The formula needs two numbers: how often a strategy wins, and how large its wins are compared with its losses. From those it gives a fraction of capital. The trouble is that both numbers are estimates, usually from a backtest or a short live record, and Kelly punishes overestimates harshly: bet too much and long-run growth falls, bet twice the true figure and it falls to nothing. This guide explains the formula, where it came from, two worked examples, why full Kelly is too aggressive in practice, how fractional Kelly works, how to estimate the inputs honestly, how it applies to option strategies and Indian lot sizes, and how to turn the result into a number of lots.
The Kelly criterion at a glance
| Part | What it means |
|---|---|
| The question | What share of capital should each trade risk? |
| The inputs | W, the chance of winning; R, the average win divided by the average loss |
| The formula | f* = W − (1 − W) ÷ R |
| A result of zero or less | No edge on these estimates: Kelly says do not trade |
| Full Kelly | The fastest long-run growth, with very deep drawdowns |
| Half Kelly | About three quarters of the growth, with much shallower drawdowns |
| In practice | A fraction of Kelly, capped at a fixed maximum such as 1% to 2% of capital |
A short history of the Kelly criterion
John Larry Kelly Jr. was a physicist working at Bell Labs, the research arm of the American telephone company, alongside Claude Shannon, the founder of information theory. Shannon had shown how much information a noisy telephone line could carry. In 1956 Kelly asked a related question: a gambler receives tips over a noisy line, tips that are right more often than not but not always. How much should the gambler bet on each tip to make their money grow fastest? His answer, published as "A New Interpretation of Information Rate", is the Kelly criterion.1 The best fraction to bet turned out to be tied to the information in the tips: the better the information, the larger the bet.
Edward Thorp, a mathematician, read the paper and used it to size bets in blackjack, where card counting gives a player a small, measurable edge at some moments in the deck. His book *Beat the Dealer* (1962) made card counting famous, and Thorp later applied the same sizing ideas to investing through his hedge fund.2 In 1961 the statistician Leo Breiman proved that, over many bets, the Kelly strategy ends up ahead of any essentially different strategy and reaches a given target fastest.3
Not everyone was convinced. The economist Paul Samuelson argued that maximising long-run growth is not the right goal for every investor, because it ignores how much a person dislikes the deep falls along the way. He made the point in 1979 in a paper written almost entirely in words of one syllable, to show how simple he thought the error was.4 The debate is still useful: Kelly tells you the size that grows wealth fastest in the long run, not the size that lets you sleep or survive a bad year.
The story of Kelly, Shannon, Thorp and their critics is told in William Poundstone's *Fortune's Formula* (2005),5 and the academic work is collected in *The Kelly Capital Growth Investment Criterion* by Leonard MacLean, Edward Thorp and William Ziemba (2011).6
The Kelly formula
For a bet that wins a fixed amount or loses a fixed amount, the Kelly fraction is:
f* = W − (1 − W) ÷ R
where:
- f* is the fraction of capital to risk on each trade,
- W is the probability of winning,
- R is the ratio of the average win to the average loss.
If the result is zero or negative, the strategy has no edge on these estimates and Kelly says do not trade it.
The formula has an intuitive reading. W is what you gain from winning often; (1 − W) ÷ R is what you give back through losses, scaled down when wins are large compared with losses. The fraction is what is left: the size of your edge relative to the size of your wins.
Why growth, not average profit
Kelly maximises the growth rate of capital, not the average profit of each trade. The difference matters because gains and losses compound. A 50% loss followed by a 50% gain leaves you at 75% of where you started, not back where you began. Betting a large share on a favourable bet raises the average profit per bet, but it also raises the chance of large falls that compounding cannot easily repair. Kelly finds the size where these two forces balance: below it you grow more slowly than you could; above it, the damage from losses grows faster than the benefit of larger wins.
