• A new NBER Working Paper uses machine learning to identify the characteristics of high- and low-performing mutual funds.
• It shows mutual fund performance can be predicted, in real time, using fund flows and fund return momentum.
• These metrics indicate mutual fund managers’ skill, and, by reallocating funds from the least to the most skilful managers, investors can make abnormal returns of 50bps a month.
This article is only available to Macro Hive subscribers. Sign-up to receive world-class macro analysis with a daily curated newsletter, podcast, original content from award-winning researchers, cross market strategy, equity insights, trade ideas, crypto flow frameworks, academic paper summaries, explanation and analysis of market-moving events, community investor chat room, and more.
- A new NBER Working Paper uses machine learning to identify the characteristics of high- and low-performing mutual funds.
- It shows mutual fund performance can be predicted, in real time, using fund flows and fund return momentum.
- These metrics indicate mutual fund managers’ skill, and, by reallocating funds from the least to the most skilful managers, investors can make abnormal returns of 50bps a month.
Almost a third of the US population rely on mutual fund performance for their pension or other financial objectives. By the end of 2020, these funds managed an estimated $24tn in assets. More than half are equity mutual funds, most of which actively trade stocks. With so many to choose from, is there a way to distinguish outperformers from underperformers?
Yes, according to researchers from Stanford and Colombia Universities. They use an artificial neural network model to uncover new evidence that fund flows and fund return momentum, which combined represent a fund managers’ skill,are the main predictors of US equity mutual fund performance. This skill leaves a trail for investors to exploit over several periods, as it can predict higher abnormal returns for up to three years.
Data and Methodology
The authors collect information on returns, expenses, total net assets, investment objectives and other fund characteristics for 3,275 actively managed mutual funds. The data is from the Center for Research in Security Prices’ Survivor-Bias-Free Mutual Fund Database and from Thomson Financial Mutual Fund Holdings. The sample period covers January 1980 to January 2019.
The authors focus specifically on predicting abnormal returns for each mutual fund, every month. This makes the prediction a relative objective because abnormal returns remove the level effect of market and other risk factors. Using abnormal returns instead of standard fund returns is therefore a better way of measuring relative mutual fund performance.
The paper calculates abnormal returns relative to the Carhart four-factor alpha model, arguably the main factor model used by mutual fund investors and in the mutual fund literature. The authors then use machine-learning techniques to connect abnormal fund returns to the characteristics of mutual funds, including the characteristics of the stocks they hold, and to variables that capture the state of the economy.
The Association Between Fund Performance and Fund Characteristics
Past research has identified 46 stock-level characteristics that can, somewhat, predict expected returns across mutual funds. The authors augment this with 13 fund-level characteristics that include information on fund momentum, fund characteristics (e.g., age, turnover, flow) and fund family characteristics (e.g., family age, number of funds in family).
This Deep Dive’s Appendix shows the full list of characteristics. The most important for the results are fund momentum (numbered 47-49), where ‘F_ST_Rev’ is the prior month’s abnormal return, ‘F_r12_2’ is the mean abnormal return over the prior 12 months, and ‘F_r2_1’ is the lagged one-month abnormal return.
To understand the association between these characteristics and mutual fund performance, the authors construct a simple portfolio: order the monthly abnormal returns of all mutual funds by the value each characteristic takes. For example, for ‘age’, this would mean sorting returns from the oldest to the youngest mutual funds. Then, go long the top decile (the oldest 10% in the above example), and short the bottom decile (the youngest 10%). Then ask, for which characteristics would this sorting of fund performance yield the best monthly abnormal returns and Sharpe ratio (SR)?
The answer is ‘F_r12_2’ and ‘F_ST_Rev’, both yielding a monthly SR greater than 0.2 and mean monthly abnormal returns of 0.3% or above (Table 1). Foreshadowing the main analysis (with neural networks), this exercise shows fund-level characteristics are much more informative than stock-specific characteristics for fund performance. Also, the state of the economy matters. The association between fund performance and fund characteristics strengthens during periods of high investor sentiment.
