Well, Week 1 of college football is in the books. What did we learn? The conventional wisdom is: not much because there was only one matchup between AP Top 25 teams. And, if your team blew out a lower-ranked team in Week 1, I bet you heard someone say something like: “It was a nice win, but we won’t really know how good they are until they play real competition.” If your rival is one of the teams that cruised against an inferior opponent, I’m sure you believe it.
Intuitively, it makes sense. Beating an FCS or Group of Six school by 30 isn’t as meaningful as triumphing in a top-10 matchup. Surely we can learn more about a top-five team and its ability to contend for a title when it plays another top-five team. Surely.
The first time I recall trying to use this intuition was when I participated in Kaggle’s March Madness prediction contest in 2019. I figured weighting games involving tournament teams more heavily or clipping blowouts would give my model an edge. Interestingly, none of them improved my model’s accuracy by much, if at all. I was puzzled, but assumed I just did not have time to capture the dynamic correctly.
I kept trying to implement different versions of this idea in other sports, including the NBA, NFL, and college football. Sometimes it improved results a tiny amount sometimes it made them worse. Eventually, I started to wonder if common sense was missing something.
Turns out the common-sense view is almost entirely wrong. Trying to gauge how good your team is after Week 1? You should pay just as much attention to the result whether it played a cupcake or a powerhouse. Here’s why.
How much should one game change our mind?
To measure what Week 1 taught us, we need to judge each performance against what was expected before kickoff. Suppose a team was favored by 28 points and won by 31. The victory itself was expected. The team performed three points better than the pregame estimate. A 28-point favorite that wins by 45 beats the expectation by 17; one that wins by 10 falls 18 points short.
There’s a concept from mathematics, Bayes’ rule, that gives us a simple framework for measuring how much we learned. The name sounds more complicated than the idea.
Before Week 1, every team has a preseason power rating. That is the prior: our best estimate using everything known about each team before the games begin. Teams play, giving us new evidence. Combining the prior and the evidence produces the posterior: our updated rating after Week 1. Bayes’ rule provides a way to update the rating, but the size of that update depends on how much confidence we place in the prior. Here, I use a simple update inspired by that idea and let past seasons determine how much of a game’s surprise to carry forward.
One result tells us something about both teams. For this test, I give them equal shares of the rating adjustment, in opposite directions. A 0% learning rate leaves both ratings unchanged. At 50%, a 14-point surprise would move each team seven points, accounting for the entire surprise between them. That is the upper limit under this equal-split rule. The data can tell us how far below it the learning rate should be.
For this exercise, I measured accuracy using mean absolute error, or MAE. It is simply the average number of points by which the predicted margin missed the final margin. If a model predicts a 14-point win and the team wins by 10, the miss is four points. Lower is better.
I tested the question using 315 Week 2 games across eight seasons from 2017 through 2025, excluding the disrupted 2020 season. Both teams in every prediction had exactly one usable Week 1 result.
For each Week 2 game, the model began with the preseason ratings, updated both teams from Week 1, and predicted the Week 2 margin. The model selected the weight without using any games from the season it was predicting. That prevents a lucky fit in one season from grading itself.
The baseline
A 14-point surprise moved each team about 2¼ points.
The best-fitting Week 1 learning rate moved each team’s rating by about 16% of the game’s surprise. With a 14-point surprise, the better-than-expected team gained about 2¼ rating points and its opponent lost the same amount.
On this per-team scale, the preseason estimate carried about 5.2 games’ worth of weight. That is a way to describe the balance between the preseason estimate and one new result, given the uncertainty about both teams.
That is a meaningful adjustment, but it leaves plenty of room to learn more. The 16% estimate comes from this simplified historical test. The published All Takes power ratings and ATS predictions use a richer model with several sources of evidence; they do not use a single fixed 16% update.
How much should each team’s rating move?
Lower is better. The curve bottoms out near a 16% learning rate per team.
That chart shows the smooth relationship across the full sample. In the stricter test, the model had to select the weight without using the season it was predicting. The updated model missed by 12.640 points on average. Ignoring Week 1 produced a 12.984-point average miss. That result is particular to the strength of my model’s prior. A separate test starting from sportsbook-based ratings also found little benefit from discounting mismatches.
Translating intuition into testable hypotheses
Now, we are in a position to more rigorously evaluate “wait until they play real competition.” However, there are different possible meanings of this phrase, and we should give all of them a hearing.
The most direct meaning is that a game between evenly matched opponents tells us more than a mismatch. A more nuanced version says that a heavy favorite’s bad performance is revealing, but running up the score tells us little. The most complex version is that results versus weaker teams have value for future mismatches, but do not help us predict what will happen when the team finally faces real competition.1
We can translate these three versions of the argument into tests of what we should learn from Week 1.
