Part 2: How much to risk · Chapter 7 of 14 · 11 min
The Kelly criterion
The Kelly criterion gives the fraction of capital to risk to make the account grow as fast as possible: f ≈ μ/σ² (mean divided by variance, in R). On our series it is 0 for LA, 2.9% for LD (with an error from −7% to +13%), about 21% for SX. But at full Kelly risk the median drawdown of SX is 94% and the probability of touching −50% is 46%: maximum growth is paid for with an account nobody can bear. Half Kelly gives 75% of the growth, and at twice Kelly capital falls even with a true edge.
Kelly's question
In chapter 5 we saw that capital grows by about f × μ − f² × σ² / 2 per trade: the expected gain minus the cost of volatility. This curve rises, reaches a maximum and then falls. The risk at which growth is maximum is called the Kelly fraction (after the physicist John Kelly, who derived it in 1956 for information theory).
For a series of results in R with mean μ and deviation σ, it is about:
> f* = μ / σ²
If results are only +b R (with probability p) and −1 R (with probability 1 − p), the exact formula is:
> f* = p − (1 − p) / b
Some examples:
| Winning share p | Win / loss b | Kelly f* |
|---|---|---|
| 60% | 1 | 20.0% |
| 50% | 1.5 | 16.7% |
| 50% | 2 | 25.0% |
| 40% | 2 | 10.0% |
| 35% | 3 | 13.3% |
| 70% | 0.5 | 10.0% |
| 50% | 1 | 0% |
| 40% | 1.5 | 0% (break-even: 0.4 × 1.5 − 0.6 = 0) |
| 60% | 0.5 | negative: do not bet |
A zero or negative value says not to trade: the expected value is not positive.
The Kelly of our series
Let us compute the fraction on the four series. You take the mean μ and the deviation σ of the series in R and divide; the numerical value is the one that maximises the mean of the logarithm of (1 + f × R) over all the results. In brackets, what the mean would give with its error (chapter 2):
| Series | Mean ± error | Deviation | Kelly (μ/σ²) | Range with the error | Numerical Kelly |
|---|---|---|---|---|---|
| LA | −0.03 ± 0.08 | 0.99 | −3.3% | from −11.7% to +5.0% | 0% |
| LD | +0.03 ± 0.10 | 1.00 | 2.9% | from −7.4% to +13.2% | 2.9% |
| SD | +0.09 ± 0.15 | 0.56 | 29.8% | from −17.7% to +77.2% | 31.4% |
| SX | +0.21 ± 0.08 | 0.98 | 21.5% | from 13.0% to 30.1% | 20.7% |
The limit of the criterion shows at once. Kelly depends on the mean, and the mean of the real series has an error equal to or larger than the value. For LD the Kelly fraction can be anywhere between −7% and +13%: you do not even know whether it is positive. For SD, with 55 trades, between −18% and +77%. For LA it is zero. Only for SX, where we built the edge ourselves, does the range sit above zero.
The growth curve
Take SX, which has a true edge, and look at what happens to capital for different multiples of the Kelly fraction (20.7%), in the true order of the trades:
| Risk | Growth per trade | Capital after 552 trades | Maximum drawdown |
|---|---|---|---|
| 1/4 of Kelly (5.2%) | +0.94% | ×175 | 38.5% |
| 1/2 of Kelly (10.4%) | +1.61% | ×7,110 | 63.8% |
| Kelly (20.7%) | +2.15% | ×141,000 | 92.5% |
| 1.5 times Kelly (31.1%) | +1.56% | ×5,570 | 99.2% |
| 2 times Kelly (41.4%) | −0.30% | ×0.19 | 100% |
| 3 times Kelly (62.1%) | −9.35% | ×0 | 100% |
(The enormous multipliers are a property of SX, an edge built by hand over 552 trades: they are not a forecast.) Three facts:
- Growth rises up to Kelly and then falls. At 1.5 times Kelly it grows as at half of Kelly (1.56% against 1.61%), with a 99% drawdown.
- At twice Kelly growth is negative even with a true edge. The cost of volatility exceeds the gain: it is the same threshold seen in chapter 5 (f = 2μ/σ²), double Kelly.
- The curve is flat near the maximum. Half Kelly gives 75% of the growth (1.61% out of 2.15%), with a maximum drawdown of 64% instead of 92%. A quarter of Kelly gives 44% of the growth with a drawdown of 38.5%.
