Practical Uses and Limits of the Kelly Criterion in Trading
Summary
The discussion describes several ways traders apply the Kelly criterion, while questioning how well its assumptions fit real trading. One approach links Kelly sizing to volatility targeting: expected return divided by a chosen Sharpe ratio sets a volatility target, and using half Kelly reduces that target to half the calculated level. The answer argues that this more conservative sizing may better reflect asymmetric risk.
Another practitioner describes using model confidence after meta-labeling to choose bet size and reports an improved Sortino ratio in their own tests. A separate reply cites a paper claiming Kelly can produce losses under certain utility functions, including potentially unbounded cumulative losses, but its author has not checked the calculations and notes errors in the manuscript. The thread offers examples rather than a validated comparative study; it gives limited detail about implementation, assumptions, or out-of-sample evidence.
Key ideas
- Kelly can be used to translate expected return and a target Sharpe ratio into a volatility target.
- Half-Kelly sizing reduces the resulting volatility target as a more conservative application.
- One practitioner reports using model confidence after meta-labeling to set bet size.
- The discussion raises concerns about Kelly assumptions and cites an unverified critique involving alternative utility functions.
Tags
Full text
# Kelly Criterion in real trading
# Kelly Criterion in real trading
Kelly Criterion seems to me very stylized and even unrealistic framework (given its assumptions), which has extremely limited applied use. I was wondering whether you are aware of any real-life trading strategies which are backed by Kelly Criterion? If so, could you please discuss a bit how Kelly Criterion was applied?
## Answer by Adam W (score 3)
https://quant.stackexchange.com/a/81054
This should be a comment, but I do not have enough reputation. The following facts may be of interest to you. In Thorp, E. O. (2011) 509–523, it is claimed that Samuelson (yes that Samuelson) has shown that the Kelly criterion will lead to a loss under the utilities: $$H(W)=-1/W$$ $$T(W)=\sqrt{W}$$ In fact following the Kelly criterion will lead the sum of losses to be infinite! I have not verified these computations and the manuscript is filled with a lot of typesetting errors.
However, Samuelson was a very staunch critic of the Kelly criterion. In fact, he once wrote and published a peer-reviewed paper criticizing the Kelly criterion with all, but one, single syllable words to prove his point.
## Answer by KaiSqDist (score 1)
https://quant.stackexchange.com/a/80499
When reading Systematic Trading by Carver, he doesn't exactly explain the mathematics behind using Kelly criterion. He just states that Kelly criterion is used in volatility targeting.
How he uses it in volatility targeting is via the Sharpe ratio. Say for example you computed an annualized expected return of 8.0%, then you would like to keep the Sharpe ratio at 1.0 and set the volatility target that does so. Therefore, the annualized volatility acceptable would be:
$$Volatility\,Target = \frac{Expected\,Return}{Sharpe\,Ratio} = \frac{8.0\%}{1} = 8.0\%$$
But then he goes onto say that this is not very good because risk itself is not symmetric and instead asymmetric, so a better use for Kelly criterion is half Kelly criterion instead, which becomes:
$$Volatility\,Target = 0.5*\frac{Expected\,Return}{Sharpe\,Ratio} = 0.5*\frac{8.0\%}{1} = 4.0\%$$
## Answer by JoLu (score 1)
https://quant.stackexchange.com/a/81048
I remember hearing Marcos López de Prado say Kelly can be used after applying meta labeling. I don't remember if it was in his book or from a lecture he was giving.
I myself have strategies that utilizes just that. Using the model's confidence levels and Kelly to determine my bet size. From my testing, I went from a Sortino of 1.7 to a 3 using this method.
Robert Carver's blog mentions using Kelly also.Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.