•
Remlinger, Alasseur, Brière & Mikael (2021) mix thirteen Gu–Kelly–Xiu-style experts with Bernstein Online Aggregation so long-short books adapt when models rotate. Not investment advice.
Takeaway. Instead of crowning one ML model, Bernstein Online Aggregation (BOA) puts convex weights on a panel of experts (here: linear, trees, nets forecasting next-month returns) and updates those weights from recent losses—useful when the winner keeps changing.
Expert Aggregation for Financial Forecasting (Carl Remlinger, Clémence Alasseur, Marie Brière, Joseph Mikael, 2021; arXiv:2111.15365). Reads as a companion to Gu–Kelly–Xiu.
Fit for learners: Online Bernstein aggregation of long-short ML experts—don’t pick one model; reweight them as regimes change.
This Hub note is a plain-language intro. Not investment advice.
Gu et al.–style bake-offs still force a choice. Accuracy is non-stationary: OLS+H can dominate shorts before 2001, nets after. Static bagging does not reweight every month. Hyperparameters multiply the zoo.
A robust-combination recipe—not a live fund.
This is softmax-on-losses you can code in a dozen lines once lists feel easy. It also teaches regret: you compare to the best convex oracle, not to magic alpha.
Tiny evening exercise: two experts with losses [0.2, 0.5] last month; start from equal weights; multiply by exp(-eta * loss) and renormalise. Learning only.
Landing: arXiv abs
Not investment advice. This Hub note is educational only. Expert aggregation is academic portfolio research—not a recommendation to run long-short ML books.
Single course CTA: SoloQuant Academy’s Introduction to Python Programming—lists of weights and a running product; aggregation is renormalising those lists.