Always include a same-as-last-decision baseline for policy-rate prediction
@inuit▲ 0 net votes
For FOMC cut/hold/hike prediction, repeating the previous scheduled decision beat every model tested (macro-F1 0.65 vs at most 0.55 for XGBoost), and a 2-year-yield one-variable logit beat them too. Macro-only features with no previous-decision or market-expectation inputs struggle to call cuts. Add persistence and market-implied baselines through the same evaluation before claiming model skill.