Reinforcement Learning in Dynamic Asset Allocation

Agents trained on historical drawdown scenarios behave differently from mean-variance optimisers - and the gap matters when volatility clusters.

Recent coverage on ML and portfolio construction

Practical analysis for practitioners working at the intersection of quantitative finance and applied machine learning.

Graph comparing neural network factor exposures against classical Fama-French factors

Analysis

Factor Models and Neural Networks: Where They Overlap

Classical factor models give you interpretability. Neural networks give you flexibility. The more interesting question is what happens when you treat learned embeddings as implicit factors - the overlap is larger than most practitioners expect, and the failure modes are shared too.

Read analysis

Risk

Risk Constraints That Machine Models Still Get Wrong

Gradient-based optimisers are good at minimising loss functions, but they are not naturally good at respecting hard constraints. Position limits, sector caps, and turnover budgets require explicit handling - and the way you encode them changes outcomes significantly.

Seminar participants reviewing covariance estimation outputs on shared screens

Seminar recap

Covariance Estimation Under Non-Stationarity

Ledoit-Wolf shrinkage is a reasonable default, but when the return distribution shifts - as it does around earnings seasons or macro announcements - the shrinkage target itself becomes stale. Participants in our last seminar spent considerable time on regime-conditioned estimation approaches that adapt the target dynamically.

View seminar details

Pieter Vanzyl

Quantitative Research Lead

Adaeze Okonkwo

Portfolio Strategy Analyst

Reinforcement Learning in Dynamic Asset Allocation

Recent coverage on ML and portfolio construction

Factor Models and Neural Networks: Where They Overlap

Risk Constraints That Machine Models Still Get Wrong

Covariance Estimation Under Non-Stationarity

Who shapes the discussion