Expected vs Actual: Find the Hitters Getting Unlucky

Here's a hard truth that separates good analysts from box-score readers: a hitter can do everything right and still come up empty. Scald a line drive straight at the shortstop and the scoreboard says "out" - identical to a weak pop-up. Over a week or a month those bad bounces pile up, and the batting average lies about how well someone actually hit. We're going to catch that lie in the act — comparing what each hitter's contact deserved against what it actually produced, using Statcast's expected-stats model. The gap between the two is a clean, defensible measure of luck.

This is an advanced step, and it builds straight on the Statcast exit-velocity leaderboard - same week of data, same groupby-and-agg backbone, same player-ID lookup. If you've run that one, this week is already cached and the pull is instant. The data is Baseball Savant via pybaseball, retrieved June 2026.

1. The two numbers, and why they're the right two

   We compare two flavors of wOBA (weighted On-Base Average), a one-number hitting metric scaled so that roughly .320 is league average and .400 is excellent. Actual wOBA is what happened - the credit a hitter earned from his real outcomes. Expected wOBA (xwOBA) is what Statcast's model says those same batted balls should have produced, based only on each ball's exit velocity and launch angle, ignoring where the fielders happened to be standing.

   The logic is simple and powerful: launch speed and angle are things the hitter controls; defensive positioning and lucky bounces are not. So if a hitter's actual wOBA sits far below his expected wOBA, he hit the ball well but the results didn't follow - he's been unlucky and is a good bet to improve. If actual sits far above expected, he's been getting away with weak contact. The gap is the luck.

2. Pull the cached week and keep batted balls

   We reuse the exact week from the leaderboard tutorial - the first week of June 2023 - so nothing new downloads. Then we keep only balls in play that have both an expected and an actual wOBA recorded.

   import matplotlib.pyplot as plt
   import pybaseball as pyb
   
   pyb.cache.enable()
   
   data = pyb.statcast("2023-06-01", "2023-06-07")
   bb = data[(data["type"] == "X")].dropna(
       subset=["estimated_woba_using_speedangle", "woba_value"]).copy()

   The column estimated_woba_using_speedangle is Statcast's xwOBA for each batted ball - the model's verdict on what that contact deserved - and woba_value is the actual wOBA credit the play earned. We filter to type == "X" (in play) and drop rows missing either value, because we can't compute a gap without both. As always, .copy() on the slice avoids the SettingWithCopyWarning when we add a column next.

3. Aggregate to one row per hitter

   This is the familiar leaderboard move, now computing two averages side by side. We group by batter, count each hitter's batted balls, and take the mean of both the actual and expected wOBA. Then we require a real sample so a single ball can't dominate.

   MIN_BBE = 12
   board = (bb.groupby("batter")
            .agg(bbe=("woba_value", "size"),
                 actual=("woba_value", "mean"),
                 expected=("estimated_woba_using_speedangle", "mean"))
            .query("bbe >= @MIN_BBE"))

   The named aggregations build all three columns in one pass: bbe counts the batted balls, actual and expected average the two wOBA flavors. The @MIN_BBE inside query pulls in our Python variable so the threshold lives in one place - and over a single week, twelve batted balls is already a generous bar; lower it and the list fills with hitters whose averages swing wildly on one or two plays.

4. Compute the luck gap

   The whole tutorial comes down to one subtraction. Luck is actual minus expected: positive means a hitter out-performed his contact (fortunate), negative means his results trailed his contact (unlucky).

   board["luck"] = (board["actual"] - board["expected"]).round(3)
   board = board.round({"actual": 3, "expected": 3})

   We round everything to three decimals because wOBA lives on a .000-to-1.000ish scale where the third place is meaningful. That single actual - expected column is the entire signal we set out to build; everything after this is turning IDs into names and drawing the picture.

