The short version
The WCH Power Index rates every Division I team by how many goals better or worse than an average D-I team it would be on neutral ice. A rating of +2.00 means a team would be expected to beat an average opponent by about two goals. Teams are ranked best to worst by that number. It is recalculated every Monday morning after the weekend's games, and every past week is kept in the archive.
It is our own model. It does not use any poll or the NCAA's selection index as an input. We use those only to check our work.
How the rating is built
- Every goal for and against, adjusted for the opponent. Each team gets an offense number (goals it scores against an average defense) and a defense number (goals it keeps out against an average offense). The rating is offense plus defense. All teams are solved at once, so beating a strong team counts for more than beating a weak one.
- Home ice. The model measures home-ice advantage from the data (historically about 0.15–0.30 goals a game; this season’s estimate so far is 0.11) and removes it, and it treats neutral-site games as neutral.
- Blowouts are capped. Winning margins count up to 6 goals, so running up the score on a much weaker team does not inflate a rating.
- Shots matter too. Goals are noisy. 30% of each game's input comes from shots on goal (converted to goals at the national shooting rate), which tracks which teams control play and smooths out hot and cold goaltending.
- A starting point from last season. Before a team has played much, its rating leans on where it finished last year, adjusted for how much of its scoring and goaltending came back. That starting point carries the weight of about 2 games, so it fades quickly once real results arrive. Each team's "Why" line shows how much of the rating still comes from last season.
How the forecast works
- We play out the rest of the schedule 10,000 times. Each game's score is drawn from the two teams' offense and defense numbers plus home ice, including overtime and ties.
- Every simulation also starts from a slightly different version of each team's rating, within the model's margin of error. That keeps the forecast honest about what we don't know yet, especially early in the season.
- After each simulated season we recompute the WCH Power Index. That gives the projected final rank (the median), its range (the middle 80% of simulations), the projected record and the odds of finishing in the top 4 or top 8.
- Regular-season conference titles use the conferences' 3-2-1-0 points. Ties after overtime go to a 50/50 shootout, and tiebreakers are random.
- NCAA tournament odds are approximate. We simulate each conference tournament in its 2025-26 format and give the five champions automatic bids. We then recompute the NCAA Percentage Index (NPI) with the NCAA's published 2025-26 settings: 25% win percentage, 75% strength of schedule, the quality-win bonus, 2/3 credit for an overtime win and the 12-win minimum. The six highest-ranked remaining teams fill the at-large spots. Our NPI matches published NPI tables to within a few tenths of a point. We don't model the committee's discretion, such as an unavailable star player, and we assume the 2025-26 formats and settings still apply.
Does it work? Backtest results
We replayed past seasons week by week. Each Monday we used only the games played before that date to predict that week's games, then scored the predictions. Settings were tuned on 2018-19, 2019-20, 2021-22, 2022-23 and tested on 2023-24, 2024-25, 2025-26, which the model never saw during tuning (2,408 games). Lower log loss is better. Accuracy counts games that had a winner.
| Model | Log loss | Picked the winner |
|---|---|---|
| WCH Power Index | 0.673 | 75.7% |
| Elo (margin, home ice, carry-over) | 0.706 | 74.8% |
| Win % (in-season) | 0.783 | 68.8% |
| Home/away base rates | 0.881 | 55.9% |
| USCHO Poll (points) | 0.783 | 70.3% |
All models are scored on the same 2,408 games. The USCHO Poll baseline predicts from the latest poll's points, including the preseason poll, with unranked teams counted as zero. Elo is a standard rating that moves after every game, with margin of victory, home ice and a carry-over between seasons. Win % uses only this season's record.
Is the forecast calibrated?
We also ran the forecast from December 1 and January 15 in each season from 2021-22 to 2025-26, with every later result hidden, and compared it with what actually happened.
- Final win totals landed inside the forecast's 10th–90th percentile range 90% of the time (432 team forecasts). The target is 80%, a little more because the range includes its end points.
