Please rotate your device to landscape mode to view the charts.

Machine-Learning the Skill of Mutual Fund Managers

Journal: Journal of Financial Economics

Date: 20231001

Author: Kaniel, Ron; Lin, Zihan; Pelger, Markus; Van Nieuwerburgh, Stijn

Abstract:
We show, using machine learning, that fund characteristics can consistently differentiate high from low-performing mutual funds, before and after fees. The outperformance persists for more than three years. Fund momentum and fund flow are the most important predictors of future risk-adjusted fund performance, while characteristics of the stocks that funds hold are not predictive. Returns of predictive long-short portfolios are higher following a period of high sentiment. Our estimation with neural networks enables us to uncover novel and substantial interaction effects between sentiment and both fund flow and fund momentum.

Link: Google Scholar

Background and Context

Market Size

U.S. mutual funds managed $24 trillion in assets by end of 2020, with over 100 million Americans relying on these funds for retirement savings and financial objectives.

Research Focus

The study uses machine learning techniques to identify which characteristics of mutual funds and their stock holdings can predict future fund performance.

Methodology

Analysis uses neural networks to examine 46 stock characteristics and 13 fund/family characteristics of actively-managed U.S. equity mutual funds between 1980-2019.

Fund Performance Spread Between Top and Bottom Deciles Shows Significant Predictability

  • Top 10% of funds generated +72% cumulative abnormal returns
  • Bottom 10% of funds generated -119% cumulative abnormal returns
  • Demonstrates strong predictability in fund performance using machine learning approach

Fund-Level Characteristics Outperform Stock-Level Characteristics in Prediction

  • Fund characteristics with sentiment achieve highest Sharpe ratio of 0.25
  • Stock characteristics alone show no predictive power
  • Demonstrates superiority of fund-level information for prediction

Long-Term Persistence of Predictability Over 36 Months

  • Predictability remains significant even after 36 months
  • Monthly Sharpe ratio only declines from 0.30 to 0.20 over three years
  • Shows unprecedented persistence in mutual fund performance prediction

Sentiment-Conditional Performance Shows Strong Interaction Effects

  • Performance is strongest during high sentiment periods
  • Monthly abnormal returns more than double from low to high sentiment states
  • Demonstrates importance of market sentiment in fund performance

Decomposition of Abnormal Returns Shows Multiple Sources of Performance

  • Between-disclosure returns account for largest portion of performance
  • Return gap and risk exposure differences contribute equally
  • Shows multiple channels through which fund managers add value

Contribution and Implications

  • Demonstrates that machine learning can consistently identify skilled fund managers before and after fees
  • Shows that fund flow and momentum are the key predictive characteristics, while stock holdings provide little value
  • Reveals important interaction effects between fund characteristics and market sentiment
  • Provides new methodology for analyzing relative performance prediction in asset management

Data Sources

  • Performance Chart: Based on cumulative returns data from Figure 5
  • Sharpe Ratio Chart: Constructed using data from Table 3
  • Persistence Chart: Based on holding period analysis from Figure 10
  • Sentiment Chart: Constructed using data from Table 7
  • Decomposition Chart: Based on data from Table 9