Build a QB Efficiency Comparison Chart

Comparing quarterbacks by passing yards or touchdowns is a trap - those stats reward volume and lucky bounces as much as skill. The analytics community settled on something better: a two-number portrait of every starting quarterback that separates how much value he adds from how accurate he is. Building it means aggregating a full season of play-by-play down to the quarterback level, requiring a fair sample of attempts, and finishing with a labeled scatter where the best passers land in the upper-right corner.

This is an intermediate step up: we're moving from analyzing plays to analyzing people, which means grouping many rows into one row per player. It builds on EPA explained by building it - we'll use a quarterback-specific flavor of EPA here - so make sure that idea is comfortable before you start. As always we read nflverse's data through the sdt_nflverse helper rather than the broken nfl_data_py library.

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

   We'll plot every qualifying passer on two axes. The vertical axis is EPA per dropback: the average Expected Points Added each time the quarterback drops back to pass. It answers "how much is this player worth on a typical passing play?" and it captures everything - completions, the costly sacks and interceptions, the explosive gains. It's the closest thing we have to a single value metric for a passer.

   The horizontal axis is CPOE, Completion Percentage Over Expected. nflverse models how likely each individual throw is to be completed given its difficulty - depth, location, pressure - and CPOE is how much better (or worse) the quarterback did than that expectation, in percentage points. A CPOE of +4 means he completed passes at a rate four points above what an average passer would on those exact throws. It isolates accuracy from the difficulty of what he was asked to do. Value and accuracy are related but not identical, and plotting one against the other reveals quarterbacks who are efficient through different means.

2. Load the dropback-level columns

   We only need five columns for the whole 2023 season: who threw the ball, his team, his per-dropback EPA, his CPOE, and a flag marking pass attempts. Pulling just these keeps the read fast.

   import matplotlib.pyplot as plt
   
   import sdt_common as sdt
   import sdt_nflverse as nfl
   
   sdt.init("build-a-qb-efficiency-comparison-chart")
   
   pbp = nfl.import_pbp_data([2023], columns=[
       "passer_player_name", "posteam", "qb_epa", "cpoe", "pass_attempt"])

   Notice we ask for qb_epa, not the plain epa from the last tutorial. They're closely related, but qb_epa is credited to the quarterback's account on dropbacks - the right column when the unit of analysis is the passer rather than the offense.

3. Aggregate to one row per quarterback

   This is the central move. We keep only pass attempts, group by passer and team, and compute three summaries at once: the count of attempts, the mean EPA per dropback, and the mean CPOE. Then we require a minimum sample so backups and emergency fill-ins don't pollute the chart.

   # One row per dropback; group to the quarterback and require a real sample.
   MIN_ATT = 250
   qbs = (pbp[pbp["pass_attempt"] == 1]
          .groupby(["passer_player_name", "posteam"])
          .agg(att=("pass_attempt", "sum"),
               epa=("qb_epa", "mean"),
               cpoe=("cpoe", "mean"))
          .query("att >= @MIN_ATT")
          .reset_index())

   The named-aggregation syntax - epa=("qb_epa", "mean") - reads as "make a column called epa by taking the mean of qb_epa," and it lets us build all three columns in a single pass. The @MIN_ATT inside query pulls in our Python variable, so the threshold lives in exactly one place. Grouping by both name and team is deliberate: it keeps a midseason-traded quarterback's stints separate, which is more honest than blending two different supporting casts.

4. Why a minimum attempts threshold matters

   Without MIN_ATT, the leaderboard would be topped by someone who threw twelve passes in garbage time and happened to complete most of them. Efficiency metrics are rate stats, and rates are wildly noisy in small samples - a single explosive play can swing a 20-attempt EPA average more than a great game swings a 600-attempt one. Setting the bar at 250 attempts (a little under half a season of starts) keeps us to genuine starters whose numbers have stabilized. It's the same discipline as the minimum batted-balls rule in the Statcast exit-velocity leaderboard: never let a tiny sample crash the party.

