Abstract
Alphas are stock prediction models capturing trading signals in a stockmarket. A set of effective alphas can generate weakly correlated high returnsto diversify the risk. Existing alphas can be categorized into two classes:Formulaic alphas are simple algebraic expressions of scalar features, and thuscan generalize well and be mined into a weakly correlated set. Machine learningalphas are data-driven models over vector and matrix features. They are morepredictive than formulaic alphas, but are too complex to mine into a weaklycorrelated set. In this paper, we introduce a new class of alphas to modelscalar, vector, and matrix features which possess the strengths of these twoexisting classes. The new alphas predict returns with high accuracy and can bemined into a weakly correlated set. In addition, we propose a novel alphamining framework based on AutoML, called AlphaEvolve, to generate the newalphas. To this end, we first propose operators for generating the new alphasand selectively injecting relational domain knowledge to model the relationsbetween stocks. We then accelerate the alpha mining by proposing a pruningtechnique for redundant alphas. Experiments show that AlphaEvolve can evolveinitial alphas into the new alphas with high returns and weak correlations.