Position sizing: a small edge can carry far less risk than Kelly says
The Kelly fraction of two small edges is 23% and 17% of the account per trade. Under a hard drawdown limit the best size is near 6%.
- What I did
- I took every trade of the two rules from my relative value study (71 on the crack spread, 107 on the Treasury butterfly) and measured it in R, units of the risk taken. I then resampled that history (redrew it at random, with replacement) at 0.25% to 25% risk per trade, under the rules of a typical funded-account challenge (the test a trader must pass to receive a firm’s capital).
- Why it matters to a trader
- A retail prop-firm evaluation, usually on simulated capital, and a multi-manager fund, with real capital, both set a drawdown limit (the largest fall from a peak they allow). Inside that limit, the size of your trades decides whether a small edge is tradeable or not.
- What I found
- Each rule wins about three trades in four, but its average loss is bigger than its average win. At 1% risk no simulated account passes within a year, because each rule trades only 3 to 4 times a year. The chance of passing and keeping the account for a year peaks at 6% per trade. Trading both rules lifts the peak from 26% and 21% to 37%.
- What I could not show
- That real trades would look like these. The trades are few, and resampling cannot create a shock worse than the worst one in the sample (-1.84 R). The two rules were flat (held no position) on most days, so their near-zero correlation says little about a real crisis. Funded-account rules differ from firm to firm: mine are typical and are no firm’s contract.
Try different risk levels
Loading the data
More settings
At 5% risk per trade, 31% of accounts pass, 14% fail and 54% run out of time.
How position sizing works
Sizing a trade means choosing how much of the account you lose if you are wrong. Here every trade has a stop: a price at which you close the trade at a loss. The distance from the entry to the stop is one R. If you risk 2% of the account per trade, a trade stopped out (-1 R) costs about 2% and a trade that gains half of that distance (+0.5 R) earns 1%.
Size matters because after losing 10% you need to make 11% to get back. The usual method is to risk a fixed share of the current balance. A common choice is a fraction of the Kelly fraction: the risk per trade that makes an account grow fastest over the long run, given the trades you have seen.
The two edges, measured in R
Data to 3 November 2026.
| Since 2008 and 1991 | Crack | Butterfly |
|---|---|---|
| Trades | 71 | 107 |
| Winning trades | 77% | 71% |
| Average win, in R | 0.62 | 0.55 |
| Average loss, in R | -1.25 | -0.99 |
| Average trade, in R | 0.20 | 0.11 |
| Worst trade, in R | -1.84 | -1.53 |
| Kelly fractionand its 90% interval | 23%5% to 39% | 17%0% to 33% |
Kelly: growth against drawdown
Growth per trade peaks at the Kelly fraction, is zero again at about 1.8 times Kelly, and is negative beyond.
Crack, five years, 5,000 paths. 5th percentile: 5% of paths (simulated histories) end lower. Data to 3 November 2026.
| Risk per trade | Median balance | 5th percentile | Median maximum drawdown | Drawdown beyond 10% |
|---|---|---|---|---|
| 2% | 1.08 | 0.96 | 6% | 10% |
| 3% | 1.11 | 0.93 | 9% | 37% |
| 6%near quarter Kelly | 1.23 | 0.86 | 17% | 93% |
| 10%near half Kelly | 1.37 | 0.75 | 28% | 99% |
| 20%near full Kelly | 1.62 | 0.46 | 51% | 100% |
A sensible size is well below Kelly, for three reasons. Kelly is estimated from few trades, hence the wide interval above. It is unreliable out of sample (on trades it has not seen). Estimated only from earlier trades, full Kelly ends the butterfly at 0.48 times the start, against 1.36 for a fixed 5%. On the crack it ends higher, 2.43 against 1.73, but only after a maximum drawdown of 90% (73% on the butterfly). And a funded account stops you at a drawdown of 5% to 10%, well before the 50% that full Kelly can reach.
The odds under a hard drawdown limit
Typical rules: reach +10% within a year, before the balance falls 10% below its start or drops 5% in one day. If you pass, you continue for another year with the same limits. The rule values are assumptions (Idea). The simulation of them is Tested.
Share of accounts that pass, then go a full year without breaking a rule. For “both”, each rule risks half of the stated amount. Quarter blocks, 5,000 paths, accurate to about 2 points. Data to 3 November 2026.
| Risk per trade | Crack | Butterfly | Both |
|---|---|---|---|
| 1% | 0% | 0% | 0% |
| 3% | 7% | 6% | 3% |
| 6% | 26% | 21% | 22% |
| 8% | 22% | 19% | 32% |
| 10% | 14% | 16% | 37% |
| 20% | 5% | 2% | 10% |
The pass rate stops rising near 44% to 56%, because at large sizes the first one or two trades decide the result. The chance of keeping the account peaks and collapses beyond 10%: at 20%, 91% of the crack accounts that pass fail within a year. That peak is no long-run size: at 6%, 93% of five-year crack paths have a drawdown beyond 10%.
The time limit matters most. With three years the best chance rises to 56% for the crack, 45% for the butterfly and 66% for both. With three months it is at most 6%. With a drawdown limit of 5% it falls to 19%, 14% and 21%.
Two rules instead of one
The daily profit of the two rules has a correlation of 0.004, but both are flat on most days, and zeros pull a correlation towards zero. On weekly totals it is -0.09 (95% confidence interval -0.15 to -0.03). It did not rise in 2008, 2020 or 2022, when both rules held a trade on only 27, 74 and 41 days. Three episodes cannot rule out a jump in correlation in a different crisis.
“Both” gives each rule half the stated risk. Over their common window, at 5%, a drawdown beyond 10% hits 31% of paths, against 82% and 73% for each rule alone. Together they also earn less, so they pass less often. At equal volatility (0.71 of the stated risk on each rule) they pass about as often as the better rule and keep the account more often (27% against 23% and 21%).
How much should you trust it?
- Tested Stops checked at the close. 21% of crack trades and 23% of butterfly trades lost more than 1 R, the loss the stop was meant to cap. That is most losing trades (15 of 16 and 25 of 31), because the test only looks at closing prices and fills there. A real stop would fill elsewhere, and in a fast market worse.
- Tested Costs. Taking a further 0.1 R off every trade (slippage: fills worse than the expected price) cuts the Kelly fraction from 23% to 13% for the crack and from 17% to 1% for the butterfly. Taking 0.2 R off removes the edge.
- Tested The edge moves. The crack rule averaged 0.26 R in its first 35 trades and 0.14 R in the last 36. Without 2008, 2020 and 2022 the butterfly’s average trade falls from 0.11 R to 0.05 R, so its edge depends on those three years.
- Tested Time under water (below an earlier peak). Replaying the real history at 1% risk, an account spent up to 6.0 years (crack) and 10.4 years (butterfly) under water. A funded account rarely gives you that long.
What I did not test or model. A central bank or oil shock that reprices a whole market. Exchange margin and financing. The bid-offer spread beyond the flat cost in the trades. Fills later than the signal close, which would reduce the results (see the relative value page). Intraday moves: a real daily loss limit counts open profit and loss, and I use closing balances.
Date of the data. This page has no live data and does not say what to trade today. No trade was open on the last day of data, 3 November 2026. Each rule’s current position is on the worked examples page.
Research only. These rules have not been traded, and nothing here is investment advice. The results describe the past, resampled at random, and do not predict the future. They are not a plan for a real account.
Source and tests at github.com/Nicolas8330/position-sizing.
Every figure comes from one run of examples/make_results.py on 5 November 2026.
I tried 44 settings and did not select the best one.