A predict- then-optimize approach for portfolio selection = Phương pháp dự báo sau đó tối ưu cho việc lựa chọn danh mục đầu tư

Abstract

Many real-world financial analytic problems involve two challenged and complexed tasks: predict and optimization, which leads to the standard paradigm is predict-then-optimize. Typically, these two tasks run separately which means that the machine learning models try to minimize the prediction error and do not consider how such predictions will be used for the downstream optimization problem. To tackle it. in this paper we introduce the Smart Predict then Optimize (SPO) framework which leverages the optimization structure for designing a better prediction model. Moreover, for the optimization structure, we apply Distributionally robust chance constrained optimization over Wasserstein ambiguity sets to define the feasible solution on a set of conic constraints. We train our model under subgradient methods, Finally, we evaluate our approach with least square loss to compare the performance between them.