- 01Close games are more informative
Close Week 1 games should receive more weight than mismatches, and discounting mismatch results should improve Week 2 predictions.
- 02Extra points in a blowout are empty
Capping a heavy favorite’s positive surprise—or the final margin itself—should improve the next prediction.
- 03Mismatch evidence does not transfer
Discounting a Week 1 mismatch should help when the next opponent is evenly matched.
Each test gives the intuition a chance to produce the pattern it predicts.
Test 1 · Information
Does a mismatch deserve less weight?
A 315-game sample is not huge, so the safest approach is to cut the data coarsely. I divided the Week 1 schedule at a 20-point spread. If “real competition” produces a more revealing game, the estimated weight should be higher below 20 and lower in the mismatches.
To my surprise, the point estimates went in the other direction. The best weight for games with spreads below 20 points was 11.8% per team; games with spreads of 20 points or more did best with 19.8% per team.
The mismatch estimate was higher, not lower.
The dot is the best estimate. The line is the 90% uncertainty range—the values supported by season-to-season variation.
Those estimates come with substantial uncertainty. They cannot establish that mismatches contain twice as much information, and the extra complexity of two separate weights made the held-out predictions slightly worse. They do answer the narrower question: the mismatch weight did not come out smaller. Paying the same amount of attention to every game appears to be more prudent.
Force the skeptical fan’s rule into the model
A more direct test is to impose the rule a skeptical fan might prefer. I forced the model to ignore Week 1 games with spreads of at least 20, give them half weight, normal weight, or double weight. If those games add mostly noise, ignoring or half-weighting them should improve the next prediction.
Giving mismatches the same weight worked best.
Each dot shows the average number of points the Week 2 prediction missed. Moving right means the predictions got worse.
One common learning rate produced the smallest average miss. Ignoring the mismatch results performed worst, half-weighting them also hurt, and doubling them brought no improvement.
I repeated the test using 14-, 21-, and 28-point definitions of a mismatch. Ignoring mismatches hurt at every threshold. Small gains from changing their weight depended on how accuracy was measured.2
Strike one against common sense.
Test 2 · Blowouts
Do blowout margins stop adding information?
There is a seemingly more sophisticated version of “wait until they play real competition” that focuses squarely on blowouts. The purpose of football is to win the game, not to win by as many points as possible. A 28-point favorite that struggles may expose a real weakness. Once the favorite is comfortably ahead, additional touchdowns may tell us little. The backup quarterback enters. The opponent gives up. The clock keeps running.
This theory predicts an asymmetry. Falling short of expectation should receive full weight, while exceeding it should eventually stop moving the rating.
At what point should we ignore blowouts?
For teams favored by at least 20 points, I left every disappointing performance untouched and capped only positive surprises. A 14-point cap, for example, means outperforming expectations by 30 counts the same as outperforming them by 14. If the extra points are empty, a modest cap should improve the Week 2 prediction.
Every tighter cap made Week 2 predictions worse.
The labels on the left show the cap. The number beside each dot is the average Week 2 miss. Lower is better.
For Week 1, every tighter cap made the predictions worse! The fitted weights also ran opposite to the expected pattern: favorites that exceeded expectations received a learning rate of about 24% per team, compared with 13% when they fell short. An asymmetric model using those separate weights performed worse when predicting a held-out season.3
What about the final margin itself?
A fan can mean something simpler by the blowout argument: once any victory gets large enough, the remaining points should stop counting. That is a claim about the absolute final margin rather than performance relative to the spread. Extreme margins are rare in the Week 1 sample, so I expanded this test to the full regular season.
The larger analysis includes 6,441 FBS-versus-FBS games across nine seasons, with 6,060 games from Week 2 onward used to judge predictions. Before every game, the rating used only games that had already been played. Each cap received its own learning rate, selected without the season being scored.
For a 28-point cap, any observed margin beyond 28 was recorded as 28. The expected margin was capped the same way; otherwise, a 35-point favorite that won by 35 would appear to have underperformed merely because of the cap.
If ordinary blowouts contain little additional information, caps such as 21, 28, or 35 should reduce the average prediction miss. The data give us a clear curve.
Ordinary blowout caps threw away useful information.
The bars show how much each cap changed the average prediction miss. Bars to the right are worse; tiny bars to the left are effective ties.
Conventional caps clearly hurt. A 28-point cap increased the average miss by 0.086 points. Even a 42-point cap moved accuracy in the wrong direction, although the difference was small.
