Estimating Mutual Fund Factor Exposures from Returns with LSTM

Summary

The study estimates a fund’s factor exposure from its net asset value returns and factor return series, aiming to provide more timely exposure tracking than holdings-based estimates from delayed periodic reports. It trains an LSTM on features describing the similarity between fund and factor returns, including rolling correlations, regression slopes, and residual sums of squares. Reported factor exposures from holdings serve as labels, with missing weekly labels filled by linear interpolation; the main example predicts exposure to a size factor.

In the test sample, the reported mean absolute error is 0.109, and the authors describe improvements over simple linear regression in matching exposure levels and responding to changes. Some funds are poorly fitted. A separate attempt to enlarge training data with simulated funds performed well on simulated examples but generalized poorly to real funds, which the authors attribute to unrealistic random trading behavior. The findings are preliminary and limited to the data, factor, and fund universe studied; the authors call for further model improvements and use of additional reporting information.

Key ideas

• The method infers fund factor exposure from fund returns and factor returns instead of relying only on delayed holdings disclosures.
• The LSTM uses rolling return correlations, regression slopes, residual measures, and a fund-boundary indicator as inputs.
• Holdings-based factor exposures are interpolated between reporting dates to create weekly training labels.
• The reported test error for the size-factor example is 0.109 MAE, though some real funds are not fitted well.
• Training on randomly simulated funds failed to generalize to real funds, underscoring the importance of realistic training data.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.