Read the chart in words
Daily gold chart with, below, three curves of the logarithm of the capital of SX with different risks.
On random paths
The real order is only one. With the bootstrap (chapter 2) I make 3,000 paths of 500 trades drawing at random from the results of SX and look at what using Kelly means in practice:
| Risk | Median final capital | Median drawdown (95th percentile) | Touches −30% | Touches −50% |
|---|---|---|---|---|
| 1/4 of Kelly (5.2%) | ×112 | 39.8% (57.8%) | 6.2% | 0.4% |
| 1/2 of Kelly (10.4%) | ×3,330 | 67.5% (85.2%) | 28.3% | 10.2% |
| Kelly (20.7%) | ×53,900 | 93.7% (99.3%) | 63.1% | 45.7% |
| 2 times Kelly (41.4%) | ×0.30 | 100% | 96.3% | 94.2% |
At the Kelly risk the median drawdown is 94%: you lose almost everything in a typical path and then recover it. The median capital at the end is enormous, but you need an account that bears the −94% and the will to hold the position while it happens. No trader does. That is why Kelly is a ceiling: it says where growth is maximum, and it does not say where you should stand.
The estimated edge is not the true one
The worst problem: Kelly is computed from the estimated mean, and if the mean is overestimated, the computed risk is too high. Suppose the edge you estimated (μ = 0.20, σ = 1, Kelly 20%) is really a fraction of that:
| True edge (relative to the estimate) | Growth at 20% risk | At 10% risk | At 5% risk |
|---|---|---|---|
| 100% (μ = 0.20) | +2.00% | +1.50% | +0.875% |
| 75% (μ = 0.15) | +1.00% | +1.00% | +0.625% |
| 50% (μ = 0.10) | 0.00% | +0.50% | +0.375% |
| 25% (μ = 0.05) | −1.00% | 0.00% | +0.125% |
| 0% (no edge) | −2.00% | −0.50% | −0.125% |
(Growth per trade, about f × μ − f² × σ² / 2.) If the true edge is half the estimate, full Kelly no longer grows; half Kelly still grows 0.50%; a quarter 0.375%. If it is a quarter, only 5%, that is a quarter of Kelly, still grows. With a quarter of Kelly you give up more than half the growth if you were right (0.875% instead of 2.00%), but you keep growing even if you overestimated the edge by three quarters.
On real data too: LD has an estimated Kelly of 2.9%. At that risk it grows 0.042% per trade, at twice Kelly (5.8%) 0.002%, at three times (8.7%) it falls 0.123%. On the real order:
Read the chart in words
Daily gold chart with, below, three curves of the capital of LD with different risks and a line at 100%.
What to do with it
- Consider Kelly a ceiling. The risk per trade must not come near that value.
- Compute Kelly from the lower limit of the mean, not from the estimated value: μ minus the error. For LD it is −7.4%: it means zero, no bet. For SX it is 13.0%.
- Use a fraction: from a quarter to half of the value obtained, and compare it with the ceiling given by drawdown (chapter 4) and ruin (chapter 6). The risk per trade is the smaller of the two.
- If the mean is not demonstrated (lower limit below zero), Kelly says zero. For that system a small risk (0.5-1%) is a choice of prudence, not a result of Kelly.
- Do not confuse growth with safety: the fraction that maximises growth also maximises drawdowns.
Procedure in four steps
- Compute the mean, deviation and error of your series in R (chapter 2).
- Compute Kelly with the lower limit of the mean: (μ − error) / σ². If it is zero or negative, there is no Kelly to use.
- Take from a quarter to half of that value.
- Compare it with the maximum risk that drawdown and ruin allow (chapters 4 and 6) and use the smaller.
Exercises
Chapter exercises
1. A system wins 50% of the time and earns on average 1.5 times what it loses. What is the Kelly fraction?
2. A system wins 40% of the time and earns 1.5 times what it loses. What does Kelly say?
3. SX has a Kelly fraction of about 21%. What happens using twice Kelly (about 41%)?
4. You estimated an edge that gives a Kelly of 20%, but the true edge is half. Which risk still grows?
Have a question or want to share your exercise?
Post in the community, or join the free signals room on Telegram.
Educational content, not financial advice. Trading involves risk.