We establish a statistical learning framework for individualized asset allocation. A high-dimensional Q-learning methodology is developed for continuous-action decision making. We show that our proposed model parameter estimator enjoys desirable theoretical properties, and our approach allows for valid statistical inference for the optimal value associated with the optimal decision making. Empirically, the proposed statistical learning framework is exercised with Health and Retirement Study data. The results show that our optimal individualized strategy improves individual financial well-being and surpasses benchmark strategies under a consumption-based utility framework. We discuss further the multiple asset allocation without resorting to individualized information. When indiv...[ Read more ]

We establish a statistical learning framework for individualized asset allocation. A high-dimensional Q-learning methodology is developed for continuous-action decision making. We show that our proposed model parameter estimator enjoys desirable theoretical properties, and our approach allows for valid statistical inference for the optimal value associated with the optimal decision making. Empirically, the proposed statistical learning framework is exercised with Health and Retirement Study data. The results show that our optimal individualized strategy improves individual financial well-being and surpasses benchmark strategies under a consumption-based utility framework. We discuss further the multiple asset allocation without resorting to individualized information. When individualized information is absent, we focus on high-dimensional minimum variance portfolio (MVP) and study the estimation of MVP under statistical factor models. A unified MVP estimator is developed and shown to enjoy sharp risk consistency theoretically. Our proposed MVP estimator performs favorably over various benchmark portfolios in numerical studies.