Earthquake precursors have long been sought as a means to predict earthquakes with very limited success. Recently, it has been suggested that a decrease in the Gutenberg–Richter b‐value after a magnitude 6 earthquake is predictive of an imminent mainshock of larger magnitude, and a three‐level traffic‐light system has been proposed. However, this method is dependent on parameters that must be chosen by an expert. We systematically explore the parameter space to find an optimal set of parameters based on the Matthews correlation coefficient. For each parameter combination, we analyze the temporal changes in the frequency–magnitude distribution for every M ≥ 6 earthquake sequence in the U.S. Geological Survey Comprehensive Earthquake Catalog for western North America. We then consider smaller events, those with a foreshock magnitude as small as 5, and repeat the analysis to assess its performance for events that modify stresses over smaller spatial regions. We analyze 25 M ≥ 6 events and 88 M 5–6 events. We find that no perfect parameter combination exists. Although the method generates correct retrodictions for some M 5 events, the predictions are dependent on the retrospectively selected parameters. About 80%–95% of magnitude 5–6 events have too little data to generate a result. Predictions are time dependent and have large uncertainties. Without a precise definition of precursory b‐value changes, this and similar prediction schemes are incompatible with the IASPEI criteria for evaluating earthquake precursors. If limitations on measuring precursory changes in seismicity and relating them to the state of stress in the crust can be overcome, real‐time forecasting of mainshocks could reduce the loss of lives.