Team Luck with PDO: the Number That Flags Hot Streaks

Every hockey winter produces a team nobody rated sitting third overall in January, and one word does more work explaining it than any other: PDO — shooting percentage plus save percentage, scaled so the league sits at 100. The folk version says PDO is luck, full stop: anything far from 100 “always regresses.” That version is half right, and this tutorial computes which half. We pull all 32 teams for the two most recent completed seasons from the NHL’s keyless public stats API, pin down why the league averages exactly 100 (it’s arithmetic, not tendency), measure how much of the standings PDO soaks — r = 0.63 in 2025-26 — and then run the only test that matters: does it carry over? Year over year, PDO persists at 0.26… and splits into a shooting half that keeps real talent (0.37) and a save half that keeps almost none (0.06). The half everyone credits to goaltending is the half that vanishes.

You have the machinery already: the NHL API basics, correlation, and regression to the mean, which this is hockey’s favorite special case of. The fetch step needs a live connection; everything after it runs from the bundled nhl_team_pdo_two_seasons.csv the script saves, so the analysis reproduces offline, byte for byte.

1. Pull two complete seasons of team summaries

   The stats endpoint takes a cayenneExp filter — season and game type — and returns one row per team: goals for and against, shots for and against per game, points. Two calls, 64 rows, no API key. One honesty detail: the Utah franchise renamed between these seasons (Hockey Club → Mammoth), so we map the names to keep it one franchise — otherwise every year-over-year join silently drops a team.

   import pandas as pd
   import requests
   
   SEASONS = {"2024-25": 20242025, "2025-26": 20252026}
   RENAME = {"Utah Hockey Club": "Utah Mammoth"}
   
   rows = []
   for label, sid in SEASONS.items():
       url = ("https://api.nhle.com/stats/rest/en/team/summary"
              f"?cayenneExp=seasonId={sid}%20and%20gameTypeId=2")
       for t in requests.get(url, timeout=30).json()["data"]:
           gp = t["gamesPlayed"]
           rows.append({"team": RENAME.get(t["teamFullName"], t["teamFullName"]),
                        "season": label, "gp": gp, "points": t["points"],
                        "gf": t["goalsFor"], "ga": t["goalsAgainst"],
                        "shots_for": round(t["shotsForPerGame"] * gp),
                        "shots_against": round(t["shotsAgainstPerGame"] * gp)})
   
   df = pd.DataFrame(rows)
   df.to_csv("nhl_team_pdo_two_seasons.csv", index=False)

   Two seasons, 32 franchises

   64 team-seasons from live pull: 2 seasons x 32 franchises
   (Utah's rename is mapped: Hockey Club and Mammoth are one franchise.)
   
                      team  points   gf  shots_for     sh_pct     sv_pct         pdo
   38         Dallas Stars     112  273       2074  13.162970  89.650350  102.813320
   58       Buffalo Sabres     109  283       2306  12.272333  89.928661  102.200994
   44        Boston Bruins     100  268       2216  12.093863  89.856263  101.950126
   48  Tampa Bay Lightning     106  286       2305  12.407809  89.538602  101.946411
   50   Montréal Canadiens     106  279       2156  12.940631  89.000876  101.941507

2. Compute PDO — and see why the league averages exactly 100

   Team shooting percentage is goals for over shots for; team save percentage is one minus goals against over shots against; PDO is their sum, in percentage points. Before looking at any team, check the league totals — because every goal scored is somebody else’s goal allowed and every shot taken is somebody else’s shot faced, the league’s SH% and SV% are the same fraction viewed from opposite benches. They must sum to 100. PDO doesn’t hover near 100 out of cosmic fairness; it averages 100 by construction, which is exactly what makes deviations from it readable.

   import pandas as pd
   
   df = pd.read_csv("nhl_team_pdo_two_seasons.csv")
   df["sh_pct"] = 100 * df["gf"] / df["shots_for"]
   df["sv_pct"] = 100 * (1 - df["ga"] / df["shots_against"])
   df["pdo"] = df["sh_pct"] + df["sv_pct"]
   
   for label, d in df.groupby("season"):
       lg_sh = 100 * d["gf"].sum() / d["shots_for"].sum()
       lg_sv = 100 * (1 - d["ga"].sum() / d["shots_against"].sum())
       print(label, round(lg_sh, 2), "+", round(lg_sv, 2), "=", round(lg_sh + lg_sv, 2))

   The league identity: SH% + SV% = 100, both seasons

   2024-25: league SH% 10.65 + league SV% 89.35 = 100.00
   2025-26: league SH% 11.07 + league SV% 88.93 = 100.00
   
   Every goal scored is a goal allowed and every shot taken is a shot faced,
   so the league's SH% and SV% are the same fraction seen from both benches.
   PDO does not hover near 100 by tendency - it averages 100 by arithmetic.

   A scope note before the leaderboards: classic PDO is a 5-on-5, on-ice stat. What we’re computing is the all-situations, team-season version — the same idea with power plays and empty-netters left in, because that’s what the summary endpoint carries. It runs a touch wider than 5v5 PDO, and everything below says so when it matters.