The version for trading returns
For a strategy whose results are spread out rather than a simple win-or-lose, an approximate form uses the average return and its variance: the fraction of capital to commit is roughly the expected return per period divided by the variance of returns per period. For a strategy that earns on average 1% a month with a monthly standard deviation of 5%, the variance is 0.0025, and the formula gives 0.01 ÷ 0.0025 = 4: four times capital, which is leverage no sensible retail trader would use. This version shows how quickly Kelly asks for more than is wise when the edge looks good on paper.
A worked example: a breakout strategy
A strategy's tested results show:
- Win rate: 55%, so W = 0.55.
- Average win: ₹1,200. Average loss: ₹1,000. So R = 1.2.
f* = 0.55 − 0.45 ÷ 1.2 = 0.55 − 0.375 = 0.175
Full Kelly says risk 17.5% of capital per trade. On a ₹10 lakh account, that is ₹1,75,000 at risk on each trade, where "at risk" means the amount lost if the trade hits its stop.
That number should feel alarming. It is mathematically optimal for long-run growth under perfect knowledge, and reckless in practice. Five losses in a row, which a 55% strategy will produce from time to time, would take about 62% off the account at full Kelly.
A worked example: a defined-risk option strategy
This is a hypothetical example.
A trader sells a weekly Nifty iron condor: the short strikes 300 points from the index on each side, and the long wings 100 points beyond them. One lot is 65 units.
- The credit received is about ₹30 a unit, or ₹1,950 a lot.
- The most the position can lose is the wing width minus the credit: (100 − 30) × 65 = ₹4,550 a lot, before costs.
- The trader closes at 50% of the credit when that is possible, so an average win is about ₹975 after costs, and an average loss, with a stop before the maximum, about ₹3,000.
- Over the test, 76% of trades won.
R = 975 ÷ 3,000 = 0.325, so f* = 0.76 − 0.24 ÷ 0.325 = 0.76 − 0.74 = 0.02.
Kelly says the edge is thin: about 2% of capital at risk per trade at full Kelly. On a ₹10 lakh account, that is ₹20,000, or about four lots at a ₹4,550 maximum loss. At half Kelly it is two lots. If the true win rate were 74% instead of 76%, f* would be 0.74 − 0.26 ÷ 0.325 = −0.06: no edge at all. A strategy whose Kelly fraction turns negative when the win rate drops two points is fragile, and the result is a warning as much as a size.
Why full Kelly is too aggressive
Your estimates are uncertain
Kelly assumes you know W and R exactly. In trading you only have estimates from a backtest or a short live record. If your true win rate is 52% instead of 55%, the correct Kelly fraction in the first example falls from 17.5% to 12%. Betting more than the true Kelly fraction reduces long-run growth, and betting twice the true Kelly fraction drives expected growth to zero. Because the penalty for overbetting is much worse than the penalty for underbetting, uncertainty about the inputs is a strong reason to bet less.
Drawdowns are severe
Even with perfect estimates, full Kelly produces deep drawdowns. In the theoretical model, a bettor at full Kelly has about a 50% chance of seeing their capital fall to half its peak at some point. At half Kelly, that chance falls to about 12.5%, while long-run growth falls by only about a quarter. Few traders would continue a strategy through a halving of their account, so the growth that full Kelly promises is rarely the growth anyone actually collects.
Trades are not independent
Kelly assumes each bet is independent of the last. Trading results often cluster: losses come together in volatile periods, and several strategies may lose on the same day. Clustered losses make drawdowns worse than the formula expects.
Gaps and fat tails
The simple formula treats the loss on each trade as fixed at the stop. Markets gap. A stock can open far below its stop after results, and an index can jump on news overnight. Losses beyond the stop are rare but large, and full Kelly sizing leaves no room for them.
Fractional Kelly
Most practitioners use a fraction of the Kelly bet, such as half or a quarter:
| Fraction | Risk per trade in the first example | Growth compared with full Kelly | Chance of halving from a peak (theory) |
|---|---|---|---|
| Full Kelly | 17.5% | 100% | About 50% |
| Half Kelly | 8.75% | About 75% | About 12.5% |
| Quarter Kelly | 4.4% | About 44% | Well under 1% |
Even quarter Kelly here is far above the 1% to 2% risk per trade that many traders use. That gap is a reminder that Kelly is a ceiling, not a target.
