Skip to content
Menu
Finabase Learn

Trading · Risk management

Kelly criterion: the fastest-growing bet size, and why to halve it

The Kelly formula worked through for trades with a win rate and a payoff, the growth curve on both sides of it, and what an overestimated edge does to a full Kelly bet.

On this page
  1. The formula
  2. Why there is a best size
  3. What full Kelly feels like
  4. Why traders halve it
  5. Kelly beside the usual risk rules
  6. How Finabase does it
  7. Frequently asked questions
  8. Summary
  9. Check your understanding

The Kelly criterion is a formula for how much of an account to risk on a bet when the odds are in your favor. It picks the fraction that makes the account grow fastest over a long series of bets. Risk less and the account grows more slowly. Risk more and, past a point, it also grows more slowly, then shrinks, even though every single bet has a positive expected value.

That second fact is the reason to learn the formula even if you never bet a Kelly fraction. It proves there is such a thing as too large a bet on a good trade, and it says where the limit is. This lesson derives the fraction for a trade with a win rate and a payoff, draws the growth curve on both sides of it, and then shows the part that matters most in markets: what happens when the edge you fed into the formula was too optimistic.

The formula

John Kelly, a researcher at Bell Labs, published the result in 1956 in a paper about information transmitted over a noisy line. His example was a gambler with a private wire giving advance but imperfect news of race results. Edward Thorp later applied the idea to blackjack and then to securities.

For a bet with two outcomes the formula is short.

Notice what "risk" means here. The fraction is the share of the account lost when the bet fails, the same quantity as risk per trade in Position sizing: how many shares to buy for a set risk. It is not the size of the position. A trade that risks 10% of an account with a stop 5% below the entry is a position worth twice the account.

Why there is a best size

The reasoning rests on one observation: an account compounds. Results multiply, they do not add. A 50% gain followed by a 50% loss leaves 75% of the starting money, not 100%.

When results multiply, what matters over many bets is the average of the logarithm of each result, which is the growth rate. For a bet risking fraction f the account is multiplied by (1 + b × f) on a win and by (1 − f) on a loss, so:

The Kelly fraction is the f that makes g largest. The figure draws g for the first example, a 55% win rate at even money.

Growth per trade by fraction risked: a bet that wins 55% of the time at even money −3% −2% −1% 0% 1% 0% 5% 10% 15% 20% 25% 30% Half Kelly 5% Kelly 10% expected log growth per trade peak: 0.50% a bet about 2× Kelly: growth is zero Fraction of the account risked on each bet Long-run growth per bet
Expected logarithmic growth per bet by fraction risked, for a bet that wins 55% of the time at even money. The peak is the Kelly fraction. Calculation by Finabase.
Fraction risked Multiple of Kelly Growth per bet Share of the maximum Typical result of 100 bets
2.5% 0.25× 0.22% 44% ×1.24
5% 0.5× 0.38% 75% ×1.46
10% 1× 0.50% 100% ×1.65
15% 1.5× 0.37% 75% ×1.45
20% 2× −0.01% below zero ×0.99
25% 2.5× −0.67% negative ×0.51

The curve is nearly symmetric around the peak, and that symmetry is the first practical lesson. Risking 5% and risking 15% produce the same long-run growth. The 15% bettor takes three times the swings for nothing. At 20% the growth is gone. At 25% a bet with a 10% edge, repeated, halves the account about every hundred bets.

Every bet in that last row has a positive expected value. Overbetting does not turn a good bet into a bad one. It turns a good series into a losing one, because the losses, taken on a larger share of the account, cost more than the wins add back.

What full Kelly feels like

Maximum growth comes with a price the formula does not show. For a bet repeated in very small steps, Thorp gives a simple rule: at full Kelly the probability that the account ever falls to a fraction x of its starting value is x. There is a 50% chance of being cut in half at some point and a 20% chance of being cut to one fifth. A bettor at half Kelly faces x cubed: a 12.5% chance of halving.

A simulation of the 55% bet agrees closely. Each setting below was run 6,000 times for 500 bets.