- Final WCH Power Index ranks landed inside the 10th–90th percentile range 93% of the time. That's a bit wide, so the ranges are on the cautious side.
- Regular-season conference title odds had a Brier score of 0.051 (0 is perfect): teams given 0.00–0.05 won 1% (avg. forecast 0%, n=311); teams given 0.05–0.25 won 18% (avg. forecast 14%, n=52); teams given 0.25–0.50 won 32% (avg. forecast 35%, n=28); teams given 0.50–0.75 won 39% (avg. forecast 64%, n=19); teams given 0.75–0.95 won 92% (avg. forecast 86%, n=12); teams given 0.95–1.00 won 100% (avg. forecast 99%, n=10).
- NCAA field odds (2025-26, the first NPI season) had a Brier score of 0.044.
What data goes in, and what each piece added
| Data | How it's used | Effect in testing |
|---|---|---|
| Game results, 2017-18 to now | Every D-I vs D-I score, with overtime and ties, solved together for each team's offense and defense (margins capped at 6 goals). | The core of the model. On 2023-26 games, log loss 0.674 and 75.7% winners picked, against 0.783 and 68.8% for a win-percentage model. |
| Home, away and neutral sites | Home-ice edge is estimated from the data, and neutral-site games are treated as neutral. | Home ice has been worth about 0.15-0.30 goals a game since 2017. Part of the core model. |
| Last season's final ratings | Each team's starting point, carrying the weight of about 2 games. | Lets the model make sensible picks from opening weekend. Tuned on 2018-23. |
| Shots on goal (box scores, 2022-23 on) | 30% of each game's input is shots converted to goals at the national shooting rate (8.6%). This tempers hot and cold goaltending. | Log loss 0.6853 to 0.6757, winners picked 75.0% to 75.4% (2023-26). |
| Returning scoring (box scores + current rosters) | Starting point moves with the share of last season's points by players still on the team. | First six weeks: log loss 0.628 to 0.619; full season 0.6903 to 0.6883 (2024-26). Returning goalie minutes added nothing, so they're left out. |
| Faceoffs, power play/penalty kill, penalty minutes | Tested as extra inputs on top of goals and shots. | No improvement (the best weight was zero or made test results worse), so not used. |
| Rest days and travel | Tested back-to-back games and cross-conference road trips. | Only 37 of about 5,000 games had a rest mismatch, and the travel effect was not significant (+0.13 ± 0.08 goals). Not used. |
| Recruiting commitments | Not used. | We only have the current class, so there's no history to test whether it helps. |
| USCHO Poll history (2017 on) | Comparison only, never an input. | Picked 70.3% of winners (log loss 0.783) on the same games. |
| NCAA Percentage Index (NPI) | Recomputed from results for the NCAA-field odds only. | Matches published 2026-27 NPI values to within 0.2 points on average. |
Game Preview and the Expert
- Win chances, not betting odds. Each rated upcoming game's preview uses the latest Power Index ratings, home ice and neutral-site flags. Expected goals feed independent Poisson score distributions; the model then accounts for overtime and the probability of a tie after OT.
- The Expert has no discretion. It picks the team with the higher modeled win probability. An edge smaller than 2 percentage points is marked Pick’em / no play.
- Picks stay locked. The forward-only W-L-T record starts on Nov. 6, 2026. Picks written before the scheduled puck-drop time are saved permanently; later rating changes never revise them. Pick’em no-plays and ties are excluded from win percentage.
- Its public Expert page lists upcoming locks and settled results. Games lacking a unique schedule match or ratings are omitted rather than guessed.
Weekly updates
Ratings and the forecast are rebuilt every Monday morning (Eastern) from all completed games. The current edition is the week of Nov. 16, 2026. Pick any earlier week from the menu on the Power Index or Forecast page. Exhibitions and games against non-D-I opponents are left out.
The WCH Power Index is an independent statistical model by womenscollege.hockey. It is not affiliated with the NCAA, any conference or any poll.