Asset Demand Systems via Data Augmentation: Competition and Differentiation in Asset Management

Many institutional investors hold portfolios with few holdings. This makes it challenging to precisely estimate their individual demand. In this paper, I seek to make two contributions. First, I propose a data augmentation technique based on the generation of data-driven and economically interpretable synthetic assets. I show that this data augmentation acts as an adaptive nonlinear shrinkage which automatically adjusts the shape of the penalty to the cost of overfitting faced by the nonlinear demand function estimator. The resulting estimation technique leads to substantial improvement in cross-out-of-sample R2 for estimation of both low-dimensional and high-dimensional demand functions. Second, I use the proposed methodology to construct a measure of investor differentiation. Using the Morningstar mutual fund ratings reform in 2002 as a shock to competition for alpha, I show that mutual funds escape the increased competition intensity by differentiating from their competitors.