Around 49 to 56 points, the differences become indistinguishable from zero. Those margins are also rare: a 49-point cap affected 123 games, and a 56-point cap affected only 38. The data support useful information through the low 40s and become unable to distinguish clipping from no clipping after that.
I also scored only future games with spreads below 10. A shallow optimum appeared around caps in the high 30s to high 40s, but it improved predictions by only hundredths of a point. Once the cap itself had to be chosen without the future season, the advantage disappeared into season-to-season uncertainty.4
Even if the 56th point carries less information than the seventh, the data never found a useful and repeatable cutoff where the remaining points became worthless. You may not like teams that run up the score, but capping ordinary blowouts threw away useful information.
Strike two against common sense.
Test 3 · Transfer
Do mismatch results travel?
Maybe your team is a flat-track bully: it piles up impressive performances against weak opponents but cannot back them up against better competition. Those mismatch results might help predict the next mismatch while telling us less about the next competitive game.
I started with Week 1, using every result to update the ratings and judging accuracy only in Week 2 games with spreads below 10 points. That left 151 close matchups. If mismatch evidence does not travel, discounting those Week 1 results should improve these predictions.
The original common-weight update missed those games by 11.906 points on average. Half-weighting the 20-point mismatches increased the miss to 11.974. Ignoring them increased it to 12.154.
The mismatch evidence still transferred.
Lower is better. Discounting the Week 1 mismatch made the next close-game prediction less accurate.
Looking only at those 151 games, the apparent best overall learning rate was about 13% per team. Once the weight had to be selected without seeing the season being predicted, that specialized approach did not improve on the original 16% update. The smaller close-game sample gives us little reason to create a separate rule.5
Strike three against common sense.
From the opener to the full season
But one game is not enough to identify a flat-track bully. An impressive opener could reflect the team’s strength, a favorable matchup, or just a good day. We need several games against different levels of competition to see whether the pattern repeats.
I therefore extended the test through the regular season, accumulating each team’s results as it played. In this version, weak opponents were FCS teams or FBS teams in the bottom fifth of the preseason ratings; strong opponents were in the top quarter.
Before another weak opponent, the model gave those earlier weak-opponent performances their usual weight. Before everyone else, it could discount them. The size of the discount was selected without using the season being predicted.
The flat-track adjustment found a tiny edge.
The best model kept 60% of the evidence in other matchups. Every held-out-season fit kept at least half; none chose zero.
Next opponent is weak100%
Next opponent is stronger60%
12.734→12.710
0.023 better12.569→12.510
0.059 better12.898→12.893
0.005 betterThe best-fitting rule kept the usual weight for weak-opponent results before another weak opponent, and 60% of that weight before everyone else. Across 5,534 games, the average miss fell from 12.734 to 12.710 points. The improvement was 0.023 points overall and 0.059 points when a team with weak-opponent history next faced a top-quarter opponent.
The exact weak-team cutoff was not specified in advance. When the cutoff also had to be chosen without the season being predicted, the overall gain shrank to 0.009 points and the uncertainty range crossed zero. Versions that counted only wins or large wins pointed in the same direction, with gains measured in hundredths of a point.6
At best, this is a foul ball. The improvement is measured in hundredths of a point, far smaller than the “wait until they play real competition” argument would lead you to expect. Every held-out-season fit kept 50% to 70% of the weak-opponent evidence. None chose zero. Even over the full season, it did not pay to disregard mismatch results entirely.
Do the best opponents reveal the champion?
You may still be skeptical that this answers what you really care about. Predicting an ordinary Saturday is one thing. The bigger question is which teams are capable of winning the national championship.
So I narrowed the test to the 19 national championship games from 2007 through 2025. For each eventual champion and runner-up, I collected every available FBS result from BCF Toys before the title game. The championship itself was excluded; earlier playoff games were included because they had already happened.
That produced 480 performances against FBS opponents. I sorted them by the opponent’s final FEI rank and divided them into three equally sized groups of 160:
- The strongest third faced opponents ranked about 10th on average.
- The middle third faced opponents ranked about 37th.
- The weakest third faced opponents ranked about 84th.
For each championship matchup, I gave both finalists the same starting point: the average of their two final FEI ratings. I then compared each earlier final score with what that shared rating and the opponent’s final FEI would have predicted. The results stay in ordinary football points: +5 means a team beat the expectation by five points on average.
Champions separated most against the weakest third.
