US housing starts · United States · 1 month ahead

2026-09: the model assigns an 80 percent probability to 1,185 to 1,423 thousands of units, seasonally adjusted annual rate

Point estimate 1,286 thousands of units, seasonally adjusted annual rate from Holt (exponential smoothing with additive trend), against 1,275 thousands of units, seasonally adjusted annual rate observed in 2026-08. holt beats last_value on out-of-sample MAE by 16.2 percent over 47 cutoffs (minimum 24, margin 5 percent).

ModeledMetric page

Distribution over the history

Fan chart

Solid: observed (thousands of units, seasonally adjusted annual rate). Dashed and shaded: model point estimates and 80 percent intervals from empirical residual quantiles. Filled dot: the record on this page. Green or red dot: the realised value, inside or outside its interval.

Quantiles: p5 1,159 · p10 1,185 · p25 1,232 · p50 1,293 · p75 1,351 · p90 1,423 · p95 1,450
Source: empirical quantiles of 36 in-window 1-step residuals, in log space; no normality assumed

The record

Immutable fields

Key: us-housing-starts:USA:1
Id: a6ab435e-d2d5-4a89-9e8d-07902641b658
Created: 2026-10-01 00:03 UTC
Data cutoff: 2026-10-01 00:03 UTC
Origin: 2026-08 = 1,275 (obs 410402)
Target period: 2026-09 (31 days)
Point: 1,286 thousands of units, seasonally adjusted annual rate
Interval: 1,185 to 1,423 at 80 percent
Model: holt models_v1.targets_v1
Baseline: last_value
Snapshot: 0736dba7ca86f833894dc43cda84fdf8141fcb4ea4591f72f87ae528292e4e6d
Data mode: ingested_at
Status: active

Drivers

Statistical association, not causal

Holt (exponential smoothing with additive trend) uses only the target's own history; there are no driver contributions.

Outcome

Not yet realised

The observation for 2026-09 has not been ingested. The weekly forecast.evaluate job attaches it when it arrives; the record above does not change.

Scenario assumptions

What the forecast takes for granted

No structural break in us-housing-starts between the data cutoff and the target period.
The source keeps publishing the series on the same definition; a rebasing or redefinition invalidates the comparison.
The smoothed level and trend at the cutoff persist over the horizon.

Falsifiers

What would show this forecast wrong

If the first published value of us-housing-starts (USA) for 2026-09 is below 1,185 thousands of units, seasonally adjusted annual rate or above 1,423 thousands of units, seasonally adjusted annual rate, the 80 percent interval is falsified.
The model assigns a 10 percent probability to a value below 1,185 thousands of units, seasonally adjusted annual rate and a 10 percent probability to a value above 1,423 thousands of units, seasonally adjusted annual rate; interval misses should occur about one time in five over many forecasts, and a run of misses well above that rate falsifies the calibration.
The point forecast is above the last observed value (1,275 thousands of units, seasonally adjusted annual rate); a realised value at or below that level falsifies the direction call.

Competing models

Scorecard for this target and horizon

Model	Status	Cutoffs	MAE	RMSE	Direction	80% coverage	Skill vs last value
Holt (exponential smoothing with additive trend) models_v1.targets_v1	champion	47	78.47	103.5	74%	77%	16.2%
Drift models_v1.targets_v1	challenger	47	93.71	118.1	51%	77%	-0.1%
Last value models_v1.targets_v1	baseline	47	93.66	118.1	0%	77%	0.0%
Linear trend models_v1.targets_v1	challenger	47	98.12	129.7	68%	79%	-4.8%
Ridge regression on lags and drivers models_v1.targets_v1	challenger	47	92.36	117.1	51%	77%	1.4%
Seasonal naive models_v1.targets_v1	challenger	47	126	168.3	64%	81%	-34.5%

Model card

Holt (exponential smoothing with additive trend)

Purpose. Point and 80 percent interval forecasts of us-housing-starts (USA) 1, 3, 6, 12 months ahead. Owner: Forecast Lab (HV 3.0).

Method

Simple exponential smoothing with an additive trend. Level and trend smoothing parameters (alpha 0.7, beta 0.05) are chosen by grid search on one-step-ahead squared error inside the training window. The series is modelled in natural logarithms and transformed back, so intervals are asymmetric in level terms.

Uncertainty

Intervals are empirical: the model is refitted at earlier origins inside the training window, the realised errors at the same horizon are collected, and the interval is the 10th to 90th percentile of those errors around the point forecast. No normality is assumed.

Training and evaluation

Training window: 812 periods, 1959-01-01 to 2026-08-01. Evaluation: rolling-origin backtest over 47 cutoffs using pseudo real-time data (the earliest ingested value per period; no vintage archive exists for this series, so revisions between the cutoff and today are not reproduced).

Known limitations

Structural breaks after the cutoff are not modelled.
Driver contributions are statistical associations, not causal effects.
Promotion to champion requires beating the last-value baseline on out-of-sample MAE by at least 5 percent over the minimum number of cutoffs; otherwise the baseline stays champion.
Retraining: every scheduled forecast.run refits on the data then available; backtests re-run from /admin/forecasts.

Version chain

Supersedes

4ca65e9b 2026-09-07 22:16 UTC · holt · 1,269 [1,159, 1,404] · superseded

this version: new data at cutoff