Share of 6,000 simulated accounts that fell to half at some point in 500 bets 1% halved; median ×3.0 0.25× Kelly 11% halved; median ×6.5 0.5× Kelly 45% halved; median ×12 1× Kelly 71% halved; median ×6.5 1.5× Kelly 86% halved; median ×0.9 2× Kelly Bet wins 55% at even money (Kelly fraction 10%). "Median ×" is the middle account after 500 bets. Fixed seed.
Share of 6,000 simulated accounts that fell to half their starting value at some point in 500 bets, by multiple of the Kelly fraction. The bet wins 55% at even money. Simulation by Finabase.
Multiple of Kelly Median account after 500 bets Worst 5% of runs ended below Fell to half at some point
0.25× ×3.0 ×1.16 0.7%
0.5× ×6.5 ×1.08 10.6%
1× ×12.2 ×0.33 45.4%
1.5× ×6.5 ×0.02 70.6%
2× ×0.9 ×0.001 86.0%

Full Kelly had the highest median, as promised. It also left one run in twenty with a third of the starting money after 500 bets with a 10% edge. Half Kelly gave up about half of the median outcome and in exchange its worst 5% of runs still finished ahead.

Why traders halve it

The curve and the simulation assume the win rate and payoff are known exactly. A casino game meets that assumption. A trading method does not. Its win rate is an estimate from a limited number of past trades, and the future may differ from the past.

The size of the uncertainty is easy to state. A win rate of 55% measured on 100 trades has a standard error of about 5 percentage points. On 400 trades it is still 2.5 points. The true figure could plausibly be 50% or 60%.

Now put that through the growth formula. Suppose you believe 55% and bet accordingly, and the truth is lower.

True win rate True Kelly fraction Growth betting 10% ("full Kelly") Growth betting 5% ("half Kelly")
55% 10% +0.50% per bet +0.38%
53% 6% +0.10% +0.18%
52% 4% −0.10% +0.08%
50% 0% −0.50% −0.13%

A three-point overestimate, well inside the error of a 100-trade sample, turns the believed full Kelly bet into 2.5 times the true Kelly fraction, and a profitable method into a losing account. The half bet is still growing. That asymmetry is the argument for fractional Kelly in one table: betting too little costs a little growth, and betting too much can cost everything.

Three more reasons point the same way.

  • Fat tails. The formula above allows one size of loss. Real trades sometimes lose two or three times the planned amount when a price gaps through a stop (Gap risk: how often stocks open through a stop, and what it costs). A larger worst case lowers the true Kelly fraction.
  • Several bets at once. The formula sizes one bet at a time. Positions held together and correlated with each other act as one larger bet (Correlated positions: when five trades are really one bet).
  • The trader. A strategy sized so that a halving is a coin flip will, sooner or later, be abandoned at the bottom. A smaller size that is actually followed beats an optimal one that is not.

Thorp, who used the criterion with real money for decades, wrote that many practitioners prefer fractions of the Kelly bet for these reasons. Samuelson's objection goes further: maximizing growth is one possible goal, not the only rational one, and a person who cares about avoiding large losses can rightly choose much less.

Kelly beside the usual risk rules

The common trading rule of risking 1% or 2% per trade can be read against the formula. A method with a Kelly fraction of 10% that is traded at 1% risk is running at one tenth of Kelly. That looks timid until the estimate error, the gaps and the correlated positions are counted. It also explains why the rule works for most people: almost any method with a real edge has a Kelly fraction above 2%, so the rule keeps nearly everyone on the left-hand, forgiving side of the curve.

Approach Sets the size from Needs to know Main weakness
Fixed fraction (1% to 2%) A chosen risk per trade Stop distance Ignores how good the edge is
Full Kelly Win rate and payoff The true odds Ruinous if the odds are overestimated
Half or quarter Kelly A fraction of the Kelly bet The odds, roughly Still relies on an estimate
Volatility targeting Recent volatility A volatility estimate Volatility can jump before the size adjusts

Position sizing inside a strategy: how much, not what shows the same curve drawn from real data: one list of QQQ signals run at nine sizes, with growth peaking and then collapsing as leverage rises.

How Finabase does it

Finabase does not compute a Kelly fraction and no tool suggests a bet size.

  • Inputs. The Backtest tool reports the number of trades, wins, losses and win rate for signals on one symbol, with every signal trading a fixed $10,000 and no costs. The average win and loss must be worked out from its transactions table. Remember the standard error: a win rate from 40 trades is a rough guess.
  • Practice. The paper account holds $10,000 of virtual cash and fills at live prices (Practise with Finabase paper trading on the web). Running the same rules at two sizes for the same weeks shows the difference in swings more convincingly than a table.

Frequently asked questions

What is the Kelly criterion in simple terms?

It is the bet size that makes money grow fastest over many repeated bets with a known edge: the edge divided by the odds. Bet less and growth is slower but steadier. Bet more than about twice that size and the account shrinks in the long run, however good each bet is.

