Objective research to aid investing decisions

Value Investing Strategy (Strategy Overview)

Allocations for December 2024 (Final)
Cash TLT LQD SPY

Momentum Investing Strategy (Strategy Overview)

Allocations for December 2024 (Final)
1st ETF 2nd ETF 3rd ETF

Realistic Expectations for Machine Learning for Asset Management

| | Posted in: Investing Expertise

Will machine learning revolutionize asset management? In their January 2020 paper entitled “Can Machines ‘Learn’ Finance?”, Ronen Israel, Bryan Kelly and Tobias Moskowitz identify and discuss unique challenges in applying machine learning to asset return prediction, with the goal of setting realistic expectations for how much machine learning can improve asset management. Based on general characteristics of financial markets and machine learning algorithms, they conclude that:

  • Machine learning typically means seeking the best model from a diverse set of models that each allow many input variables and/or complex non-linear associations between inputs and output. Machine learning thrives in big data environments with strong signals and high signal-to-noise ratios.
  • To suppress in-sample overfitting, machine learning includes regularization, constraining model size/complexity.
  • Perhaps the clearest difference between machine learning and traditional statistics is the variety of computationally efficient optimization methods for the former.
  • However, as applied to asset management:
    • The core task, return prediction, is usually a small data problem. At the monthly frequency most reasonable for the bulk of asset management, there are only a few hundred to a few thousand return observations per asset, and asset returns may be correlated. In other words, the number of observed independent return outputs, not the number of predictor inputs, limits asset allocation model robustness. (Machine learning is suited to high-frequency trading, which offers many more return observations, but these trades must be small, limiting who can participate.)
    • Interesting new data sources such as social media and images have especially short histories.
    • Return signals are weak and signal-to-noise ratios low, with market forces constantly driving them toward zero. Employing economic theory (such as limiting inputs to widely accepted risk premiums) as a complement to machine learning can mitigate this concern.
  • Other portfolio implementation issues, such as risk management and trading frictions suppression, are more amenable than asset returns to machine learning solutions.

In summary, the inherent nature of financial data suggests that gains from machine learning in asset management will be evolutionary, not revolutionary.

Cautions regarding conclusions include:

  • As noted in the paper, financial industry marketing generally oversells the potential of machine learning.
  • Most investors bear fees for access to machine learning, and it is not obvious that gains offset such costs.
Login
Daily Email Updates
Filter Research
  • Research Categories (select one or more)