By Yu Huang
With the development of super-computers and sophisticated statistical models, automating the alpha search becomes possible. Automated alpha search is a method using computer algorithms to find signals out of the huge data cloud. When implemented properly, it can significantly boost the efficiency of alpha search, producing thousands of alphas in a single day. This comes with a price: not all signals found are real alphas. Many of the seemingly great alpha signals found by automated search are noise and don’t have any predictive power. Thus, special efforts have to be made in the input preparation, search algorithm, signal testing algorithm, etc., to improve the robustness of these alphas.
In general, there are three components in an automated search: input data, search algorithm, and signal testing (see Figure 17.1).
First, we select a group of meaningful financial data, such as price, earnings, news, etc., as input predictor variables. The predictor variables are usually pre-processed to remove outliers before feeding into the search algorithm. Then we select a target function Y, which represents the future stock returns or its variants. The fitting algorithm then finds the parameters of a group of pre-selected family of functions (the simplest example being the linear functions), which approximate ...