This paper theoretically analyzes the operating principles of Decision-Driven Learning (DFL), which emerged within Markowitz's mean-variance optimization (MVO) framework to solve the problem of estimating the expected value, variance, and covariance of uncertain asset returns. 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). We demonstrate that DFL incorporates correlations between assets into the learning process by weighting the MSE-based forecast errors by multiplying them by the inverse covariance matrix. In this process, DFL creates systematic biases that overestimate the returns of assets included in a portfolio and underestimate those excluded. We show that this bias is the reason DFL achieves superior portfolio performance despite higher forecast errors. In other words, we emphasize that strategic biases are a feature, not a defect.
