López de Prado on machine learning in finance

Marcos López de Prado

Introduction

Marcos López de Prado, whom we have featured in previous Math Scholar articles (see Article A, Article B and Article C), has been invited to present a keynote presentation at the ACM Conference on Artificial Intelligence in Finance, to be conducted virtually October 14-16, 2020.

López de Prado is a faculty member of Cornell University and also CEO of True Positive Technologies, LP, a private firm that provides machine learning techniques techniques for finance applications. He is also the author of two books in the field: Advances in Financial Machine Learning, published by Wiley (2018) and Machine Learning for Asset Managers, published by Cambridge University Press (2020).

López de Prado’s talk

López de Prado has graciously provided the viewgraph file for the talk he is scheduled to present at the ACM Conference on AI in Finance: Viewgraph file. Here are some of the key points of the talk:

• Financial machine learning offers unique opportunities to gain insight from data: extracting non-linear relationships in high-dimensional data, analyzing unstructured, asynchronous or categorical data, learning complex hierarchical, nonparametric patterns, as well as controlling for statistical overfitting.
• Among the advantages of machine learning techniques, when properly applied to finance, is that they can learn without being specifically directed, often find patterns that cannot easily be represented by a simple set of equations, and can find solutions involving a large number of variables and interactions.

• Effective machine learning methods in finance are, for the most part, not “plug-and-play” methods — considerable insight and care must be exercised.
• Ten particularly promising applications of machine learning in finance are:
  1. Price prediction: Finding nonlinear relations, hierarchical relations and categorical variables.
  2. Hedging: Reinforcement learning methods involve very few assumptions.
  3. Portfolio construction and risk analysis: Machine learning methods outperform classical mean-variance portfolio optimization, with Sharpe ratio gains often exceeding 30%.
  4. Structural breaks / outlier detection: Machine learning techniques can reduce the percentage of incorrect buy/sell signals due to structural breaks and outliers.
  5. Bet sizing / alpha capture: Meta-labeling methods can improve buy-sell decisions.
  6. Feature importance: As mentioned above, machine learning methods can identify patterns in high-dimensional space, identifying key features that are often missed as a result of a model’s misspecification.
  7. Credit ratings / analyst recommendations: Machine learning algorithms have been successful in reproducing the recommendations produced by bank analysts and credit rating agencies, which often involve a number of models and subjective heuristics.
  8. Unstructured data: Machine learning tools are often effective in analyzing unstructured data, as often arises in such arenas as analyses of news articles, corporation announcements, sales figures, transportation activity and more.
  9. Execution: Machine learning may be used in analyzing trading and execution activities.
  10. Detection of false investment strategies: Backtest overfitting and other forms of statistical error are the bane of the finance world. Machine learning strategies, in conjunction with other techniques, are effective in guarding against statistical overfitting and other forms of false discoveries.

For full details, see López de Prado’s viewgraphs at the above link.