The ensemble of approaches behind every ParallaxEdge forecast, and how they are weighted.
No single model captures every dimension of a match. A Poisson goal model is excellent at scoring rates but blind to momentum; an Elo system tracks strength over time but ignores expected goals. We blend several, each weighted by its demonstrated contribution to out-of-sample accuracy.
Weights are not fixed. As models prove themselves across more fixtures, their influence is adjusted, and the full weighting is documented openly.
Our primary engine is a Bayesian hierarchical model in the Dixon-Coles tradition, which estimates team attack and defense strengths while accounting for the correlation between low-scoring outcomes that simpler Poisson models miss.
The Bayesian framing matters because it carries uncertainty through the whole pipeline. We do not just produce a number; we produce a distribution, which is what makes honest volatility and confidence measures possible.
Around the core sit an Elo rating layer, an xG-projection layer, and adjustments for venue, rest days, and weather. Each is a transparent input, not a black box, so the reasoning behind a forecast can always be traced back to specific factors.