with fat tails costing growth at high exposure, though changing volatility could also contribute; this comparison cannot separate the two.

Borrowing costs bend the curve above 100%. With no spread over bills the realised peak was 225% at 13.98% a year; at 1.5 points, 191%; at 2.5 points, the spread listed for the same broker's IBKR Lite plan, 166% at 11.27%.

At 3 points it was 153%, at 4 points 127%, and at 5 points 100%: from there on, borrowing to hold more of the market added no growth over this century. A household's actual borrowing rate is the input that matters, and a broker's benchmark rate can sit above the bill rate used here.

Rebalancing has a cost and a direction. Charging 0.10% on every amount traded at month-end moved the peak to 188% and cut its growth to 12.06%. Above 100%, the reset sells after falls and buys after rises; below 100%, it does the reverse. Resetting daily instead of monthly gave a peak of about 215% on a 5-point grid, so the frequency also changes the answer.

Taxes are not modelled. In a taxable account the monthly reset realises gains, while a 100% index holding need not sell at all, and the tax treatment of interest on borrowing differs by country. All figures are nominal; deflating by a price index lowers every exposure's log growth by the same amount and would not move the peak.

What can a household copy from Kelly sizing?

The method, applied with its own inputs and limits; not a full-Kelly figure, which this history shows is unstable and was reached through near-total losses. The checks below are educational and use your own figures.

  1. Count total exposure. Add direct shares, equity funds, the multiple of any leveraged fund and any borrowing, and divide by investable wealth.
  2. Write down the fall you could hold through. Compare the worst falls above with the cash you would need to keep spending and contributing without selling.
  3. Price your borrowing. Use your actual margin or loan rate, not the cash rate, and compare its spread over bills with the roughly 5 points at which borrowing stopped adding growth here; check the maintenance level at which a broker can sell your positions.
  4. Test a lower expected return. Re-run the formula with an excess return several points below the historical 8.33%; in the formula, the optimum falls in proportion.
  5. Count your rebalancing rule and taxes. Decide how often exposure is reset and what each reset costs after tax in your account.
  6. Check whether the rest of the portfolio falls at the same time. More of one index is not a second return engine.

A tail-hedging allocation like Spitznagel's rests on the same geometric-growth argument: it keeps most of the portfolio in stocks and asks whether a small, costly hedge raises whole-portfolio compounding. That question needs the hedge's costs and payoffs, which a single-index Kelly calculation does not contain. Copio has not identified an optimal exposure, and the figures above are not a recommended level of exposure or borrowing.


Sources and method

Checked on 18 November 2026. Market data: Kenneth R. French Data Library, Fama/French 3 Factors, monthly and daily files built from the September 2026 CRSP database, July 1926 to September 2026 (1,203 months). Mkt-RF is the value-weighted return of CRSP firms incorporated in the US and listed on the NYSE, AMEX or Nasdaq, minus the one-month Treasury bill rate RF (Ibbotson Associates to May 2024, then the ICE BofA US 1-Month Treasury Bill Index).

We used it because one primary file gives monthly market and matched cash returns from 1926, long enough to include several deep crashes. A world or euro view would need a different and shorter series, so the results are a US, US-dollar view. Kelly problem and drawdown quotes: Busseti, Ryu and Boyd, Risk-Constrained Kelly Gambling (arXiv 1603.06183, 2016), sections 1 and 2.2.

Borrowing spread: Interactive Brokers' margin-rates page (interactivebrokers.com, Trading, Margin Rates), retrieved 18 November 2026: benchmark plus 1.5% for the first USD 100,000 and EUR 90,000 on IBKR Pro, benchmark plus 2.5% on IBKR Lite. Margin rules: Regulation T, 12 CFR 220.12 (50% for margin equity securities) and FINRA Rule 4210(c)(1) (25% maintenance).

All growth rates, drawdowns, peaks, intervals and out-of-sample results are copio calculations. Exposure f is reset each month-end: return = RF + f × (Mkt − RF) − max(f − 1, 0) × spread ÷ 12. CAGR is (ending ÷ starting wealth)^(1 ÷ years) − 1; drawdowns use month-end wealth; peaks use a 1-point grid from 0% to 500%. Formula inputs are 12 times the monthly mean and variance of excess returns; the formula is shown as exp(g) − 1.

Half the 191% peak is 95.5%; we report the 95% row. The 95% interval for f* divides the interval for the mean excess return by the full-period variance, before borrowing costs. The out-of-sample test chooses the realised peak on all months before each window and holds it unchanged inside the window. The 0.10% trading cost is an assumption, not a quoted price.

Limits: all results are historical and in-sample unless labelled out-of-sample; nominal US dollars; before taxes, fund fees and, unless stated, trading costs. The spread is today's published rate applied to the whole century, and the T-bill rate stands in for a broker's benchmark rate. Month-end resets ignore falls within the month and margin calls; one country's market is a single history. Downloads: chart values, all exposures, peaks by period, out-of-sample test.

This article is educational, not personalised investment advice.

FAQ

What is the Kelly criterion in investing?

It is the rule of choosing the exposure that maximises the expected logarithm of wealth, which is the same as maximising expected long-run compounded growth. For one stock index and cash, a common approximation gives the growth-maximising exposure as the expected excess return over cash divided by the variance of returns. The answer is only as good as those two inputs, and the expected return is the one nobody observes in advance.

What stock exposure maximised growth in US history?

In our calculation on the Kenneth French data from July 1926 to September 2026, a constant 191% exposure to the US market, reset monthly and financed at the Treasury-bill rate plus 1.5 points, compounded fastest, at 12.15% a year against 10.36% at 100%. That is a historical, in-sample result: the same mix fell 98.3% from August 1929 to June 1932.

Why do investors use half-Kelly?

Because growth changes little near the peak while losses keep deepening. In the approximation, half the optimal exposure keeps three quarters of the extra growth over cash. In the 1926-2026 data, 95% exposure kept 6.8 of the 8.9 points of extra growth a year that 191% earned, and its worst fall was 81.8% instead of 98.3%. Half of a wrong estimate is still wrong, and 81.8% is not crash protection.

How reliable is a Kelly estimate for the stock market?

Not very. In four 25-year windows since 1926, the realised growth-maximising exposure was 116%, 273%, 237% and 285%. A rough 95% interval around the 100-year estimate of the excess return gives formula exposures from 141% to 354% before borrowing costs.

Does Kelly sizing protect a portfolio against crashes?

No. It maximises expected log growth under assumptions; it does not limit losses. In our 1926-2026 calculation every exposure from 15% to 300% had its worst fall in the same August 1929 to June 1932 decline, and from 347% the September 1931 fall wiped the account out. Drawdown limits have to be added as a separate constraint.

Optional cookies help us measure visits and advertising. Choose each purpose, or decline both. You can change your choice using Cookie settings in the footer. Cookie policy.