Why use half Kelly?

Because the edge is never known exactly. Half Kelly keeps about three quarters of the maximum growth, cuts the chance of halving the account from about 50% to about 12%, and still grows if the real edge turns out to be only half of what was estimated.

Can the Kelly fraction be more than 100%?

For a win-or-lose bet it cannot exceed the win probability. For a continuous investment such as a stock index, the corresponding formula (expected excess return divided by variance) can give a figure above 100%, which means borrowing. Such results are very sensitive to the return assumed, which is the least certain number in finance.

Is the Kelly criterion the same as optimal f?

They are close relatives. Optimal f, described by Ralph Vince in 1990, finds the fraction that would have maximized growth over a specific list of past trades, scaled by the largest loss in that list. It reaches the same answer as Kelly for simple bets and shares the same weakness: the best fraction for the past is not known to be the best for the future.

Summary

The Kelly criterion gives the fraction of an account to risk on a favorable bet so that the account grows as fast as possible: win probability minus loss probability divided by the payoff. Its growth curve rises to that point and falls beyond it, reaching zero at about twice the Kelly fraction. Full Kelly betting is rough, with about an even chance of halving the account at some point, and it depends on knowing the odds exactly. Trading odds are estimates from samples, and overestimating them by a few points turns a full Kelly bet into an overbet that loses over time. That is why practitioners bet a half or a quarter of the calculated fraction, and why the plain 1% to 2% rule keeps most traders on the safe side of the curve.

Check your understanding

4 questions on this lesson. Each answer comes with the reason, and nothing is scored.

  1. A system wins 50% of its trades and its average win is twice its average loss. What is the Kelly fraction?

    1. 25%
    2. 50%
    3. 12.5%
    Show the answer

    A. 25% f = p − q ÷ b = 0.50 − 0.50 ÷ 2 = 0.25. The fraction is the share of the account lost when a trade fails, not the size of the position.

  2. A bet has a Kelly fraction of 10%. What does the growth curve say about risking 20% on each bet instead?

    1. The account grows twice as fast
    2. Long-run growth is about zero, with far larger swings
    3. Each bet now has a negative expected value
    Show the answer

    B. Long-run growth is about zero, with far larger swings Growth peaks at the Kelly fraction and returns to zero at about twice that size. Each bet still has a positive expected value; it is the compounding of larger losses that removes the growth.

  3. You estimate a 55% win rate at even money from 100 trades and bet the full Kelly fraction of 10%. The true win rate is 52%. What happens over many bets?

    1. The account still grows, only more slowly
    2. The account shrinks, because 10% is 2.5 times the true Kelly fraction of 4%
    3. Nothing changes: Kelly adjusts itself
    Show the answer

    B. The account shrinks, because 10% is 2.5 times the true Kelly fraction of 4% At a true 52% the Kelly fraction is 4%. Betting 10% is beyond twice Kelly, where growth is negative. Betting half the believed fraction (5%) would still have grown. This asymmetry is the case for fractional Kelly.

  4. What does half Kelly give up, and what does it buy?

    1. It gives up about half the growth rate and removes all drawdowns
    2. It gives up about a quarter of the growth rate and cuts the chance of halving the account from about one in two to about one in eight
    3. It gives up nothing; the curve is flat near the peak
    Show the answer

    B. It gives up about a quarter of the growth rate and cuts the chance of halving the account from about one in two to about one in eight Half Kelly keeps about 75% of the maximum growth rate. In the continuous-time rule the chance of ever falling to half is 50% at full Kelly and 12.5% at half Kelly; the simulation gave 45% and 11%.

Sources and further reading

  1. J. L. Kelly Jr., “A New Interpretation of Information Rate,” Bell System Technical Journal 35(4), 1956
  2. Edward O. Thorp, “The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market,” Handbook of Asset and Liability Management, vol. 1 (2006)
  3. L. C. MacLean, E. O. Thorp and W. T. Ziemba, “Long-term capital growth: the good and bad properties of the Kelly and fractional Kelly capital growth criteria,” Quantitative Finance 10(7), 2010
  4. Paul A. Samuelson, “The ‘Fallacy’ of Maximizing the Geometric Mean in Long Sequences of Investing or Gambling,” Proceedings of the National Academy of Sciences 68(10), 1971
  5. Finabase calculation and simulation — 6,000 runs of 500 bets per setting, fixed seed (scripts/learn/figures/trading-risk-depth.ts)