The Upend of Theory: Correlationism and Stock Prediction in the Era of Machine-Driven Big Data Analytics

In his now famous 2008 Wired Magazine article ‘The end of theory: the data deluge makes the scientific method obsolete’, Chris Anderson, the magazine’s editor-in-chief at the time, provocatively proclaimed that scientific practice was about to be disrupted by big data analytics approaches to problem-solving, which would reduce established scientific methods to relics of the pre-big data science era. Since its formulation, Anderson’s end of theory hypothesis has been met with much criticism and even ridicule by STS, critical data studies, and other scholars studying big data and machine learning use in science, industry, and society. However, as acknowledged by most critics of Anderson’s hypothesis, theory-void big data-analytics approaches are increasingly being presented as superior in problem-solving, regardless of whether problems lie within the realms of science, business, politics, or elsewhere. In this paper, I do not want to argue an impending obsolesce of theory, but instead call to caution against a flattening of categories and instrumentalization of theory in- and outside the confines of the world of science. Evidence of this epistemological shift from theory- and analytical category-driven approaches to research towards more instrumental big data-driven correlationism, exudes from many places—Bruno Latour’s almost entirely uncritical hailing of big data network analysis as a window to the real workings of the social being an illustrative example from STS—, yet I focus on experimental deep neural network-models for stock prediction. Examining the way theory—more specifically, the so-called Elliott Wave Principle (a scientifically questionable theory developed by the American accountant Ralph Elliott in 1934)—is perceived and used in academic articles leveraging such models in stock price prediction experiments. The theory-use in these data science stock-prediction studies dilutes, I argue, the concept of theory and reduces it to a mere name in a recognizable pattern in a dataset.