The majority of journal articles in the investment literature make associational claims, and propose investment strategies designed to profit from those associations. For instance, authors may find that observation X often precedes the occurrence of event Y, determine that the correlation between X and Y is statistically significant, and propose a trading rule that presumably monetizes such correlation. A caveat of this inductive reasoning is that the probabilistic statement “X often precedes Y” provides no evidence that Y is a function of X, thus the relationship between X and Y may be coincidental or non-causal. The relationship is coincidental when variables X and Y appear to have been associated in the past, however that association is a statistical fluke, e.g. due to backtest overfitting. Backtest overfitting leads to false investment strategies that perform poorly out-of-sample. The relationship is non-causal when X and Y are associated even though Y is not a function of X, e.g. due to a confounding variable Z which researchers have failed to control for. Under-controlling makes it likely that the correlation between X and Y will change over time, and even reverse sign, exposing the investor to systematic losses.

Science is more than a collection of observed associations. While the description and cataloguing of phenomena plays a role in scientific discovery, the ultimate goal of science is the amalgamation of theories that have survived rigorous falsification. For a theory to be scientific, it must declare the falsifiable causal mechanism responsible for the observed phenomenon. In the context of investments, knowledge of the causal mechanism allows researchers to hedge unrewarded risks, decommission fading investment strategies before they accumulate losses, and anticipate the possibility of black-swans. Backtests and regressions cannot explain why a strategy may be successful. Only a falsifiable causal theory can give investors some confidence that the claimed performance is not based on spurious patterns, and it is hence replicable going forward.