The science behind the map

What the analysis found

• Demography predicts vote similarity far better than winners. A riding's top-10 demographic neighbours are also its top-10 vote twins of the time — about random chance.
• Persistence is the strongest filter. Only ~a quarter of vote-twin pairs survive between elections ( overlap), yet stable pairs are far more predictable from demographics than one-election pairs — at riding level AND at poll level, where stable twins double predictability.
• The signal replicates within ridings. Between polls of the same riding — zero regional confounds — demographic similarity predicts vote similarity at ~2× random in every election, in all four provinces.
• What carries it: between ridings, rural/exurban economic structure, language and incumbency; within ridings, an affluence-and-diversity gradient.
• Honest negatives: turnout adds nothing; open seats are not easier to read demographically; riding-level financing data has coverage artefacts.

Method in one paragraph

"Voting twins" = pairs of ridings whose six-party vote-share profiles are close (cosine similarity). The "demographic lens" compares ridings across ~80 census characteristics (2021 Census, 2023 Representation Order). We measure how often a riding's demographic neighbourhood coincides with its vote neighbourhood — precision@K — against the K/(n−1) random baseline. Full reports: voting_twin_variables_v2.md, voting_twin_panel.md, voting_twin_poll_level.md.

Which variables carry the signal

Nonzero learned weights, GE45, all provinces (voting_twin_variables_v2.md).

Across elections (both representation orders)

Within ridings, at poll level