Track Record
Backtested against the 2024 General Election. Each model improvement narrows the error. Transparency about what we get wrong builds trust in what we get right.
Improvement journey
Basic UNS model
Uniform national swing from polling averages. Wrong data for some seats, no house effect correction.
House effects + calibration
Added house effect correction, poll recency weighting, and regional calibration. All 12 remaining errors were tight three-way marginals.
Full model pipeline
Vote source models, turnout adjustment, tactical voting. The 4 remaining errors are all independent wins that no national model can predict.
The 4 we still get wrong
All four errors are independent wins driven by hyperlocal dynamics invisible to any national polling model. This is the honest limit of what polling-based forecasting can achieve.
Hyperlocal independent candidacy with strong community support
Independent candidate with local profile beyond polling capture
Local independent campaign invisible to national model
Strong independent challenger not captured by any polling
What the backtest tells us
The model correctly assigns seats in 99.4% of cases using only pre-election polling data. The multi-layer approach (regional swing + vote source + turnout + tactical) adds 94 correct seats over simple uniform national swing.
The remaining 4 errors represent the genuine floor of national-model accuracy. Independent candidates are a local phenomenon that no amount of polling sophistication can predict. We show this honestly rather than claiming perfect accuracy.