The two views
Basic is the default guardrail against overfitting. It uses the same CMO expectation as Fundamentals+, but shrinks the district adjustment four-fifths of the way toward zero. The current generic-ballot environment is D+7.55 across 6 quality-gated pollsters.
The CMO expectation uses district racial and educational composition, prior presidential trend and its availability, chamber, and 2008/2016 era terms. It deliberately excludes candidate incumbency, fundraising, ideology, and prior CMO so downstream consequences of candidate strength are not mistaken for independently known candidate effects.
National environment
The polling signal comes from quality-gated generic-ballot polling. Historical tests support transferring none of the swing before 2018, half in 2018, and the full swing in 2022 and 2026. The 125% continued-nationalization case remains a separate experimental scenario, not a selected model.
Forward validation
Every holdout is trained only on earlier cycles. The table reports cycle-balanced mean absolute error across the expanding-window tests, the post-2016 holdouts, and 2022 separately.
| View | Mean MAE | 2018–22 | 2022 | Status |
|---|---|---|---|---|
| Basic | 22.88 | 10.98 | 9.54 | Default |
| Fundamentals+ | 22.85 | 10.86 | 9.16 | Experimental |
Fundamentals+ applies the complete historical expectation and is the richer substantive view. Basic remains the default because its 20% adjustment is a transparent hedge against selection noise and structural change in a small number of election cycles.
Probabilities and chamber summaries
A recent-era Southern calibration converts the selected expected margin into a win probability using P(D win) = Φ(expected Democratic margin / 6.0). The six-point scale was fitted on 1,188 contested legislative races outside Alabama. Forward-cycle validation produced a 0.0310 Brier score and 96.1% winner accuracy; leave-state-out validation produced a 0.0332 Brier score and 95.8% winner accuracy.
The middle-80% district margin band is the corresponding conditional predictive interval. Chamber distributions combine district probabilities as independent Bernoulli outcomes and hold single-major-party seats fixed. They are conditional on the displayed national environment; national polling uncertainty belongs in explicit shared scenarios.
Limitations
- Only a small number of forward election environments—and two after 2016—identify these relationships.
- The CMO expectation is contextual; it does not identify an individual candidate effect.
- National race and education crosstabs are not joint race-by-education samples and Alabama polling remains sparse.
- Polling house effects and mode effects are not yet estimated explicitly.
- Probabilities remain provisional rather than fully calibrated long-run frequencies.