Neural Network Analysis
We have previously written on the forecasting power of machine learning, and here the authors use a feedforward network to predict abnormal returns. The inputs are the 59 characteristics specific to each fund and the two measures of the state of the economy. The output is abnormal returns. Between is a single ‘hidden layer’ in which non-linear relationships are created among the inputs (Chart 1).
At the end of the neural network estimation, the authors are left with a prediction of a mutual fund’s abnormal returns for the next month based on its 59 characteristics. They then sort funds into deciles based on these predicted returns, where each decile represents a portfolio. They give weights to each fund in the portfolio. In all, there are 10 such portfolios – one for each decile.
The Importance of Fund Momentum and Fund Flows
The cumulative abnormal returns of investing in the top 10% of mutual funds, based on the neural network’s prediction, would have earned a cumulative abnormal return of 72% over the 40-year period (Chart 2). Investing in the bottom 10% would have yielded abnormal returns of -119%, a 191% difference. In other words, the neural network excelled at predicting both mutual fund winners and losers!
So, what makes abnormal returns predictable? The authors run the neural network on a subset of either stock-level characteristics or fund-level characteristics to find out (Charts 3 and 4). It shows that the best model for predicting abnormal mutual fund returns includes fund-specific characteristics. You can achieve the highest cumulative abnormal returns and SR by using the predictions from a neural network that only includes information on fund momentum (‘F_r12_2’) and fund flows (‘flows’) and is interacted with investor sentiment.
The Winning Mutual Fund Strategy
For investment strategy, you can achieve the best monthly returns like so. First, collect information on the state of the economy, fund momentum and fund flows. Then predict (using machine learning methods) the abnormal returns of mutual funds using only those inputs. Sort funds from highest to lowest in terms of predicted returns for the next month. Go long the top 10% and short the bottom 10%. This would achieve an annualised abnormal return of 6% per year and an SR of 1.5.
Why Do Fund Momentum and Fund Flows Proxy Fund Manager Skill?
The authors decompose abnormal returns into two components. One reflects trading between quarters (between disclosure), and the other reflects trading within quarters (within disclosure).
The between-disclosure abnormal return is the abnormal return of an investor who invests in the most recently disclosed stock positions of a fund and holds that portfolio until next fund disclosure. It captures the long-term stock-picking skills of mutual funds. A positive value means that the fund can pick stocks with positive alpha at a quarterly frequency.
Meanwhile, a high value of within-disclosure abnormal returns indicates that the fund is adding value by actively trading between two adjacent disclosure dates. We can break this down further by the return gap (a measure of the returns a fund has created from trades within the quarter) and the risk exposure differential (a measure of the portfolio’s systematic risk).
Fund momentum and fund flows can predict all three components. That is, a fund with high rankings on these characteristics is trading in ways that create returns both between quarters and within quarters, while reducing the systematic risk of the portfolios. In other words, higher fund momentum and fund flows represent investors reallocating funds to managers that have been navigating stock markets well over the last 12 months.
The authors show we can predict mutual fund performance. The key lies not in the characteristics of stocks but in the momentum and flows of the fund. These metrics are proxies for the skill of mutual fund managers. Funds with high momentum and high flows are typically creating positive abnormal returns between quarters and within quarters, all while managing systematic risk.
If you can find these skilful fund managers, it really pays off. The difference between picking a winner over a loser was nearly 5% per year in abnormal returns. Also, when adjusting for fund fees, only the top 20% of mutual funds make positive abnormal returns, so picking the crème-de-la-crème of funds is an important undertaking!
Table 2 provides a full list of fund- and stock-specific characteristics used to predict the abnormal returns of mutual funds in any given month. Characteristics 1-46 are categorised as ‘stock-specific’, and 47-59 are ‘fund-specific’ characteristics.
Sam van de Schootbrugge is a Macro Research Analyst at Macro Hive, currently completing his PhD in international finance. He has a master’s degree in economic research from the University of Cambridge and has worked in research roles for over 3 years in both the public and private sector.