Filed · Sourced · Walk-Forward Backtested

Mortgage rate: the short horizon, honestly

Multi-year rate prediction does not work, we proved that and stopped pretending. This is a short-horizon model: it forecasts the 30-year mortgage rate ~1 and ~3 months ahead from the rate's lagged catch-up to Treasury moves that already happened. Every prediction below was made walk-forward, fit only on data available at the time, and the chart overlays each prediction on what actually happened so you can see it track reality. Where the model does not beat “rate stays flat,” we say so.

30-yr Mortgage loading… … FRED MORTGAGE30US
10-yr Treasury … FRED DGS10
Spread (30y − 10y) … Mean-reversion anchor
Model edge @ ~1mo … MAE reduction vs “rate stays flat”
Near-term forecast, ~1 and ~3 months ahead
Weekly model · OLS · lag-catch-up
5-year forward projection, rates to 2031
Central path + honest p10–p90 band
Left of the line is tested history. Right of the line is the projection. The chart shows the actual 30-year mortgage rate up to today, then a 5-year forward central path with a p10–p90 uncertainty band out to 2031. The central path uses the short-horizon model for the first ~3 months (where it has real, tested skill), then glides toward a long-run anchor. The band is the empirical p10–p90 of how much the mortgage rate has actually moved over each horizon in 50+ years of FRED history, no assumed volatility. It fans out because rates genuinely move more over longer spans. This is a plausibility range, not a point prediction.
30-yr mortgage, tested history + 5-year forward cone to 2031 Actual MORTGAGE30 Model 13-wk prediction Forward central path Empirical p10–p90 cone
How we know this: model validation & methodology Historical chart · walk-forward proof · backtest scoreboard ▾
Historical rates
Weekly · FRED
30-yr mortgage vs 10-yr Treasury MORTGAGE30US DGS10
Walk-forward proof, the short-horizon backtest
Predictions vs actuals · ends at today
Walk-forward proof, every ~3-month prediction overlaid on what actually happened Actual MORTGAGE30 Model 13-wk-ahead prediction (placed at its target week)
Residual distribution, ~3-month model prediction error (predicted − actual) count of weeks

Backtest scoreboard: model vs the baselines it has to beat

Every row is a full walk-forward backtest: at each week the model is fit only on data available at the time, predicts ahead, and is scored against what actually happened. naive = “rate stays flat.” momentum = “recent trend continues.” A real short-horizon model should beat naive; if it does not, the cell is plain, not green.

Horizon / modelMAE (pp)RMSE (pp)Directional hit-ratevs naive MAE

What the model is

The 30-yr mortgage rate is the Freddie Mac PMMS weekly survey average, a smoothed series that trails the Treasury market by a few weeks. So mortgage ≈ Treasury + spread, and when the 10-yr Treasury has already moved, the survey-based mortgage rate has a predictable near-term drift to catch up. The model is an ordinary least-squares regression on three lag-catch-up features:

  • DGS10 1-week momentum, the most recent Treasury move not yet absorbed.
  • DGS10 4-week momentum, the longer recent Treasury move.
  • spread deviation, the mortgage-minus-Treasury spread relative to its own 52-week rolling mean (the mean-reversion term).

mortgage_forecast(h) = mortgage_now + OLS(features). We tested the textbook decomposition that also forecasts the 10-yr Treasury, it added noise, because rate changes are near-unforecastable even at a week. So the model holds the Treasury level fixed and predicts only the part that is real: the catch-up. We also ran this model at monthly frequency, it does not beat naive monthly. The edge is a weekly phenomenon; monthly sampling washes it out. That is why the model runs weekly.

  • No AI claim. This is ordinary least squares, transparent coefficients, shown live below.
  • Gradient boosting was an optional cross-check “if sklearn is available.” sklearn is not installed in the build environment, so OLS stands alone, stated, not hidden.
  • No lookahead leakage: at each week, training uses only pairs whose outcome was observed strictly before that week.

Live coefficients (fit on all history): …

Sources: FRED MORTGAGE30US, DGS10, DGS2, FEDFUNDS, DFF, T10Y2Y (all public domain). Model: ordinary least squares (numpy). Built …. Not investment advice.
Beyond ~6 months, we do not predict
Empirical range · not a forecast
This is not a forecast. Multi-year rate moves are genuinely unforecastable, our earlier multi-year ensemble honestly backtested to ~zero edge. Instead of pretending, the band below is the empirical range: where the rate would land if each future year repeated the 10th / 50th / 90th percentile of historical year-over-year mortgage-rate moves. It widens with time because uncertainty does. Use it as a plausibility envelope, not a prediction.
30-yr mortgage, empirical 5-year envelope from historical YoY moves Today → p50 path p10–p90 empirical range