OP here. I spent the last 2 years building a risk management framework to fix the fragility of Naive Kelly betting.
The core issue is that standard Kelly assumes p_hat = p_star (perfect estimation). But due to Cramér-Rao bounds, there's always irreducible epistemic uncertainty. My solution uses Wasserstein-2 ambiguity balls to define a radius of uncertainty around the prediction, and then solves a convex dual problem to find the allocation that maximizes log-growth for the worst-case distribution inside that ball.
The Stack:
- Python 3.9+
- Optimization: Exact Convex Formulation using `cvxpy` + `SCS` splitting conic solver.
- Stochastic Model: Dynamic Bivariate Poisson (captures covariance μ between teams).
- Validation: Walk-Forward with strict temporal purging.
I released the stable version (v1.0) as a commercial Asset Kit for deployment. The repository linked above serves as the *public documentation*, detailing the mathematical proofs and the architectural overview for those who want to study the logic.
Let me know if you have questions on why I chose SCS over ECOS for this specific problem.
liripe•49m ago
The core issue is that standard Kelly assumes p_hat = p_star (perfect estimation). But due to Cramér-Rao bounds, there's always irreducible epistemic uncertainty. My solution uses Wasserstein-2 ambiguity balls to define a radius of uncertainty around the prediction, and then solves a convex dual problem to find the allocation that maximizes log-growth for the worst-case distribution inside that ball.
The Stack: - Python 3.9+ - Optimization: Exact Convex Formulation using `cvxpy` + `SCS` splitting conic solver. - Stochastic Model: Dynamic Bivariate Poisson (captures covariance μ between teams). - Validation: Walk-Forward with strict temporal purging.
I released the stable version (v1.0) as a commercial Asset Kit for deployment. The repository linked above serves as the *public documentation*, detailing the mathematical proofs and the architectural overview for those who want to study the logic.
Let me know if you have questions on why I chose SCS over ECOS for this specific problem.