30-year fixed mortgage and 10-year Treasury since the dataset starts. The forecast is an honest ensemble — four named models (AutoARIMA, AutoETS, Theta, SeasonalNaive) fit on annual data since 1995, median-blended, with bands drawn from the empirical 10th/90th percentile of historical YoY changes. Every model is shown and every model is backtested below.
The gold center line is the median of four named forecast models, each fit
independently on annual MORTGAGE30US since 1995:
AutoARIMA, AutoETS (additive trend, no seasonality),
Theta, and SeasonalNaive. All four are shown faintly so
nothing is hidden. We do not claim AI — these are classical statistical
forecasters from the statsforecast library.
The shaded envelope is the asymmetric empirical band:
centert + p10 · √t below,
centert + p90 · √t above, where
p10 and p90 are the 10th and 90th percentile of
historical year-over-year changes in MORTGAGE30US over the last 30 years.
This replaces the old symmetric ±1.5σ Normal assumption with
something that reflects actual asymmetric historical moves.
For each model, we fit on data through year t, predict the next 1/3/5/10 years, and check whether the realized rate fell inside that model's p10/p90 band. Coverage is the share of walk-forward observations that landed inside. Any 5-yr coverage below 70% is flagged in red as narrower than honest.
| Model | 1y | 3y | 5y | 10y |
|---|