KellyBoost gets the objective right. That can still mean betting too much.

Public research brief

An optimizer can solve the problem it was given and still deliver a portfolio we would not want to own.

KellyBoost trains boosted trees to choose portfolio weights directly by optimizing log wealth growth. Instead of forecasting returns and then allocating, it learns the allocation itself.

• The paper finds better point estimates for the growth objective than a classification surrogate in four paired comparisons.
• The same paper reports a major boundary: the direct models can concentrate aggressively, and a conventional two-stage pipeline can perform better.

We have reviewed the study, not replicated it. The paid section examines why the negative finding is as informative as the new method.

AI assists curation and drafting. The research status distinguishes paper reviews from house tests. Read our research approach.