Is There a Bye-Week Advantage?

Test 1 · The raw numbers

Every game where exactly one team was coming off a bye (13+ days of rest), seen from the rested team's side. "Expected" is what the two teams' ratings and the venue predicted, before accounting for rest.

Test 2 · Accounting for who was playing

A regression of the final margin on the rating gap, home field, and whether either team was coming off a bye or playing on a short week (5 days or fewer, like a Thursday game after a Saturday). Each rest effect is what's left once the rating gap and venue are priced in. The 95% intervals come from resampling teams (2,000 times), and the p-values from shuffling the rest labels between games (2,000 times).

Test 3 · Margin against expectation, by rest

Every team's side of every game, grouped by how many days it had off, showing how far its margin beat or missed what the ratings and venue expected. Lines are 95% intervals; a bar is highlighted when its interval clears zero.

Four groups are compared here, so one clearing zero by chance is more likely than the intervals alone suggest. Treat a single highlighted bar as a lead, not a result.

Who do teams play after a bye?

Every regular-season game a team played right after a bye, against every game it played on a normal week (6-8 days' rest), by how the opponent was ranked going into it.

Teams face tougher opponents after a bye: 3.7 rank places better on average than on a normal week (95% CI +0.40 to +6.89, p = 0.024).

Based on 718 post-bye and 4,485 normal-week team-games.

Season by season

The Test 2 bye effect fitted to each season on its own. One season has far fewer bye games than the pooled total, so expect these to swing.

More questions

One thing to know first, because it touches all of them: in any week, a team coming off a loss beats what the ratings expect by +0.1 points (95% CI -0.97 to +1.27) - the ratings overreact to a single loss. Every test below holds that constant. With it held constant, the headline bye effect is +0.78 points instead of +0.79.

Asking 10 questions means about one could clear p < 0.05 by chance, so each also shows q, its Benjamini-Hochberg adjusted value. A result is only called real when its interval excludes zero and q is under 0.05. Bars show the bye's worth in points of margin, with 95% intervals.

The postseason, tested separately

176 bowl and playoff games (28 in the College Football Playoff), kept out of every test above. Byes barely exist here - the typical team had 27 days off - so rest is measured as how many more days one team had than the other.

No detectable effect: each extra day of rest is worth -0.0 points of margin, well within what chance alone produces (95% CI -0.46 to +0.55, p = 0.971).

Do the ratings predict bowls as well as the regular season?

"Favourite" and "miss" use a model fitted to the regular season and applied to the postseason. Opt-outs and motivation are the usual suspects when bowls go off script; this data can't tell them apart.

When one team had a week or more extra rest

The more-rested team won 17 of 42 such games, and 3 of 12 in the playoff.

Margins are from the more-rested team's side. These are a handful of games - a story, not a statistic.

How this was done

• Rest is the days since a team's previous game that season, including games against FCS opponents. A team's first game of the season has no rest value and isn't included.
• Team strength is each team's rating going into the game, built only from earlier games, so no game helps explain itself. It uses the v2-iterative-adjusted model, which has weekly ratings for every season in the study.
• Only FBS-vs-FBS games where both teams had a rating: 2,806 games, with 192 left out for having no rating yet (mostly the first week). Bowls and the playoff are only in their own section, never in the regular-season tests; a postseason game is rated with each team's final regular-season rating.
• This is observational: byes aren't assigned at random, and schedule makers may place them around big games. The regression controls for team strength and venue, not for everything.