Extracting Stock Predictive Information in Mutual Fund Managers’ Portfolio Decisions Through Machine Learning with Hypergraph

Abstract This paper proposes a machine learning framework that incorporates mutual fund managers’ portfolio decisions to predict stock price movements. It employs a weighted hypergraph to connect the information of the funds’ portfolio weight changes to the corresponding stocks in the portfolio. This information is further aggregated with price data extracted through classical machine learning methods. The framework then predicts stock price movements as a classification problem. We tested this framework using Taiwanese stock and mutual fund data, discovering that the managers’ decisions provide predictive information about stocks in addition to the classical technical data. Moreover, we determined that the level of information provided by the managers’ decisions is asymmetric under different market conditions. This finding supports existing literature on mutual fund behavior under varying market conditions. Further discussions on incorporating other expert trading behaviors in Taiwan and model comparisons are provided.