Build NHL Combos With Data-Driven Analytics

NHL combos come from one goal distribution per game, built around the confirmed starting goaltender. The builder reads the moneyline, puck line and total from the public feed, then adds shot share, special teams, rest and back-to-back context and recent form. Legs — moneyline, puck line and game total — are priced from that shared distribution.

Hockey is low-scoring, which makes correlation sharp: a favourite backed by a hot goalie tends to win close, low-scoring games, so the moneyline and the under share a script; the puck line at −1.5 usually needs an empty-net goal, which only comes in games the favourite is already winning late. Saved tickets are tracked and graded from the official final, including overtime and the shootout for moneyline legs. Goalie confirmations often land close to puck drop, and the model works from what the feed has reported.

A home favourite at −160 with the total at 6.0 and a confirmed starting goalie who has stopped everything for a month. “Favourite wins” and “under 6” are strongly positively correlated: the goalie is the reason for both. The naive product prices the pair as if unrelated and undervalues it.

Now add the puck line instead of the moneyline. Winning by two often requires an empty-net goal in the last two minutes — which adds a goal to the total. Puck line −1.5 and under 6 pull against each other, and the joint probability falls below the naive number. The builder prices every combination from a single goal distribution for this game, so it sees both relationships. A late goalie change moves the whole model; the game page shows the latest confirmation the feed has.

The general version of this argument — why one score model beats multiplied legs — is on the combos overview, and the rules used to grade every saved ticket are on the pick tracker.