AI vs Physics in Portfolio Optimisation
summary
Backtests on 2019–2024 sector and factor ETFs (11 S&P sectors, nine Fama‑French-like factors) use a 6‑month lookback (125 trading days) and show a physics‑inspired optimizer that computes a multiscale covariance by normalizing sigma(Δt) ∝ Δt^Δ and averaging across scales. The method uses multifractal scaling and fractional PDE intuition to infer a Delta critical exponent, feeds that covariance into a minimum‑variance optimizer, and achieves ~0.53 Sharpe and ~30–31% max drawdown versus lower Sharpe for Markowitz and worse performance for an off‑the‑shelf RNN‑based AI optimizer.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:12Portfolio optimization approaches
- 2:37Physics vs AI and Markowitz
- 6:32Scaling laws in finance
- 12:54Dynamical critical exponent Delta
- 17:47Why Markowitz fails
- 21:00Multiscale covariance
- 24:29Weight dependencies and risk
- 26:16Backtest setup
- 29:12Results and conclusion
related talks
30mPredictive trading
23mComplex AI Forecasting Methods for Investments
29mComplex AI Forecasting Methods for Investments Portfolio Optimization
59mAI in Finance: Scope and Examples
19mAI & Digital Twins for Financial Risk & Asset Optimization- 59mAI in Finance: Scope and Examples
33mQuantamental Investing - the future of finance
30mHow to build a stock market money printing machine with AI