UCLA

This thesis studies the relationship between return prediction and portfolio optimization for illiquid assets. Traditional predict-then-optimize models first predict asset returns and then use those predictions as inputs to a portfolio optimizer. However, a model with lower prediction error does not always lead to better portfolio decisions. Due to the noisy signals, limited data, and liquidity-related frictions, this issue is especially serious for illiquid assets, where small prediction errors may lead to different allocation decisions.To address this problem, this thesis develops a Decision-Focused Fine-Tuning (DFF) framework based on a pretrained return prediction model. In this framework, a correction layer is inserted between the prediction model and the portfolio optimizer. This layer makes controlled adjustments to the original return predictions, and the adjustment size is controlled by the parameter ε, which bounds the maximum relative adjustment applied to the original predictions. Then the adjusted predictions will be passed into a differentiable Markowitz-style portfolio optimization layer, and the correction layer is trained to reduce downstream decision regret.Experiments are conducted on illiquid stock portfolios, corporate bond portfolios, and mixed stock--bond portfolios. The results show that decision-focused fine-tuning can improve downstream optimization quality compared with a traditional two-stage baseline. The experiments also demonstrate that different settings of ε lead to different levels of improvement, highlighting the importance of this parameter in balancing predictive stability and decision quality. Overall, this thesis shows that the Decision-Focused Fine-Tuning framework provides a practical approach for linking return prediction with portfolio optimization in illiquid financial markets.