Portfolio optimisation sounds like a solved problem. You have assets, you have returns, you crunch the maths and out pops the perfect allocation. If only. The dirty secret of mean-variance optimisation (MVO) is that it treats your estimated covariance matrix as gospel — and in live trading, that matrix is about as reliable as a weather forecast printed three weeks ago.
This question matters enormously because the gap between backtested Sharpe ratios and live performance is often not a strategy problem — it's an estimation problem. MVO is exquisitely sensitive to input errors. Small mistakes in your covariance estimates get amplified into wildly unstable weights. Marcos Lopez de Prado demonstrated this clearly in his 2016 SSRN paper introducing Hierarchical Risk Parity (HRP), and the quantitative community has been rethinking portfolio construction ever since.
The core problem with MVO comes down to matrix inversion. To solve for efficient frontier weights, you must invert the covariance matrix. Inversion massively amplifies estimation errors — assets with similar return profiles get assigned extreme opposing weights to exploit tiny perceived differences. Think of it like adjusting a recipe where you accidentally misread 1 tablespoon as 1 cup. The optimiser doesn't know the measurement was dodgy. It just acts on it, confidently and catastrophically.
HRP sidesteps matrix inversion entirely. Instead, it uses hierarchical clustering to group assets by their correlation structure — think of it as sorting your portfolio into families of related assets. Risk is then allocated top-down through the tree, with each branch receiving weight inversely proportional to its variance. No inversion. No error amplification. The structure of the data guides the allocation rather than fighting it. Lopez de Prado's out-of-sample simulations showed HRP consistently delivered lower drawdowns and better risk-adjusted returns than MVO and naive equal-weighting when estimation noise was present — which, in real markets, it always is. For a rigorous grounding in the mathematics, the hierarchical risk parity explainer on Investopedia covers the mechanics cleanly, while the foundational theory behind covariance instability is well-documented in the Modern Portfolio Theory article on Wikipedia. If you want to go straight to the source, Lopez de Prado's original methodology is grounded in hierarchical clustering techniques that have been standard in machine learning for decades.
The practical takeaway is simple: if your covariance matrix is estimated from fewer than several hundred observations — which describes almost every realistic trading strategy — MVO is optimising noise as confidently as signal. HRP won't always beat MVO on clean, stationary data, but clean stationary data doesn't exist in live markets.
Run both on your next backtest and compare out-of-sample weight stability. If your MVO weights are jumping around like a cat on a hot tin roof, HRP is probably worth the switch.
This content is for educational purposes only and does not constitute financial product advice. Past performance is not indicative of future results. Profit Logic Ltd (ACN 688 669 936) accepts no responsibility for errors or omissions in this content or anywhere on this website. Always seek advice from a licensed financial adviser before making investment decisions.