Skip to content
All library documents

Why Distribution Estimation and Integral-Based Kelly Sizing Are Uncommon

Article Quant Q&A · Author: thankfulperson
Read the original

Summary

The document asks why trading research does not more often estimate conditional return distributions from data, update them with Bayesian methods, and use those distributions to solve for expected-value-optimal position sizes. It notes that normality assumptions can be unrealistic and that generalized Kelly sizing can be expressed through an integral over the return distribution.

The text poses the idea as a question rather than presenting an implementation, evidence, or a review of existing methods. It does not identify a specific computational barrier or establish whether distribution-based sizing is absent from practice. Its central takeaway is therefore an open research question: how to make conditional distribution estimates reliable enough for position sizing, given that both the distribution and its estimates may change across market conditions.

Key ideas

  • Generalized Kelly sizing can use an integral over an assumed return distribution.
  • Financial return distributions may change with market conditions.
  • The document asks whether data-driven Bayesian distribution updates could support adaptive position sizing.
  • It offers no evidence or specific answer about computational or practical barriers.

Tags

StatisticsPosition sizingRisk management
Full text
# Why don’t methods focus on constructing expected distributions and solving the integrals


# Why don’t methods focus on constructing expected distributions and solving the integrals












Everyone knows assumptions of normality etc are bad and that the expected distributions of financial quantities (such as returns) change depending on the circumstances.

We know that we can compute the asymptotic optimal EV position sizing by solving the integral associated with a distribution (generalized Kelly). So why doesn’t the literature, firms, etc. focus on systematic data-driven Bayesian updates of expected probability distributions in various circumstances, and then just applying the optimal position sizing.

Surely this isn’t a novel idea. Is there severe computational issues or something I’m missing.

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.