A simple linear fit looks stable in a quiet window, then breaks across zero-lower-bound policy, shifting inflation dynamics, and post-2020 regime change.
The fit looks excellent in the calm 2005-2007 window (RMSE 0.32 pp), which creates the illusion that a simple linear mapping from inflation and unemployment to the policy rate is stable.
Once the economy hits the 2008-2015 zero-lower-bound regime, the model keeps wanting much lower rates than the policy framework can deliver, pushing GFC-era RMSE to 7.85 pp.
After 2020, the same coefficients are exposed again: unemployment, inflation, and the funds rate move through combinations that were not present in the quiet sample, driving the largest miss to 30.532 pp in Apr 2020.
The first three charts are time-series views: policy-rate fit, rolling CPI-unemployment correlation, and residuals. All three repeat the same regime shading so the breakdown is easy to compare across views.
The regime error bars quantify the overfitting problem directly, while the scatter plot shows how the inflation-unemployment feature space itself shifts across periods. Marker size scales with absolute prediction error, so the biggest points are the months where the naive fit breaks most visibly.
| Date | Regime | Actual Fed Funds | Naive Prediction | Prediction Error |
|---|