Election forecasts should be judged by long‑run calibration and true out‑of‑sample performance rather than memorable single‑event 'hits' or backtests. A model that systematically predicts probabilities consistent with observed frequencies (and is tested on genuinely new data) is more trustworthy even if it 'misses' one high‑profile race.
— Shifting attention from headline 'right/wrong' calls to calibration reduces misinterpretation of forecasts, improves media reporting, and protects civic trust during tightly watched contests.
Nate Silver/FLIPR's discussion of calibration, backtesting versus true out‑of‑sample predictions, and the 18‑year track record of published, scrutinized forecasts.