This paper theoretically analyzes the operating principles of Decision-Driven Learning (DFL), which emerged to address the challenges of estimating the expected value, variance, and covariance of uncertain asset returns within Markowitz's mean-variance optimization (MVO) framework. We highlight the limitations of existing machine learning-based forecasting models, which fail to account for correlations between assets when minimizing the mean squared error (MSE), and demonstrate how DFL overcomes this limitation. By analyzing the gradient of DFL, we demonstrate that DFL incorporates correlations between assets into the learning process by weighting the MSE-based errors by multiplying them by the inverse covariance matrix. This induces systematic forecast biases that overestimate the returns of included assets and underestimate those of excluded assets. However, we demonstrate that these biases actually contribute to achieving optimal portfolio performance.
