Ron's Puck Tool
Loading…
Loading…
Loading…
Loading…
How Picks Work
▶
- What the price is. The modelled chance he scores at least one goal, shown as American odds with the raw probability beneath. These are fair odds with no vig — a real sportsbook will always price worse. A price assumes he dresses, which is how books settle it: a scratch is void.
- It is calibrated, not a rating. Out of sample across three seasons the number means what it says — the top tenth of the board was priced at 34.8% and scored 34.6% of the time, the bottom tenth 3.8% and 3.7%.
- Poisson, not a coin flip. Goals per skater-game run variance/mean 1.004 — as close to Poisson as real data gets — so the model projects expected goals and the anytime price is 1 − e−expected. That is also what the NFL touchdown board is, one link function over.
- What drives it. His career scoring — goals, shot volume and shooting percentage together — takes the model from a 0.61 ranking to 0.69 on its own. Recent ice time is next, then power-play minutes, recent shots, the opponent's goals allowed, his team's offence and home ice. Final ranking accuracy (AUC) 0.706.
- What is deliberately absent. Recent goals (his last ten games) add nothing once recent shots are in — goals are what shots turn into by luck. Opponent shots allowed and penalty kill were tested and are nulls. A separate even-strength and power-play model, summed, was tested and is worse than one model, because power-play minutes already carry that channel.
- It is recalibrated. A monotone (isotonic) map fitted on out-of-sample results trims the top end, where the raw model runs a little hot. It halves the calibration error without changing the order of the board.
- Only the current roster is priced. A skater's stats are filed under the club he finished last season with, so the board reads each team's roster directly — otherwise anyone who moved this summer would be priced for his old team.
- Seven markets off one model. Pick a market in the filters. 2+ goals, the hat trick and a 1st-period goal come straight from his expected goals — goals are Poisson, so the tails follow (29% of goals are scored in the 1st). First and last goal are a race: exactly one skater scores each, so his price is his share of the expected goals of everyone who dresses in his game — the same anytime number is a better first-goal bet against a weak lineup. The healthy scratches are left out of that race, or every price would run about a fifth long. Power-play goal has its own model: his power-play minutes and power-play scoring, on a log scale, against the opponent's penalty kill. Out of sample every market lands within a third of a point of what happened.
- 🔒 Chalk = every skater, shortest price first. Narrow it with the cut lines in Filters.
- 🎯 Value = Bueno's screen, experimental. Longshots (+400 or longer, anytime) whose form (shots over his last ten, against his position) and the opponent's goals allowed both sit in the top 40% of tonight's slate, sorted by Bueno, the two multiplied: MLB's P. Vuln × Contact × Platoon for hockey (there's no platoon here). Graded in Results on its own until it proves itself, or doesn't.
- Cut lines. Each maximum-price filter shows the hit rate that cut actually produced out of sample — so tightening the board is a measured trade, not a guess.
- Honest limits. Built on 182,608 skater-games across four seasons, never yet graded against a live night, and not measured against sportsbook prices — a fair number is not the same thing as an edge. No starting-goalie or line-combination data yet: those need a lineup feed the free NHL API does not have.
Commas match any of them — narrows whatever the filters already show
Loading…
How Shots Works
▶
- It is a volume market, and volume is what hockey predicts well. Goals are a share market and mostly luck game to game; shots are a role. A skater's shots per game splits half-against-half at 0.82, and his ice time at 0.98 — higher than the snap share that carries the whole NFL touchdown board. That is why this is the market we built first.
- His own shot rate carries it. Career shots per game alone takes the model most of the way — from a league-average guess to −10.8% log loss. Recent ice time is the next largest term, then his last ten games, then power-play minutes.
- Prices are negative binomial, not Poisson. Per player, shot counts run variance/mean 1.13 against the 1.00 a Poisson assumes, so a Poisson pushes too much probability past the line. The wider distribution fixes it — the same reason the NFL receptions board is priced this way.
