AlphaPortfolio for Investment and Economically Interpretable AI

We propose reinforcement-learning-based portfolio management, an alternative that improves upon the traditional two-step portfolio-construction paradigm a la Markowitz (1952), to directly optimize investors’ objectives. Specifically, we enhance cutting-edge neural networks such as Transformer with a novel cross-asset attention mechanism to effectively capture the high-dimensional, non-linear, noisy, interacting, and dynamic nature of economic data and market environment. The resulting AlphaPortfolio yields stellar out-of-sample performances even after imposing various economic and trading restrictions. Importantly, we use polynomial-feature-sensitivity (and textual-factor) analysis to project the model onto linear regression (and natural language) space for greater transparency and interpretation. Such “economic distillations” reveal key characteristics/topics (including their rotation and non-linearity) that drive investment performance. Overall, we highlight the utility of reinforcement deep learning and provide a general procedure for interpreting AI and big data analytics in finance and beyond.