Raphaël Glucksmann

Polls from different pollsters are not directly comparable: each has its own method and usual bias, which the model corrects in part. The spread between dots on the same day reflects this diversity as much as statistical uncertainty.

The three nested bands hold 50, 80 and 95% of the draws of the estimated state, carried forward with the same random walk as the opinion model; they contain neither future polls nor the vote itself. Validation shows this layer's intervals are too narrow: read them as a lower bound on uncertainty. The table under the chart gives the same values as numbers.

A falling probability does not say the candidate is "losing": it says that, given the known polls, a smaller share of simulations has them winning. The model knows neither manifestos nor events, only the numbers.

Head-to-head polls test specific pairs on a given date; the model, for its part, simulates all possible pairs from the first round and vote transfers. The two figures can therefore differ without contradicting each other.

The number of mentions measures press attention, not opinion: an article may be favourable, critical or purely factual, and the site does not judge. The outlets tracked cover the whole spectrum; their list is on the "Press" page.

A betting market reflects money stakes, with their biases and liquidity. It is neither a poll nor the truth: it is another source, shown next to the model so that everyone can compare.