5. Turn IDs into names and find the extremes

   Statcast gives us batter IDs, so we reverse-look-up the whole index at once and merge the names back on. Then we sort by the absolute size of the luck gap to surface the most extreme stories in either direction.

   names = pyb.playerid_reverse_lookup(board.index.tolist(), key_type="mlbam")
   names["player"] = names["name_first"].str.title() + " " + names["name_last"].str.title()
   board = board.merge(names.set_index("key_mlbam")["player"], left_index=True, right_index=True)
   
   extreme = board.reindex(board["luck"].abs().sort_values(ascending=False).index).head(8)
   print("Biggest gaps between results and contact quality (June 1-7, 2023):")
   print(extreme[["player", "bbe", "expected", "actual", "luck"]].to_string(index=False))

   The biggest luck gaps that week

   Biggest gaps between results and contact quality (June 1-7, 2023):
          player  bbe  expected  actual   luck
       Ryan Noda   13     0.413   0.762  0.349
    Bryce Harper   17     0.588   0.306 -0.283
   Eddie Rosario   16     0.396   0.666  0.269
     Nolan Jones   12     0.462   0.721  0.259
      Bobby Witt   21      0.54     0.3  -0.24
     Pavin Smith   17     0.398   0.171 -0.228
    Ha-Seong Kim   12     0.289   0.508  0.219
   Jake Mccarthy   14     0.343   0.129 -0.215

   The pattern to learn is board["luck"].abs().sort_values(ascending=False).index: we sort by the magnitude of luck, grab that ordering as an index, and reindex the board onto it so the most extreme hitters - lucky or unlucky - rise to the top together. The results read true: Ryan Noda's .762 actual wOBA towers over his .413 expected for a +0.349 gap - he got hits on contact that usually doesn't reward you. At the other pole, Bryce Harper crushed the ball (a .588 expected wOBA, elite contact) but managed only a .306 actual - a -0.283 gap, the unluckiest line on the board and a near-certain bet to bounce back.

6. Plot expected against actual

   The right chart for "predicted vs realized" is a scatter with a diagonal reference line where the two are equal. Every hitter is a dot; the dashed line is expected = actual. Points above the line out-hit their contact; points below it are the unlucky ones whose results trailed what they earned.

   fig, ax = plt.subplots(figsize=(7.8, 7))
   ax.scatter(board["expected"], board["actual"], s=42, color="#B23A3A",
              edgecolor="#FBF7EE", linewidth=0.4, zorder=3)
   lo, hi = 0.1, board[["expected", "actual"]].values.max() + 0.05
   ax.plot([lo, hi], [lo, hi], color="#6C7079", linestyle="--", linewidth=1,
           label="expected = actual")
   for _, r in board.reindex(board["luck"].abs().sort_values(ascending=False).index).head(6).iterrows():
       ax.annotate(r["player"].split()[-1], (r["expected"], r["actual"]),
                   fontsize=8, xytext=(4, 3), textcoords="offset points")
   ax.set_xlabel("expected wOBA (what the contact deserved)")
   ax.set_ylabel("actual wOBA (what they got)")
   ax.set_title("Who's out-hitting - or trailing - their contact?")
   ax.legend(frameon=False, fontsize=9)
   fig.savefig("xwoba_vs_woba.png", dpi=144, bbox_inches="tight")

   

   The diagonal is the trick that makes this readable. Drawing the line from lo to hi - the same range on both axes - means any dot's vertical distance from the line is its luck, no arithmetic required. We label only the six most extreme hitters with annotate, using r["player"].split()[-1] to print just the last name and an xytext offset to nudge it clear of the dot. Harper sits well below the line; Noda well above. Setting zorder=3 keeps the dots drawn on top of the reference line.

7. Read it like an analyst

   The instinct to fight is treating one week's gap as a verdict on a hitter's talent - it isn't. Over seven days even an excellent MIN_BBE sample is small, and luck is, by definition, noisy. What the chart genuinely earns you is a shortlist: the hitters whose underlying contact and surface results have pulled apart, which is exactly where reversion tends to happen next. A pro would read Harper's dot as "buy" and Noda's as "don't expect this to last," then widen the window to a month before betting anything real on it. That discipline - trusting the process metric over the small-sample result - is the whole reason expected stats exist.

Troubleshooting