- Individual Corsi and Fenwick are in the feed, were tested, and are not in the model. Total shot attempts correlate 0.93 with career shots per game and unblocked attempts 0.98. Both are worse than plain shots per game on their own, and adding Corsi back to the finished model makes it worse, not better. They measure the same thing less directly.
- The opponent matters here, which is unusual. Shots allowed per game by the other team is the largest gain after usage — worth more than power-play minutes. On the NFL boards defence has failed this test five times; in a shot-volume market it passes.
- Built on 182,608 skater-games across four seasons, fitted walk-forward — every season predicted by a model that never saw it. Out of sample the ladder is honest to within a point at every line.
- Rates are on a log scale — and that fixed the stars. The first version fed shot rates in raw, which makes a projection grow exponentially with a skater's rate. Out of sample it priced MacKinnon at 5.76 shots a game when he took 4.21, Pastrnak at 4.91 against 3.65. On a log scale the projection is a power law in the rate, and the league's top 25 shooters now come in 2% under their actual volume instead of 5% over. Quoted-line calibration error fell from 1.7 points to 0.6. An isotonic correction on top was tested and made it worse, so there isn't one.
Loading…
How Points Works
▶
- His own scoring rate carries it. Career points per game alone takes the model most of the way — −8% log loss from a position-only guess. Ice time and power-play minutes add the next largest piece, then his last ten games, his team's scoring and the opponent's goals against.
- Assists are the hardest of the three. An assist is half a teammate's event — he can make the pass and the shot misses — so a player's assist rate is noisier than his points rate. The model still separates skaters (AUC .674), just less sharply than points (.692).
- PP points is the sharpest model on the site. A skater is either on a power-play unit or he isn't, and his power-play minutes say which — AUC .825. Its projection is recalibrated at the top, where the raw model ran hot.
- Poisson, and tested. Pooled across players, points look over-dispersed, but that's the mix of players; for one player on one night a Poisson fits. A wider curve was tested at every line, including over 2.5 points, and did no better.
- What was tested and left out. His share of his team's goals ("involvement" — it's points per game by another name), a separate goals-plus-assists build, a centre-versus-winger flag, and shot attempts (Corsi and Fenwick). None added anything.
- Built on 182,053 skater-games across four seasons, fitted walk-forward — every season predicted by a model that never saw it. Out of sample the lines land on what happened: over 0.5 points priced 34.7%, went 34.7%; over 1.5 points 8.9% and 8.9%.
Loading…
How Pairings Work
▶
- What it counts. How often each group of skaters scored on the same night — or, tighter, in the same game — across the season, ranked by how many times. Until this season has two weeks of games, the page runs on last season's goals, and a chip lets you switch once both exist.
- Basis. Every goal, or one of the markets the Picks board prices: the game's first goal, power-play goals, or 2+ goal games. The narrower bases are rarer, so the co-occurrence floor drops to 2.
- The floor steps down as groups grow. A pair has to have scored together 6 nights (3 games) to show; trios 4 (2), quads 3 (2), quints 2 (2). Five skaters scoring on the same night twice is rarer than two doing it six times — one floor for every size left the big groups empty. Early in a season every floor is clamped to the nights played.
- SGP. When 2+ members play for the same club they share a game — a true same-game stack, flagged SGP·TEAM.
- The price, and the correlation. When every member is on tonight's Picks board, the card carries a fair price next to the naive product, struck through. + Parlay drops every leg into your slip, where the same correction applies.
- Teammates barely compete — until the group gets big. Measured over three seasons of goals, two skaters from the same club land at 0.98× the naive product: nearly independent. A team that scores four spreads them around, so a pair doesn't cannibalise the way two NFL receivers do. It compounds as the group grows: a same-club trio lands 11% below, a four-man stack 31% below. Skaters on opposite sides run about 3% below (0.97×).
- Why we trust the number. Two skaters from different games on the same night are independent by construction — that control lands at 1.00, so the gaps above are correlation, not a pricing error. The fit (2023-24 and 2024-25) and the check season (2025-26) agree to within a point for every group size.