3. Read a season’s leaders and laggards

   Sort 2025-26 by PDO and look at both tails next to the points column. The whole league lives inside about six points of PDO — a standard deviation of 1.4 — and yet the correlation between PDO and standings points is 0.63. Meaning: a big share of what the standings call “good” in any single season is teams sitting on the friendly end of two percentages.

   cur = df[df.season == "2025-26"].set_index("team")
   
   print(cur.nlargest(5, "pdo")[["pdo", "sh_pct", "sv_pct", "points"]].round(2))
   print(cur.nsmallest(5, "pdo")[["pdo", "sh_pct", "sv_pct", "points"]].round(2))
   print("r(PDO, points):", cur["pdo"].corr(cur["points"]).round(2))

   2025-26 PDO leaders, laggards, and the same-season correlation

   2025-26 PDO leaders and laggards (all situations, team level):
   
     Dallas Stars             PDO  102.8   SH% 13.16  SV% 89.65  pts 112
     Buffalo Sabres           PDO  102.2   SH% 12.27  SV% 89.93  pts 109
     Boston Bruins            PDO  102.0   SH% 12.09  SV% 89.86  pts 100
     Tampa Bay Lightning      PDO  101.9   SH% 12.41  SV% 89.54  pts 106
     Montréal Canadiens       PDO  101.9   SH% 12.94  SV% 89.00  pts 106
     Vancouver Canucks        PDO   97.0   SH%  9.86  SV% 87.16  pts  58
     New Jersey Devils        PDO   98.0   SH%  9.32  SV% 88.73  pts  87
     Anaheim Ducks            PDO   98.1   SH% 10.49  SV% 87.62  pts  92
     Florida Panthers         PDO   98.2   SH% 10.72  SV% 87.53  pts  84
     Calgary Flames           PDO   98.6   SH%  9.02  SV% 89.56  pts  77
   
   spread: mean 100.0, sd 1.36, range 97.0 to 102.8
   same-season r(PDO, points) = 0.63 - the standings are soaked in it

   Notice how teams get to the same PDO: Montréal’s 101.9 is shooting-heavy (12.94, with an ordinary save percentage), Buffalo’s 102.2 leans on the netminding. The sum treats those identically. The decomposition in the next step is why you shouldn’t.

4. The test that settles it: does PDO carry over?

   “Luck” has a measurable signature: it doesn’t repeat. Join the two seasons on franchise and correlate everything with its own next-year value. If the folk claim were fully right, PDO’s year-over-year r would sit near zero. It doesn’t — and the split is the real lesson.

   cur = df[df.season == "2025-26"].set_index("team")
   prev = df[df.season == "2024-25"].set_index("team")
   both = cur.join(prev, lsuffix="_26", rsuffix="_25", how="inner")
   
   for a, b, label in [("pdo_25", "pdo_26", "PDO"),
                       ("points_25", "points_26", "points"),
                       ("sh_pct_25", "sh_pct_26", "shooting half"),
                       ("sv_pct_25", "sv_pct_26", "save half")]:
       print(label, both[a].corr(both[b]).round(2))

   Year-over-year persistence, 2024-25 to 2025-26

   Year-over-year persistence, 32 franchises, 2024-25 -> 2025-26:
   
     r(PDO                 ) =  0.26
     r(points              ) =  0.22
     r(SH% (shooting half) ) =  0.37
     r(SV% (save half)     ) =  0.06
   
   The half of PDO everyone credits to goaltending is the half that vanishes.

   PDO persists at 0.26 — not zero, because rosters carry real finishing talent, and the shooting half shows it at 0.37. But the save half — the part broadcast crews narrate as a goalie’s ability — correlates with itself at 0.06 across seasons. One season of team save percentage is close to uninformative about the next. That is regression to the mean wearing a goalie mask, and it’s why analysts treat a save-percentage-driven standings run as borrowed points.

   

5. Use it the way it deserves: as a watchlist

   The payoff table. Take 2024-25’s five highest-PDO teams and five lowest, and check their point totals a season later. The high five include the season’s two loudest collapses — and the low five improved five for five, by an average of 21 points.

   top5 = both.nlargest(5, "pdo_25")
   bot5 = both.nsmallest(5, "pdo_25")
   
   print(top5[["pdo_25", "points_25", "points_26"]].round(1))
   print(bot5[["pdo_25", "points_25", "points_26"]].round(1))

   2024-25's PDO extremes, one season later

   2024-25's five highest-PDO teams, one season later:
   
     Winnipeg Jets            PDO  103.4 -> pts 116 ->  82  (-34)
     Tampa Bay Lightning      PDO  103.1 -> pts 102 -> 106  (+4)
     Dallas Stars             PDO  102.5 -> pts 106 -> 112  (+6)
     Washington Capitals      PDO  102.4 -> pts 111 ->  95  (-16)
     Toronto Maple Leafs      PDO  102.1 -> pts 108 ->  78  (-30)
   
   2024-25's five lowest-PDO teams, one season later:
   
     Nashville Predators      PDO   97.1 -> pts  68 ->  86  (+18)
     San Jose Sharks          PDO   97.5 -> pts  52 ->  86  (+34)
     Philadelphia Flyers      PDO   97.7 -> pts  76 ->  98  (+22)
     New York Islanders       PDO   98.3 -> pts  82 ->  91  (+9)
     Boston Bruins            PDO   98.7 -> pts  76 -> 100  (+24)
   
   scoreboard: 3 of 5 high-PDO teams fell, 5 of 5 low-PDO teams improved.

   Read it with both eyes open. Winnipeg’s 103.4 rode the league’s best goaltending to 116 points and gave back 34 of them; Toronto gave back 30. But Dallas and Tampa held — high-PDO teams that were also simply good. That’s the correct final shape of the idea: PDO is not a verdict, it’s a flag. An extreme value tells you which teams’ records are running ahead of (or behind) their underlying play, and the save-percentage half tells you which flavor of extreme is least likely to survive the summer. Luck isn’t the absence of skill; it’s the part of the skill-shaped number that doesn’t come back.