Filter Garbage Time from Play-by-Play Data (Python)

When a game is decided by the third quarter, the snaps that follow — backups in, prevent defense, clock running — aren't really competitive football. Left in your data, this "garbage time" quietly distorts efficiency stats. Cleaning it out is a small, important habit. Here's how to define and filter it from play-by-play. Full code: scripts/clean-garbage-time-pbp-python.py.

Define garbage time

There's no single official definition; a clean, common one is: the 4th quarter with a margin of 22 or more points (a three-possession-plus lead late). You can also use win probability (say, above 97%). We'll use the score-and-period rule because it needs nothing but the play data:

def is_garbage(play):
    period = play["period"]["number"]
    margin = abs(play["homeScore"] - play["awayScore"])
    return period >= 4 and margin >= 22

Filter and re-measure

Pull a blowout, score every scrimmage play's success (the 50/70/100 rule), and compute success rate twice — all plays, then competitive plays only:

for p in plays:
    if not scrimmage(p): continue
    good = successful(p)
    all_total += 1; all_ok += good
    if not is_garbage(p):
        comp_total += 1; comp_ok += good

The result

On the most lopsided 2024 CFP first-round game — Penn State's rout of SMU:

Game: SMU Mustangs at Penn State Nittany Lions
All scrimmage plays: 138  | success rate 42.8%
Garbage-time plays removed: 29
Competitive plays only: 109 | success rate 45.0%
Difference: -2.2 pts of success rate

Nearly one in five plays (29 of 138) happened after the game was effectively over. Removing them moved the game's success rate by 2.2 points — here it went up, because garbage-time football (clock-killing runs, vanilla calls, backups) tends to be less efficient than competitive football. The direction varies by game; the point is that those plays are measuring something different from the contest you actually want to analyze. Over a season, garbage time can meaningfully flatter or punish a team's numbers depending on how many blowouts it played.

Important caveats

• Not all feeds carry running scores per play. If homeScore/awayScore aren't on each play, fall back to a period-only rule or align with the win-probability series. Our script notes when no plays met the test.
• Don't over-filter. Too aggressive a threshold throws away real competitive snaps. 22 points in the 4th is conservative.
• Be consistent. Whatever rule you pick, apply it everywhere so teams are compared on the same basis.
• Filter per analysis, not per dataset. Keep the raw data; filter in code so you can change the definition later.

This is the unglamorous data-cleaning step that separates a number you can trust from one you can't. The best analysts spend more time here than on the modeling.