The EZ Dub Election Forecast

How to read this

• What's driving this prediction: open any race (What's driving this?) for the inputs that move the model most in it, which candidate each helps and by how much, and both candidates' values. Every input the models use is listed above.
• Win probability comes from the model's win classifier, normalized so the candidates in a race add up to 100%. It is trained on Senate, governor and House races since 1998 and judged on how well it would have called 2018-2024 using only earlier data.
• Projected margin comes from a separate margin model (half the gap between the top two candidates' projected results). The two models can occasionally disagree on a close race.
• Ratings: the leader at 95%+ is Safe, 80-95% Likely, 60-80% Lean, under 60% Tossup.
• Markets are the average of Kalshi and Polymarket for the same outcome, shown only for comparison. They are not an input to the model.
• No polls: races no pollster has covered (most House seats) use the no-polls model, trained on the same races since 1998 without their polls. It weighs which party holds the seat, the seat's past results (the new map's lean where districts were redrawn for 2026), the national environment, fundraising and the candidates' experience. Most of these are safe seats, and its picks match the markets' in nearly every one.