Average points above or below the margin expected from the shared finalist rating. Higher is better.
| Opponents faced | Average rank | Games | Eventual champion | Runner-up | Champion’s advantage |
|---|---|---|---|---|---|
| Strongest third | 10 | 160 | +5.4 points | +2.4 points | +3.0 points |
| Middle third | 37 | 160 | −3.1 points | −2.3 points | −0.8 points |
| Weakest third | 84 | 160 | +1.3 points | −6.2 points | +7.4 points |
The eventual champions beat expectation by three points more than the runners-up against the strongest third of the schedule. The middle third was essentially even. But the largest difference appeared against the weakest opponents: a 7.4-point advantage for the eventual champions.
If elite opponents were the uniquely revealing test, the strongest third should have produced the clearest separation. Instead, the largest gap appeared against opponents ranked around 84th.7
This does not prove that weaker opponents are a better test. Nineteen championship games are far too few for that, and the exact size of each gap is noisy. The narrower conclusion is enough: weak-opponent games did not become empty evidence when the question was who could win the national championship.
That brings us to the final question: if no single opponent reveals a team, how quickly does the full picture come into focus?
The season-long view
So when do we know how good they are?
The matchup-weight estimates leaned in the opposite direction, the common update predicted Week 2 best, and every conventional blowout cap made predictions worse. Only extreme caps produced effective ties. The full-season transfer adjustment bought only hundredths of a point and still kept most of the mismatch evidence. Even among championship finalists, the eventual winners separated themselves most against the weakest third of their schedules. “We won’t know how good they are until they play good competition” remains wrong; “we won’t know how good they are until they play more competition” is right.
The findings above bring us back to the assumption buried inside the original claim. A matchup against a good opponent feels like a definitive test because the game is harder to win and matters more. What the data is saying is that neither quality guarantees a cleaner measurement of team strength.
A football game contains more uncertainty than we want to accept. The ball bounces a certain way. A tipped pass becomes an interception. A fourth-down conversion turns into a touchdown two plays later. One team’s best pass rusher draws an opponent who can handle him; another finds the weak tackle who cannot. Players get hurt. Coaches change the plan. Teams improve and decline during the season.
That uncertainty exists when Ohio State plays Ball State and when it plays Texas.
The final score of a mismatch is easy to predict in one sense because the favorite will probably win. By contrast, its performance relative to expectation remains uncertain. That surprise is where the information lives. A game with a 30-point spread can therefore carry similar information to a game with a 3-point spread.
How much clearer does the picture get each week?
To measure that progress, I compared the prediction made before each game with a second prediction made using the rest of that season. The second version gets the benefit of hindsight, but leaves out the game being predicted. How quickly does the live model catch up?
Across 6,441 FBS-versus-FBS games, the picture cleared gradually. On the study’s measure, about half the progress toward the hindsight estimate came in the first three games. Reaching 80% took about seven, and 90% took about ten.
Learning about a team takes almost the entire season.
Progress toward the season’s hindsight estimate, using prior FBS-versus-FBS games.
Even after accounting for team strength, final margins typically vary by about 14 points around the model’s expectation because of everything unique to that game: matchups, turnovers, injuries, play-calling, and plain luck.
Even after ten games, one season still leaves unanswered questions. If a highly ranked team destroyed an overmatched opponent in Week 1, we learned something. The amount depends on how it performed relative to a reasonable expectation, not whether the opponent’s logo looked impressive.
The next time someone dismisses a Week 1 blowout, ask whether the favorite performed better or worse than expected—and by how much. That evidence remains incomplete, but it is the best the game gave us.
We are never waiting for the one game that will tell us how good a team really is. We are accumulating noisy evidence, one Saturday at a time.
Methodology and disclosure
How the tests were run.
The Week 1 analysis covers 315 Week 2 games in eight scored seasons: 2017–2019 and 2021–2025. Each team had exactly one usable Week 1 result. The preseason power rating supplied the initial estimate, and Week 1 performance was measured relative to the pregame expectation. Weights were fit while holding out the entire season being scored.
This exercise starts from the football-only preseason ratings. For FBS opponents, the surprise is the final margin minus the preseason model’s expected margin. For FCS opponents, a historical point spread supplies the expectation because the model has no FCS rating. These tests evaluate later score predictions, rather than the profitability of the published ATS policy.
The Week 1 learning rate is the fraction of the original game surprise applied to each team, with equal and opposite changes. The full-sample estimate is about 16%, with a season-resampled 90% range of roughly 12%–20%. It was fitted by minimizing squared prediction errors; fitting directly for average absolute miss gives 16.2%. The chart shows selected points on the full-sample curve, while the policy comparisons use seasons excluded from fitting.