5. Read the leaderboard

   Let's sort by EPA per dropback and look at the most valuable passers of 2023. We label the rows 1 through 8 and round for readability.

   with sdt.snippet("leaders"):
       top = qbs.sort_values("epa", ascending=False).head(8).round(3)
       top.index = range(1, len(top) + 1)
       print(f"Most efficient passers (min {MIN_ATT} attempts, 2023):")
       print(top[["passer_player_name", "posteam", "att", "epa", "cpoe"]].to_string())

   The most efficient passers of 2023

   Most efficient passers (min 250 attempts, 2023):
     passer_player_name posteam    att    epa   cpoe
   1            B.Purdy      SF  584.0  0.280  3.947
   2       T.Tagovailoa     MIA  632.0  0.170  3.150
   3         D.Prescott     DAL  695.0  0.169  3.845
   4             J.Love      GB  667.0  0.168  1.402
   5            J.Allen     BUF  678.0  0.151  4.466
   6             J.Goff     DET  758.0  0.142  1.349
   7         M.Stafford      LA  592.0  0.123 -1.946
   8           C.Stroud     HOU  595.0  0.120  0.102

   This passes the smell test, which is how you know the pipeline is sound. Brock Purdy tops the list at +0.280 EPA per dropback with an excellent +3.947 CPOE - both accurate and enormously valuable. Tua Tagovailoa and Dak Prescott follow, then Jordan Love, Josh Allen, and Jared Goff. The numbers are real season aggregates, retrieved June 2026.

   The most instructive row is Matthew Stafford at #7: a strong +0.123 EPA per dropback despite a negative CPOE of -1.946. That combination tells a story the single number would hide - he added value while completing passes below expectation, which usually means aggressive, high-value downfield throws rather than easy checkdowns. Two quarterbacks can reach similar efficiency by opposite routes, and the two-axis view is what lets you see it.

6. Build the labeled scatter

   Now the payoff chart. Each quarterback becomes a dot at his (CPOE, EPA) coordinates. We add dashed reference lines at the mean of each axis so the plot splits into four quadrants, and we label every dot with the passer's name.

   fig, ax = plt.subplots(figsize=(8.2, 6.2))
   ax.scatter(qbs["cpoe"], qbs["epa"], s=46, color=sdt.sport_color("football"),
              edgecolor="#FBF7EE", zorder=3)
   ax.axhline(qbs["epa"].mean(), color=sdt.SPORT_COLORS["foundations"], linestyle="--", linewidth=1)
   ax.axvline(qbs["cpoe"].mean(), color=sdt.SPORT_COLORS["foundations"], linestyle="--", linewidth=1)
   for _, r in qbs.iterrows():
       ax.annotate(r["passer_player_name"], (r["cpoe"], r["epa"]),
                   fontsize=7.5, xytext=(4, 3), textcoords="offset points", color="#20242B")
   ax.set_xlabel("CPOE  (completion % over expected)")
   ax.set_ylabel("EPA per dropback")
   ax.set_title("QB efficiency, 2023: accuracy vs value added")
   sdt.save_fig(fig, "qb_efficiency", source="nflverse via nfl_data_py")

   

   The two design choices that make this readable: the annotate call with an xytext offset nudges each name a few points up and to the right of its dot so the labels don't sit on top of the markers, and the dashed mean lines turn raw coordinates into meaning. The upper-right quadrant - above-average value and above-average accuracy - is where the best quarterbacks live; the lower-left is where struggling starters cluster. The off-diagonal quadrants are the interesting ones, holding passers who are accurate but low-value (lots of safe completions) or inaccurate but high-value (aggressive playmakers like Stafford).

7. How to actually read the four quadrants

   Train your eye to read position, not just rank. A dot far to the right is a precision passer hitting throws others would miss. A dot high up is moving the ball and avoiding the catastrophic sacks and picks that tank EPA. The dream is up and right - Purdy's corner. A quarterback who is high but left is winning with explosiveness over accuracy; one who is right but low is accurate on a diet of low-value throws. None of that is visible in a passer-rating column, and that's the whole point of trading a single ranking for a two-dimensional map.

Troubleshooting