Very good points in this article, Thomas. I did not study all of them in detail, but the essence of your message is crucial. Portfolio construction is not just "running the optimizer". In many ways that is the easiest step if you have a fast and stable implementation that can solve the practically relevant problems including all the important nuances like risk budgets, transaction costs and parameter uncertainty.
Thanks for mentioning our work with the Time- and State-Dependent Resampling. This can indeed be very powerful when applied properly.
Just a few comments:
1. The Time- and State-Dependent Resampling article is written by both Laura Kristensen and me.
2. There is a [insert link] after the initial reference to the article that probably should be adjusted somehow.
3. In relation to Entropy Pooling's limitation to scenarios in the prior, I view it as a prior problem rather than an Entropy Pooling problem if you prior model cannot generate all the scenarios that you can currently imagine.
4. If you look at the Black-Litterman model and think about all the questionable engineering necessary to make it work, as well as the logical inconsistencies in edge cases, I think it is safe to say that Entropy Pooling is always better, even in the CAPM prior case. In that case, I think there even exist analytical solutions in Meucci's original article, but I have never used them myself because I have just witnessed so many issues with the normal/elliptical assumption that I think it is a recipe for disaster.
Thanks a lot, Anton — I really appreciate the thoughtful comments.
You are absolutely right on the attribution. I will correct the reference to the Time- and State-Dependent Resampling article and properly mention both Laura Kristensen and you. I will also fix the remaining “[insert link]” placeholder.
I also fully agree with your point on Entropy Pooling and the prior. My formulation should be more precise: if an economically relevant scenario is missing from the support of the prior distribution, that is not an Entropy Pooling limitation but a prior model design problem. In that case, the modeller should revisit the scenario generation layer rather than expect the view-integration method to compensate for an incomplete scenario universe.
I agree that once we move beyond simple return views and normal or elliptical assumptions, Entropy Pooling is the more general and practically more robust architecture compared to BL. Especially in a scenario-based framework, the ability to express views on regimes, tail risks, dependencies, factors, and non-linear portfolio features is a major advantage.
Related to this, I am currently experimenting with optimal transport / Wasserstein-distance-based approaches for robustness diagnostics or distributional uncertainty in optimization problem. Have you worked with these methods in a portfolio construction context? I would be very interested in your experience or thoughts.
It's a well written article. I enjoyed reading it and learned from it. Thanks for sharing and for keeping the newsletter free.
Very good points in this article, Thomas. I did not study all of them in detail, but the essence of your message is crucial. Portfolio construction is not just "running the optimizer". In many ways that is the easiest step if you have a fast and stable implementation that can solve the practically relevant problems including all the important nuances like risk budgets, transaction costs and parameter uncertainty.
Thanks for mentioning our work with the Time- and State-Dependent Resampling. This can indeed be very powerful when applied properly.
Just a few comments:
1. The Time- and State-Dependent Resampling article is written by both Laura Kristensen and me.
2. There is a [insert link] after the initial reference to the article that probably should be adjusted somehow.
3. In relation to Entropy Pooling's limitation to scenarios in the prior, I view it as a prior problem rather than an Entropy Pooling problem if you prior model cannot generate all the scenarios that you can currently imagine.
4. If you look at the Black-Litterman model and think about all the questionable engineering necessary to make it work, as well as the logical inconsistencies in edge cases, I think it is safe to say that Entropy Pooling is always better, even in the CAPM prior case. In that case, I think there even exist analytical solutions in Meucci's original article, but I have never used them myself because I have just witnessed so many issues with the normal/elliptical assumption that I think it is a recipe for disaster.
Thanks a lot, Anton — I really appreciate the thoughtful comments.
You are absolutely right on the attribution. I will correct the reference to the Time- and State-Dependent Resampling article and properly mention both Laura Kristensen and you. I will also fix the remaining “[insert link]” placeholder.
I also fully agree with your point on Entropy Pooling and the prior. My formulation should be more precise: if an economically relevant scenario is missing from the support of the prior distribution, that is not an Entropy Pooling limitation but a prior model design problem. In that case, the modeller should revisit the scenario generation layer rather than expect the view-integration method to compensate for an incomplete scenario universe.
I agree that once we move beyond simple return views and normal or elliptical assumptions, Entropy Pooling is the more general and practically more robust architecture compared to BL. Especially in a scenario-based framework, the ability to express views on regimes, tail risks, dependencies, factors, and non-linear portfolio features is a major advantage.
Related to this, I am currently experimenting with optimal transport / Wasserstein-distance-based approaches for robustness diagnostics or distributional uncertainty in optimization problem. Have you worked with these methods in a portfolio construction context? I would be very interested in your experience or thoughts.