Policy Decisions as Data: History-Only Benchmarks for Central-Bank Rate Forecasting

Central-bank policy decisions leave a persistent record of policy cycles. We develop forecasting benchmarks that use this record alone, without macroeconomic indicators, market prices, or communication text. A Gaussian recency score provides a transparent local baseline. A state-conditioned policy network combines domestic histories with lagged foreign decisions and their forecast innovations; a causal error filter then reallocates reversal predictions within each month. Using BIS policy-rate histories for 37 economies, we evaluate cut, hold, and hike forecasts on 2,490 eligible country-months from January 2019 to August 2026. The balanced pipeline attains macro-F1 of 0.614 and correctly predicts 37.7% of policy reversals. These results provide a history-only reference for richer information sets, while showing that reversals remain difficult. The analysis is exploratory: the evaluation period informed model development, and the incremental contribution of foreign histories has not been isolated by a matched ablation of the final pipeline.