Prévisions
Où se dirige l'appui des partis — pas seulement où il en est aujourd'hui. Construit sur les fondamentaux économiques et les sondages récents, avec des marges d'incertitude calibrées.
How the Forecast Works
A plain-English summary of the Trajectory Model and how it produces the projections above.
The core idea
Polls are noisy snapshots of public opinion. The Trajectory Model treats each party's true support as a hidden variable that drifts gradually over time, pulled toward a “fundamental” level set by the economy. This is a Bayesian state-space model (think of it as a Kalman filter with economic covariates and full posterior uncertainty).
Economic fundamentals
The model learns from 20+ years of monthly data how GDP growth, inflation (CPI), unemployment, and time in office historically relate to governing-party support. These coefficients anchor a long-run equilibrium: if the economy improves, the model expects support to drift upward toward that level over time.
Mean-reversion (AR(1))
Support doesn't jump to the fundamental level overnight. A mean-reversion parameter κ controls how quickly it converges, typically over several months. Parties far from their fundamental tend to drift back; those close to it stay put. This dampens extreme poll swings in the forecast.
Valence signal
News sentiment around the governing party — measured daily from Canadian media and smoothed with a 28-day EWMA — enters the Bayesian trajectory model directly as a fitted covariate (γ_valence ≈ 0.05). Favourable coverage nudges the fundamental support level up; negative coverage nudges it down. The effect is modest by design: economic fundamentals dominate over the medium term, and the sentiment coefficient is calibrated from data, not assumed.
Blended forecast
The trajectory model's 3-month projection is blended with the current poll aggregate. At short horizons the polls dominate; at longer horizons the economic model carries more weight. This “blended” estimate is what drives the seat projection and riding win probabilities on the map.
Prediction intervals
The shaded bands on the chart are 80% credible intervals from the Bayesian posterior. The p10–p90 intervals in the accuracy tracker are further calibrated using split conformal prediction: interval widths are fit on a held-out historical period and evaluated prospectively, so coverage statistics are out-of-sample.
Implemented in PyMC · trained on Statistics Canada + Bank of Canada economic series · poll aggregate from Polling Canada
Literature: Lewis-Beck (1988), Erikson & Wlezien (2012), Daoust & Dassonneville (2018), Vovk et al. (2005)