The growth figures follow from a simple rule of the model: betting a fraction c of the Kelly amount gives c × (2 − c) of the full-Kelly growth rate. Half Kelly gives 0.5 × 1.5 = 75%; quarter Kelly gives 0.25 × 1.75, about 44%. Betting twice Kelly, c = 2, gives zero.
Estimating the inputs honestly
The quality of a Kelly figure is the quality of W and R. Some ways to keep them honest:
- Use trades the strategy was not designed on. A backtest on the data used to pick the rules overstates both W and R. See avoiding overfitting.
- Use results after costs. Brokerage, STT, exchange charges, stamp duty and slippage reduce every win and enlarge every loss.
- Count enough trades. With 100 trades and a measured win rate of 55%, the standard error of the win rate is about 5 percentage points, so the true rate could plausibly be anywhere from about 45% to 65%. At 45% the first example's Kelly fraction is negative.
- Use a conservative estimate. A simple habit is to use the lower end of the plausible range for W, and a slightly lower R than measured.
- Check the inputs against live results once the strategy trades, and recalculate as the record grows.
A table of Kelly fractions
| Win rate | R = 0.5 | R = 1.0 | R = 1.5 | R = 2.0 |
|---|---|---|---|---|
| 40% | No edge | No edge | 0.0% | 10.0% |
| 50% | No edge | 0.0% | 16.7% | 25.0% |
| 60% | No edge | 20.0% | 33.3% | 40.0% |
| 70% | 10.0% | 40.0% | 50.0% | 55.0% |
| 80% | 40.0% | 60.0% | 66.7% | 70.0% |
The table shows how quickly full Kelly reaches sizes that no trader should use, and how small changes in win rate or R move the answer. It is a map of the edge, not a set of sizes.
Applying Kelly to trading strategies
- Use out-of-sample results for W and R, not the best backtest.
- Use costs-included figures for wins and losses.
- Shrink the estimate. Assume your edge is smaller than the backtest shows.
- Take a fraction, such as a quarter of the result.
- Cap it. Apply a maximum risk per trade, such as 2%, whatever the formula says.
- Check with Monte Carlo. Simulate the drawdowns at your chosen size. See Monte Carlo simulation.
- Translate to lots. Divide the rupee risk by the loss per lot at the stop, and round down.
Translating to position size
Suppose quarter Kelly with a 2% cap gives 2% risk on a ₹10 lakh account: ₹20,000. If a trade's stop would lose ₹6,000 per lot, the position is three lots (₹18,000 at risk), because four would exceed the limit. The position size calculator does this step.
Lot sizes make sizing coarse
Indian derivatives trade in fixed lots, and lots are large. A Nifty option lot is 65 units, and SEBI raised the minimum value of index derivative contracts from 20 November 2024.7 For a small account, the smallest possible position may already be above the Kelly fraction. If one lot of a strategy risks ₹6,000 and the account is ₹2,00,000, one lot is 3% risk, above a 2% cap. The honest answers are to trade a smaller-risk version of the strategy, such as a narrower spread, to wait until the account is larger, or not to trade it. Taking the one lot anyway, because it is the smallest available, quietly sets the risk above the limit the trader chose.
A worked example: sizing a ₹3 lakh account step by step
This is a hypothetical example.
A trader with ₹3,00,000 runs a Bank Nifty option-buying breakout. The out-of-sample test, after costs, has 140 trades: 38% wins, an average win of ₹4,400 and an average loss of ₹1,900 a lot.
- Kelly fraction. R = 4,400 ÷ 1,900 = 2.32. f* = 0.38 − 0.62 ÷ 2.32 = 0.38 − 0.267 = 0.113, or 11.3%.
- Shrink the inputs. With 140 trades, the standard error of the win rate is about 4 points. Using 34% instead of 38%, and R of 2.1: f* = 0.34 − 0.66 ÷ 2.1 = 0.34 − 0.314 = 0.026, or 2.6%. The edge shrinks sharply with modest changes, which is a sign to be careful.