- 🎯 Goal → Assist. A different relation: a scorer and the teammates who assisted his goals in the same game, ranked by how many games they connected. A trio is the scorer and both setup men.
- These price the other way. When he scores, someone gets the assist, so “A scores + B gets an assist” runs 19% above the naive product even for teammates who have never connected — and 47% above once B has set A up in 5+ games this season (28%, 31%, 36% at 1, 2 and 3-4 games). Held on the check season to within five points. Early in a season, before the connections build, stacks price at the base 19%.
- What this is not. A fair, no-vig number isn't a claim you beat the book. It tells you what the bet is worth, so you can see how far a book's same-game price sits from it.
- Search. Type names, comma-separated, to require specific skaters in every group.
Loading…
How Hits Works
▶
- His own rate carries it. Career hits per game, then his last five and ten — a hitter is a hitter. Ice time adds the rest of the opportunity: the model works on a log scale, so his rate and his minutes multiply, the way hits-per-60 × ice time would.
- The rink's scorekeeper. Hits and blocks are logged live by off-ice officials at each arena, and some count far more than others — the same play is a hit in one building and nothing in the next. The board scales every skater by tonight's rink, measured on this season's games there only: scorekeepers change, and a rink's reputation goes stale fast (Pittsburgh counted 46% above average in 2022-23 and below average the three years after). A rink with no games yet this season is treated as average.
- Game state. Underdogs expect to trail, and a trailing club chases the puck — so the closing line's win chance is in (the weaker side hits more). The opponent's shot attempts per game are in too: the more they have the puck, the more there is to hit.
- What was tested and left out. How many attempts his own club allows (possession): nothing once his own rate is in — his rate already carries his team's style. Same-division rivalry games: nothing. Penalty-kill time: nothing for hits.
- Negative-binomial prices. Hit counts swing more than a simple average suggests, so the tails are wider than Poisson — tested on every line, it prices them better.
Loading…
How Blocked Shots Works
▶
- Position and role. A defenceman blocks about three times what a forward does, and his own rate — career, last five, last ten — carries most of the rest. That rate is the “willingness” habit, and it's very repeatable.
- The penalty kill. Shots get blocked on the kill, so his penalty-kill minutes are in (t +4.3), next to his ice time.
- The opponent's volume. A club that throws a lot of pucks at the net gives him more to block (t +8.0).
- The rink's scorekeeper. Hits and blocks are logged live by off-ice officials at each arena, and some count far more than others — the same play is a hit in one building and nothing in the next. The board scales every skater by tonight's rink, measured on this season's games there only: scorekeepers change, and a rink's reputation goes stale fast (Pittsburgh counted 46% above average in 2022-23 and below average the three years after). A rink with no games yet this season is treated as average.
- Game state. Underdogs spend more time defending — the closing line's win chance is in.
- What was tested and left out. Home ice; how many attempts his own club allows; the club's blocking system — real and stable (Montreal blocks 27% of attempts against, Carolina 23%) but already in its players' own rates; and the opponent's shot profile. None added anything on top of the rest.
Loading…
How Due Works
▶
- Who's here. Everyone on tonight's Picks board who has gone five or more games without a goal, with the shots he has put on net and the ice time he has played since his last one. A drought that began last season carries over the summer.
- It is not an edge, and I checked. Across three seasons (134,304 skater-games), the goals model was graded against how long each skater's drought had run. Nobody is owed anything. Players 30+ games dry scored 5.2% of the time on a 5.7% price — if anything a long drought runs slightly colder than the price, because it often means a lost role or a quiet injury. Shots since his last goal, ice time since, and the expected goals he “should” have scored each added nothing to the model.
- So why show it? Because it is fun, and because it flags cold players — which is worth knowing too. The default order is tonight's price, which already carries what predicts: his shot volume, his ice time and his power-play role. Sort by games or shots since to see the longest droughts.
