The NHL’s stats are everywhere.
The analyst isn’t.
So RinkIQ does two things at once. Its own models do the maths an Elo rating for every club, a homegrown expected‑goals model that grades shot quality straight from the play‑by‑play, and a blend of machine learners that turns it all into a win probability and a summarizer does the talking, turning the result into a read a smart fan can follow, so you know which edge holds up over a season and which is just a hot goalie you shouldn't trust. It keeps honest score, too about 57.7%, tracked out‑of‑sample and built on our own features with no third‑party stats, a hair under the simulation‑driven heavyweights that run thousands of seasons. But the number was never the point; the point is never being handed a figure without being told what it means
Not a black box that prints a number and shrugs. Every pick sits next to the measurable inputs behind it — the ones you can open up and check for yourself.
Every read is anchored to real, current numbers — never an invented stat, never a guessed injury. And a good analyst tells you their confidence, not just their pick: RinkIQ says plainly when a game is a coin flip, and treats its rating as a rough proxy, not a prophecy.The read shows up where it matters — on tonight’s games, on any matchup you open, and on the news around it. It calls what the numbers mean on the ice, flags which edge holds up and which is a small-sample mirage, and names the factors that drove the pick — all anchored to the data on the page, so it explains rather than invents.
The pick isn't a single formula. An Elo rating that updates game to game and a rolling read on recent form feed a stack of learners, gradient‑boosted trees and regressors that read the goal‑ and expected‑goals difference weighted by how each has actually performed out‑of‑sample. The expected‑goals model underneath is ours too, learned from shot locations, angles and types in the NHL play‑by‑play, then calibrated so a 0.10 chance really scores about one time in ten. Every model's method, track record and calibration opens up on the accuracy page current‑season signal only, no team‑name shortcuts to lean on.
Player roles, a goalie’s workload weighed against the save percentage he actually posts, and expected-goals shot quality all feed the matchup — and so do injuries, weighed by how much a player actually plays, so a top-pairing defenceman or a starting goalie counts for far more than a depth call-up.
Pick any two teams and get an overall comparison, not just a game preview: Elo and league rank, offense and defense, special teams and form, plus side-by-side shot-location maps showing where each team creates and allows its chances — with an AI verdict that reads it all and calls the edge, honestly.
Where it came from
It began with a question that wouldn’t let go after reading Nate Silver’s The Signal and the Noise: how far could feature engineering get on its own, without a simulation engine to fall back on? The only way to answer it was to keep asking why the model landed where it did — which is how a lone prediction script sprouted a set of dashboards, and then the summarizers that let the predictions and the analytics finally talk to each other.
The friend I showed it to didn’t just nod along — he pulled up a chair and built out the deployment side so the thing could actually ship. An evening side-project ended up with two names on it, a mountain of CSVs behind it, and a real debt to the book that started the whole chase.
Everything is open to everyone through mid-October, no signup gymnastics. Once the season is underway a token $3/month kicks in — not a business plan, just enough to cover the API bills and keep the servers running.