2026 House midterm forecast

This project is a district-level 2026 U.S. House forecast with:

• a full generic-ballot archive
• a daily state-space national filter
• a weak Trump-approval cross-check in the national layer
• structural district priors
• Monte Carlo chamber simulation
• a Streamlit UI built around auditability, not just toplines

What changed in this build

This version fixes the main comparability problem in the prior trend view and expands the national polling layer.

Main upgrades

• keeps the current district and campaign inputs fixed through the reconstructed daily odds line so the endpoint no longer jumps because the model definition changed
• adds Trump approval as an explicit national cross-check in the model rather than only a display metric
• writes a dedicated Trump approval curve with separate approve and disapprove lines plus 90% bands
• surfaces both the raw filtered generic-ballot line and the approval-adjusted national House environment used in simulation
• records endpoint-gap diagnostics in the audit output so any future discontinuity is visible immediately

Files

• app.py — Streamlit app
• src/data_sources.py — source ingestion and seed/live loaders
• src/model.py — national filter, district priors, simulation engine, output writer
• scripts/run_update.py — batch runner
• data/seed/generic_ballot_polls_master.csv — bundled generic-ballot archive
• data/seed/trump_approval_recent_polls.csv — bundled recent Trump approval rows for seed mode
• data/runtime/latest/summary.json — latest topline output
• data/runtime/latest/run_audit.json — latest audit manifest
• data/runtime/latest/trump_approval_curve.csv — latest approval curve for the UI
• data/history/forecast_history.csv — daily comparable odds history

Methodology in one pass

1. Generic-ballot archive

The main national input is a dated generic-ballot archive assembled from RealClearPolling / RealClearPolitics pages. In seed mode, the model reads the bundled archive directly. In live mode, it refreshes the current RealClearPolling average and appends newly published rows when possible.

2. Latent national House environment

Each generic-ballot poll row is treated as an observation of a latent national House environment. The model uses a local-level Gaussian filter:

• latent margin today = latent margin yesterday + process noise
• poll row = latent margin + observation noise

Observation uncertainty is larger when:

• the row is adults instead of LV
• the row lacks exact field dates
• the row lacks sample size or population metadata
• the row is flagged as partisan or internal

3. Trump approval as a weak cross-check

Trump approval is now explicitly in the model, but only as a weak prior on the national House environment. The model converts Trump net approval into an approval-implied House margin using a shallow slope, then combines that with the filtered generic ballot using a large prior SD.

That is deliberate. Approval and the generic ballot are strongly correlated, so letting approval hit the model too hard would double-count the same national mood. In this build, approval mostly matters when generic polling is sparse or unusually noisy.

4. Comparable daily trend line

The continuous odds line is not a historical archive of exact saved forecasts. It is a reconstructed daily curve. To make that line internally comparable, the project now holds today's campaign structure fixed through the full trend reconstruction:

• current open seats
• current expert ratings
• current finance inputs
• current approval cross-check

What changes day by day is the national generic-ballot state and any dated district polls. Exact saved runs are still preserved separately in run_history.csv.

5. District priors and chamber simulation

District forecasts start from:

• 2024 House results
• presidential vote by district when available
• persistence rules for incumbent, successor, and open seats
• optional ratings, district polling, and finance

The chamber forecast then runs Monte Carlo draws with:

• national uncertainty
• correlated state error
• district-specific uncertainty

Outputs include:

• GOP odds to hold the House
• expected GOP seats
• interval estimates
• full seat distribution
• district-level win probabilities and projected margins

Running locally

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python scripts/run_update.py --seed-only --no-fec --no-district-polls
streamlit run app.py

Notes on accuracy

This build is materially better than a thin recent average or a one-point saved-run chart, but it is still not a full multi-cycle calibration system. The strongest next upgrade would be cycle-by-cycle backtesting so pollster house effects, horizon error, process variance, and approval-to-House link strength can be estimated empirically rather than set conservatively.