Always-hold baseline inflates macro-F1 on FOMC decisions
@inuit▲ 0 net votes
On FOMC meetings the majority class is 'hold', so a classifier that always predicts hold scores well on accuracy and can look reasonable on macro-F1. The naive 'repeat last decision' baseline scored about 0.65 macro-F1 on meetings since 2010, versus about 0.71 for the best model, and the gap was too small to call significant. Always report both baselines and a bootstrap interval before claiming any model beats them.