- Take a fraction. Half of the shrunk figure is 1.3%.
- Apply the cap. The trader's cap is 2%, so 1.3% stands: ₹3,900 at risk per trade.
- Translate to lots. A losing trade loses about ₹1,900 a lot on average, but the stop can lose up to ₹2,600 a lot on a fast move. ₹3,900 ÷ ₹2,600 = 1.5, so one lot.
- Check the streak. At one lot, a run of eight losses at ₹1,900 each costs ₹15,200, about 5% of the account. The trader can sit through that.
The full-Kelly figure of 11.3% would have suggested more than a dozen lots. The honest answer is one.
What the Kelly fraction tells you about a strategy
Even if you never size by Kelly, the number is a useful diagnostic:
- A large Kelly fraction from a short test usually means the test is too optimistic, not that the strategy is extraordinary.
- A Kelly fraction that changes sign with a small change in inputs means the edge is thin and fragile.
- A Kelly fraction below your fixed risk per trade means you are betting more than your edge justifies, even before uncertainty.
- A Kelly fraction that falls month by month as live results come in suggests the edge is fading.
Writing the Kelly figure next to each strategy, and updating it as results come in, turns it into a simple health check. It costs nothing to calculate, it uses numbers you already have from your backtests and your live record, and it gives an early, objective warning when a strategy's edge starts to slip, often well before the drawdown itself becomes obvious on the equity curve.
Kelly and losing streaks
Kelly does not reduce the number of losing trades; it reduces how much each one costs. A 40% strategy will lose eight times in a row now and then, whatever the size. At full Kelly those eight losses might take half the account; at a quarter of Kelly, a modest share. The streak itself is a feature of the strategy. The size decides whether you are still trading when it ends.
Kelly for options strategies
Option strategies often have a high win rate and a large average loss, such as credit spreads. Kelly handles this through R. For example, a strategy that wins 75% of the time with an average win of ₹4,000 and an average loss of ₹12,000 has R = 1/3:
f* = 0.75 − 0.25 ÷ (1/3) = 0.75 − 0.75 = 0
Kelly says there is no edge: the high win rate exactly balances the large losses. Any costs would make it negative. This is a useful check on strategies that look attractive because they win often.
For defined-risk strategies, use the realistic loss on a losing trade, not only the maximum loss. Many traders exit before the maximum, so the average loss is smaller; but on gap days the maximum is what you lose. Check the Kelly figure with both, and size for the worse one.
For undefined-risk strategies, such as a naked short straddle, the largest losses are much larger than the average. A Kelly figure based on the average loss will be too large. These strategies need sizing by the loss on a severe day, not by Kelly alone.
Kelly across several strategies
Most traders run more than one strategy. Kelly for each one on its own overstates the total, because strategies that trade the same index often lose on the same days. Two Nifty option-selling strategies that each deserve 5% at quarter Kelly do not together deserve 10%: on a sharp move they behave like one strategy. A practical rule is to set a total risk budget for the account, for example 4% across everything open at once, and to share it between strategies according to how differently they behave. Strategies on different instruments, or with opposite exposures, can share a budget more freely than strategies on the same index.
Kelly vs fixed-fraction sizing
| Kelly | Fixed fraction | |
|---|---|---|
| Basis | Your estimated edge | A risk level you choose, such as 1% |
| Adapts to edge | Yes | No |
| Sensitivity to estimates | High | Low |
| Typical use | As a ceiling or a guide | As the everyday rule |
Many traders use fixed-fraction sizing day to day and use Kelly to check that their fraction is well below the theoretical maximum. If the Kelly figure is close to the fixed fraction, the edge is thin and the size should come down.
Sizing and limits in QuantPower
QuantPower does not size trades for you, and it does not tell you which strategy to trade. It gives you the numbers and the controls to apply your own sizing:
- The inputs. A backtest reports the win rate, the average winning and losing day or trade, the largest loss, the maximum drawdown and the results by month, all after charges. These are the W and R you need, from data you can check trade by trade.