- The puck doesn't know he's been waiting. A shooter's odds tomorrow are his shot rate and his role, whether he scored last night or not in a month.
Loading…
How Milestones Works
▶
- What it counts. Regular-season career totals, all the way back, plus this season as it's played. Playoffs don't count toward the round numbers the league celebrates, so they don't count here.
- The rungs. Goals every 50, within 10. Assists and points every 100, within 15. Games every 100, within 8. Goalie wins every 50, within 6. Each is a distance someone covers in a few weeks, so the board moves.
- Active players only. Everyone on a current NHL roster, nobody else.
- Results. Every milestone reached this season, logged the night it happens — a goal milestone carries the goal itself, clip and all, and how long he sat on this list first. Last season's 910 sit underneath.
- It's a watch list. Being one goal from 200 doesn't make him likelier to score tonight. Where a chaser is on tonight's board, his price is shown so you can see what the model thinks independently.
Loading…
How Birthdays Work
▶
- The week. Every player on a current NHL roster whose birthday falls between three days ago and a week from now, with the age he's turning and whether his club plays that day. If he's on tonight's Picks board on his birthday, his goal price is shown.
- His own record. Games he has played on his birthday since 2022-23, and how many he scored in.
- Does it mean anything? No. Graded against the goals model's price over three seasons, birthday boys scored in 57 of 377 games — 15.1% against a 15.3% price. That is as close to nothing as a test gets. It is here because it is fun.
Loading…
How Saves Works
▶
- Saves are a shot-volume market, and the other team supplies the volume. The biggest single input is how many shots his opponent takes, season-long and over its last ten games. Then his own saves per start and how many shots his club lets through. Save % adds nothing to a saves projection — a goalie stops the shots he faces at nearly the same rate as the next one, and the volume is what moves.
- The league keeps getting quieter, so the level floats. Saves per start fell every season we trained on, 27.3 to 24.1. A model fitted on older seasons prices newer ones low, so the projection is the league's saves per start over the last 30 days, times how this goalie, his defence and tonight's opponent compare to the rest of the slate. Opening night, before there are 30 days, uses last season's level.
- The price is two curves, because goalies get pulled. About one start in 18 ends with the goalie pulled, and a pulled goalie keeps about half a game's saves. That gives saves a long left tail; a single curve bends the other way and makes every over near the middle look worse than it is. So the price is a full-game curve plus a small pulled-goalie one. Out of sample that puts every line from 20.5 to 32.5 within half a point of what happened.
- Starters are projected, not confirmed. The NHL doesn't publish starters ahead of time. The board prices the goalie with most of his club's recent starts, shows how many, and Results voids a start he didn't make.
- Built on 10,076 starts across four seasons, fitted walk-forward — every season predicted by a model that never saw it. The busiest 2% of projections still run hot: those are the nights a goalie is most likely to be pulled.
Loading…
How Goals Allowed Works
▶
- Goals allowed is mostly the game total. What a goalie gives up tracks how many goals the game is expected to have far more than anything about him. Without a betting line the best inputs are team goals for and against, home ice and back-to-backs — and together they only just beat the league average.
- So the projection leans on the average, on purpose. Fitted raw, the model's spread ran about twice too wide out of sample: the leakiest-looking spots averaged 3.1 goals against a projected 3.6. Its projections are pulled toward the league mean by the amount that held up on seasons it never saw.
- The price is tighter than a Poisson. Starts run variance/mean 0.82, not 1.00: a goalie who leaks gets pulled before the long tail happens. The price uses a distribution matched to that, and it lands — over 1.5 is priced at 80.0% and went over 79.0%, where a Poisson says 76.5%.
- His save % barely moves it. Over a season, a goalie's save % is mostly noise around the league's; it's shrunk hard, and it adds almost nothing.
- Starters are projected, not confirmed. The board prices the goalie with most of his club's recent starts, shows how many, and Results voids a start he didn't make.
- Built on 10,076 starts across four seasons, fitted walk-forward. The edge here is thin, and the page says so.
Loading…