AlphaEvolve

AlphaEvolve is a framework that automatically mines novel alphas with high returns and low correlations with an existing set of alphas.

We first look at an example of a pair of formulaic alphas in a hedge fund. Based on experts’ domain knowledge, alphas are designed manually from different perspectives. The formulaic alphas are backtested on the markets and used for investment. The low correlation is achieved by domain experts’ knowledge.

AlphaEvolve can evolve a well-designed alpha automatically. An overview of AlphaEvolve architecture is shown below. A well-designed alpha is first transformed into the formulation of AlphaEvolve with operators and operands. During evolution, validity checks (i.e. AFEC) are performed to keep the valid (part of) alphas. Then alphas are hashed to avoid repetitions. After AFEC, mutated alphas are evaluated and eliminated if they are correlated. Finally a novel alpha is discovered with low-correlated returns.

Next we demonstrate an example alpha mined by AlphaEvolve below. Each operand (i.e. a value in the equations) is marked with a letter to represent the data type and a number starting from 0. For example, operand s5 is the sixth scalar operand. Special operands are the input feature matrix m0, the output label s0, and the prediction s1. The alpha makes a prediction at time t by the equation in Equation Set M. The parameter M2_t-2 is updated recursively with the input feature matrix in Equation Set U. S2_t-2 is a trend feature based on the comparison between high_price t-4 and a recursively compared feature of high_price_t-5 in Equation Set P. Thus the alpha makes trading decision based on the volatility of the historically summarized features in the training period, the trend feature based on high prices and the return.

View our Raptor publication on the ACM Special Interest Group on Management of Data (SIGMOD) 2021:

AlphaEvolve: A Learning Framework to Discover Novel Alphas in Quantitative Investment [ paper link]

(A patent on AlphaEvolve is being filed.)