The effective sample size uses the same per-team learning rate: preseason weight divided by new-game weight, or (1 − learning rate) ÷ learning rate, is about 5.2. This shorthand treats a game as evidence about one team while allowing for uncertainty about its opponent. It describes this first update; it does not imply that every later game adds an identical amount of information.
The full-season clipping analysis covers 6,441 regular-season FBS-versus-FBS games from 2017–2025, with 6,060 Week 2-or-later predictions scored. Ratings were updated chronologically. Each candidate cap received a separately selected learning rate, with the scored season excluded from selection. A non-garbage-margin version produced the same broad conclusion.
The flat-track extension uses only the article model and scores 5,534 Week 2-or-later games from the same eight seasons. Ratings were updated chronologically, and every transfer weight was selected with the scored season held out.
All prediction-error comparisons use mean absolute error: the average number of points by which the predicted margin missed the final margin. Lower is better. The reported research is historical model evaluation, not a claim that any individual future result is predictable.
Notes and robustness checks
None of the alternative mismatch weights or blowout caps significantly improved our ability to predict games out of sample. The full-season transfer test found a small improvement worth watching. The details below show how modest the effects were. Uncertainty ranges resample entire seasons, so repeated games involving the same teams are not treated as independent evidence.
- This version is closest to EvanMiya’s relative-rankings idea. See EvanMiya ↗ ↩
- Mismatch weighting. Extending the test through Games 2–5 produced 1,726 held-out predictions. Equal weighting missed by 12.733 points on average, compared with 12.752 for half weight and 12.749 when the mismatch weight was allowed to vary. A separate 2014–2025 spread-only model reached the same conclusion. In the original Week 2 test, doubling the weight of 28-point mismatches reduced average absolute miss by 0.032 points, but worsened root mean squared error by 0.024 points.↩
- Positive-surprise caps. No cap was best across all later games in both samples. A 14-point cap made the average miss 0.042 and 0.046 points worse. Its apparent 0.036-point gain among close games in the first sample was uncertain and shrank to 0.004 in the longer sample. ↩
- Absolute-margin caps. Caps at 14, 21, 28, and 35 points made the average miss worse by 0.290, 0.156, 0.086, and 0.041 points. The apparent 0.002-point gains at 49 and 56 were effective ties, and choosing the cap without the test season did not improve accuracy. ↩
- Transfer to close games. In 862 later games with spreads below 10, half-weighting earlier mismatches lowered the average miss by just 0.043 points, with an uncertainty range that crossed zero. In the longer 1,134-game sample, the gain was 0.002 points—an effective tie. ↩
- Flat-track model. The primary bottom-fifth definition improved all-game accuracy by 0.023 points, with a season-resampled 90% range of −0.0004 to +0.046. The 0.059-point gain against top-quarter opponents had a +0.027 to +0.091 range. When the weak-team cutoff was also chosen outside the test season, the overall gain fell to 0.009 points, with a −0.009 to +0.027 range. Across held-out seasons, the selected transfer weight ranged from 50% to 70%; zero was never chosen. ↩
- Championship-strength test. Both finalists in each title game were assigned the average of their two final FEI ratings, so the comparison does not begin by assuming that the champion was stronger. A version using BCF’s opponent- and venue-adjusted performance measure produced champion advantages of +4.0 points against the strongest third, +0.3 against the middle third, and +6.5 against the weakest third. Resampling complete title-game seasons put the 90% ranges for the primary final-score gaps at −1.0 to +6.8, −5.2 to +3.6, and +4.0 to +10.7 points. Comparing the finalists within each title-game season, the champion performed better in 12 of 18 available strongest-third comparisons, 9 of 18 middle-third comparisons, and 15 of 19 weakest-third comparisons. The championship game itself was excluded; earlier playoff games were included. Opponent groups use final FEI rank, and FCS games are omitted because BCF does not publish Game Ratings for them. Because final FEI is retrospective, this is a historical comparison rather than a pregame forecasting method. ↩
- Season learning. The season-learning curve uses the 6,441 regular-season FBS-versus-FBS games from 2017–2025. For each game, I compared the live predicted margin with a hindsight prediction from ratings fitted using the other games in that season. I fitted an exponential decline to the squared differences between those predictions. The chart reports the share of that fitted initial disagreement removed. Game counts use the smaller of the two teams’ numbers of prior usable FBS-versus-FBS games. Resampling whole seasons put the 50% landmark between 2.72 and 3.27 games, the 80% landmark between 6.33 and 7.60, and the 90% landmark between 9.05 and 10.87. The hindsight estimate is what this season can support, not a direct measurement of true team strength. ↩