Generative Scenario Design for Explaining Predict-Optimize Portfolio Pipelines Across Tactical and Strategic Asset Allocation

This thesis develops a scenario-based explanation framework for learned portfolio decision pipelines. Starting from the Predict-Optimize-Explain perspective, the thesis treats a trained prediction-optimization system as a fixed mapping from economically interpretable inputs to portfolio decisions, and explains this mapping by solving inverse questions in the macroeconomic input space. Instead of asking only which variables predict returns, the thesis asks which macro-financial conditions would induce specific portfolio behavior, such as meeting a return target, becoming more diversified, shifting toward defensive assets, or making alternative decision pipelines diverge. Empirically, the thesis applies this explanation logic to two complementary portfolio settings. The first is a tactical/firm-level asset allocation setting based on firm characteristics, macroeconomic variables, and their interactions, comparing predict-then-optimize and decision-focused pipelines. The second is an institutional strategic asset allocation setting, where macro-financial states are used to predict long-horizon asset-class returns, construct a robust long-only portfolio, and generate plausible macroeconomic scenarios. The thesis contributes a unified empirical account of how generative scenario design can support decision-level explanation in portfolio optimization.