- The size. Each algo has a quantity multiplier from 1 to 50 that scales every leg's lots, and rupee targets and stops scale with it.
- The caps. A daily loss limit refuses new orders once the day's loss reaches it, Smart Exit squares off every position and stops the algos at a set loss or profit for the whole account, and a cap on orders a day limits how much the account can trade. These hold the account inside its budget whatever an individual strategy does.
- The test. Forward test runs the same algo on live prices with test money, so you can watch the drawdowns at your chosen size before real money is involved.
Common mistakes
- Using full Kelly. It assumes perfect knowledge and produces drawdowns few traders can bear.
- Using in-sample results. The best backtest overstates the edge, so the Kelly figure is too large.
- Ignoring costs. A small edge before costs can be zero after them.
- Treating each strategy separately. Strategies that lose together need one shared budget.
- Rounding up lots. If the size works out to 2.6 lots, it is 2 lots.
- Not recalculating. As live results come in, the inputs change, and so does the right size.
Where position sizing is going
Two trends are clear. Testing tools have made W and R easy to measure, which makes Kelly easy to calculate and also easy to misuse on overfitted results; the growing use of out-of-sample tests and Monte Carlo checks is the counterweight. And account-level risk controls have become standard: limits on daily loss, total open risk and the number of orders sit above every strategy. Kelly works best inside that frame, as a guide to how much of a fixed risk budget each strategy deserves, rather than as a free-standing rule.
Start with 500 free tokens
Connect your broker, build a strategy and test it on paper before a single live order goes out.
Frequently asked questions
What is the Kelly criterion in simple words?
A formula that suggests what share of your capital to risk on each trade, based on your win rate and how large your wins are compared with your losses.
Should I use full Kelly?
Almost never. Your estimates are uncertain and full Kelly produces deep drawdowns. Most practitioners use half or less, with a hard cap.
What if Kelly gives a negative number?
The strategy has no edge on your estimates, and Kelly says not to trade it.
Does Kelly work for intraday trading?
The maths applies to any repeated bet, but intraday strategies often have small edges and high costs, which make the Kelly fraction small or zero. Calculate it on results after every charge, including slippage.
Should I change my size after a big loss or a big win?
Kelly sizes are a share of current capital, so after a loss the rupee size falls and after a win it rises, automatically. What should not change is the fraction itself, unless the inputs have changed. Raising the fraction to win back a loss, or because a winning run feels like proof, is exactly the kind of decision a fixed rule is meant to prevent.
How many trades do I need to estimate Kelly inputs?
At least 100, and more is better. With fewer trades, the uncertainty in the win rate alone can move the Kelly fraction from positive to negative.
Is Kelly useful if I never bet the full amount?
Yes. It tells you the most the edge could justify. If your fixed size is close to the Kelly figure, your edge is thinner than your size assumes.
Related guides
- Lot size meaning: how lots set the smallest position you can take
- Max drawdown calculator: the fall you must be ready to sit through
- Walk-forward testing: estimating an edge on data the rules were not built on
References
- J. L. Kelly Jr., "A New Interpretation of Information Rate", *Bell System Technical Journal* 35(4), 917–926 (1956)
- E. O. Thorp, "The Kelly Criterion in Blackjack, Sports Betting and the Stock Market", in *Handbook of Asset and Liability Management*, Elsevier (2006)
- L. Breiman, "Optimal Gambling Systems for Favorable Games", *Proceedings of the Fourth Berkeley Symposium on Mathematical Statistics and Probability* (1961)
- P. A. Samuelson, "Why We Should Not Make Mean Log of Wealth Big Though Years to Act Are Long", *Journal of Banking and Finance* 3(4), 305–307 (1979)
- W. Poundstone, *Fortune's Formula*, Hill and Wang (2005)
- L. C. MacLean, E. O. Thorp and W. T. Ziemba, *The Kelly Capital Growth Investment Criterion*, World Scientific (2011)
- SEBI, circular "Measures to strengthen Equity Index Derivatives framework for increased investor protection and market stability", 1 October 2024 (sebi.gov.in)