Research working paper · Revised October 4, 2026
Do learned weights help when election polls are scarce?
We tested several ways to combine polls with background information. The more complicated method has not earned its extra complexity.
The study covers six past elections, from 2012 to 2022, in both the House and Senate. Each test used earlier elections to fit the model. However, we developed the methods while knowing past outcomes, so these are development results. They are not proof of future accuracy.
What we found
Adding candidates’ past federal elected experience explains much of the improvement over a simpler regression. A fixed blend often beat the learned blend on probability accuracy. The learned blend’s weights barely varied. Its apparent advantage in uncertainty ranges also needed a fairer calibration check.
We will test the simpler experience regression as the main research method. We retain the old results and report revised comparisons separately. The separate chamber experiment shows possible seat totals, but does not publish a chance of controlling Congress.
Read the paper (PDF) · Word copy · Markdown source
Download the revised NotebookLM source pack. Unzip it and upload the revised guide and paper as sources for a new audio overview.
What changed in this revision?
Chamber audit update: Appendix D documents repaired candidate-name matches, additional ballot exclusions and the model’s historical tendency to overpredict Democratic House seats. The older editorial call is under review. Reproducing a calculation is not the same as proving forecast accuracy.
The paper now explains Compass, the Tide challenger, the separate poll-weighting tests and the whole-chamber scenarios. The original six-election study is unchanged. A new appendix reports the later work separately, including a different seven-election check on subsets of seats.
More weighting did not clearly help. Shared-error scenarios show how several races can miss together, but their full-chamber ranges are not yet validated. Google search attention does not enter the forecasts. No 2026 outcome test is complete.
The new appendix also records a low-confidence editorial call: a Democratic House and Republican Senate. This earlier judgment is now under review following the chamber audit; it is preserved as a dated record, not a current call.
Explore the seat ranges and maps · Compare the two approaches
Written and checked with AI
Codex: AI builder and manuscript writer. Claude Code: AI auditor. People at Lampbotics set the goals, oversee the work and remain responsible. Claude Code verified the earlier numerical results and raised the issues that prompted this revision; its October 4 chamber audit reproduced the v2 scenario numbers but found input and reporting problems. The v3 repairs await follow-up review.
This is a working paper, not peer-reviewed research. Naming the AI contributors describes their actual roles; it does not claim journal-approved authorship, human review or demonstrated forecasting superiority.