Introducing MVO vs HRP vs RL Portfolio Comparison for after-work quants

Takeaway. One paper, three allocators, same Indian names: mean–variance (MVO/MVP) needs means and a cov and often blows up; HRP clusters the corr matrix and bisects risk without inverting a full Σ; deep Q-learning treats weights as a control problem with a neural Q-table. Sen et al. report Sharpes, returns, and risk on NIFTY50 and sector baskets—a bake-off template, not a live engine.

A Comparative Analysis of Portfolio Optimization Using Mean-Variance, Hierarchical Risk Parity, and Reinforcement Learning Approaches on the Indian Stock Market (Jaydip Sen and co-authors, 2023; arXiv:2305.17523).

Fit for learners: Same Indian equity universes, three allocators: mean–variance, hierarchical risk parity, and deep Q-learning—read as a bake-off, not a winner-takes-all product.

This Hub note is a plain-language intro. Not investment advice.

The problem

After-work notebooks jump from mean/cov/minimize to “I trained RL.” Those are different failure modes: MVO hates estimated means; HRP ignores means on purpose; RL can overfit the sample path and needs a state you actually understand. Comparing them on one market makes the trade-offs visible.

The idea / core intuition

• MVO: max Sharpe or min variance given constraints—inherits Ledoit–Wolf lessons if you use raw S.
• HRP (López de Prado 2016): hierarchical clustering on correlations → recursive bisection of risk budgets; no quadratic programme.
• RL: deep Q-learning; large discrete state space → DQN (network as Q-table) in an environment built from NIFTY50 and ten sector groups.
• Metrics: annualised return, risk, Sharpe across mixed and sector portfolios.

Their tables are one study. RL winning a Sharpe column is not a licence to skip HRP homework or costs.

Why after-work Python learners care

You can code HRP-style clustering after you can sort a list; DQN can wait. The paper is a map of what to implement first (MVO with shrinkage → HRP → maybe RL). Indian names also remind you US notebooks are not the universe.

Tiny evening exercise: for three toy assets with corr 0.8, 0.8, 0.1, write by hand a 2-cluster split (assets 1–2 vs 3) and assign 50/50 risk between clusters. Learning only.

One-evening reading checklist

1. Why MVO weights swing when means are noisy.
2. What HRP uses instead of Σ⁻¹.
3. What the RL agent’s environment is (prices → state/reward).
4. One reason a higher in-sample Sharpe for RL may not travel.

Download PDF

Download PDF

Landing: arXiv abs

Important disclaimer

Not investment advice. This Hub note is educational only. Sen et al. is an academic comparison on historical Indian equities—not a recommendation to trade MVO, HRP, or DQN portfolios.

What this is not

• Not CEF / subsidy framing
• Not Quant or Macro course CTA
• Not a robo-advisor or NIFTY product pitch

Next step

Single course CTA: SoloQuant Academy’s Introduction to Python Programming—nested lists for a tiny corr matrix come before clustering or Q-learning.