In what areas does machine learning have advantages over conventional financial/investment analysis? In his June 2018 presentation entitled “Nine Financial Applications of Machine Learning”, Marcos Lopez de Prado summarizes investing-related areas in which well-supervised machine learning outperforms conventional methods. Based on relevant research and his experience, he asserts that:
- Machine learning beats linear factor models for price prediction because it focuses on price rather than factor selection and factor coefficients.
- Similarly, machine learning is effective in controlling for a potentially confounding variable directly rather than via an assumption of linear effect and coefficient.
- Portfolios constructed via machine learning beat equal weight and mean-variance optimization (which does not beat equal weight).
- Machine learning is particularly effective at identifying regime changes via cluster analysis.
- Machine learning is effective at position sizing via two-step analysis that first determines direction of trade and then sets the size of the trade.
- Machine learning directly relates a large number of data features to outcomes, thereby exposing which features are most important. Essentially, machine learning sacrifices closed-form specifications (theoretical formulas) for breadth of variables considered.
- Machine learning algorithms can replicate, and may be able to predict, a large percentage of analyst and credit rating agency recommendations.
- Machine learning is well-suited to extraction of sentiment via natural language processing.
In summary, well-supervised machine learning is often more effective for investment analysis than conventional modeling because it accesses a large number of data features without restrictive assumptions.
Whether they have access to machine learning or not, investors are increasingly in competition with it.
Cautions regarding assertions include:
- Application of machine learning requires compute power, sophisticated software and skilled supervision. These capabilities are costly and beyond the reach of most investors, who would bear fees for delegating them to a fund manager.
- As noted in the presentation:
- Overlapping data confounds determination of the connection between outcome and data features.
- Financial series may have properties inconsistent with machine learning assumptions, promoting discovery of patterns when there are none.
