Watching the low scoring game on Thursday night between Minnesota and Washington got me to thinking whether teams might perform better or worse when they play on certain days of the week. For example, it makes sense that professional athletes might not be expected to be at the peak of their athletic ability when they have less turnaround time between games, like they would when playing on a Thursday night with only four days recovery after playing the previous Sunday. Conversely, they might be expected to play even better when given an extra days rest when playing on a Monday night. I decided to check this out myself using the NFL season data I've been collecting for the past 3.5 seasons.
To assess each teams relative performance on a given day, I calculate the average amount of points their offense scores and I compare it to the average amount of points the other team's defense allows up until that point in the season*. I then take the mean of these two numbers, which gives me a bench mark of what I expect both the offense and defense to achieve for that week. Any points scored/allowed above or below this mark represents either an improvement or a regression in performance, depending on whether it's the offense or defense. For example, in the game between Minnesota and Washington, the Minnesota offense on average scores 23.7 points/game, while the Washington defense on average allows 25.1 points/game. I therefore expect Minnesota to score the average of these two numbers which is 24.4 points, and the fact they only scored 19 points represents a regression of 5.4 points for their offense, or an improvement of 5.4 points for the Washington defense, depending on how you look at it.
Given in Figure 1 are the performance values for all games played since the beginning of the 2016 NFL season. Each dot on the graph represents the results from one team, while the box plots are the distribution of these values as a function of day. The min/max, quartile and median information is given at the top of each day of the week. Most teams by far play on Sunday (1527 green dots), with the next amount of games played on Monday (120 purple dots), then Thursday (115 blue dots), and finally games played on Saturday (66 red dots). The blue dashed line in the middle represents what we would expect each team to score based on their season averages and who they were playing that day.
Based on the distribution of the relative performance, I'd say that there is about an equal chance of a team playing better or worse on Thursdays (median=0.10) and definitely on Sundays (median=0.00). I would also say that NFL teams tend to score more points than expected on Mondays (median=0.93) and less points on Saturdays (median=-0.86) based on the plots. I'm not subjecting these data to statistical analysis, I'm simply observing how these data lay out across all the different days.
One variable that I know that plays an important role in team performance is location of the game. Given above in Figure 2 is the same observed minus predicted scores in the previous figure, but broken down by whether that team was playing at home (right side of graph) or away (left). You can see that the median score for all days is well into the positive category when playing at home, but the median score for all away games, with the exception of Monday, is negative when playing away. In fact, the disadvantage of being the away team seems to decrease the farther they get away from their last game. I do think fatigue is a real factor for the away team, and I think there is evidence this increases over the season (more on that later).
Given above in Figure 3 is the performance values for the New Orleans offense, a side that usually scores more points than expected when playing at home. You can see the median performance values for all days increase when playing at home, especially on Monday nights. In the last 3.5 years, the Saints have scored an average of 10.58 points less than you would expect on the road on Monday night, while the median for home games on the same night is 8 points more than you would expect. I have noticed a similar trend for offenses that have played with the same quarterback for many years (Ben Roethlisberger and the Steelers offense comes to mind), so maybe there is an advantage to your signal caller being comfortable in a place he has played many, many times. The home effect for New Orleans has decreased with the injury to Drew Brees, which seems to support this theory.
As you would expect, this home and away effect is seen not only in scoring ability, but also the ability to hold your opponents to less points. Given in Figure 4 are the performance values of the Minnesota defense when playing away (left) or at home (right). You can see that there is a clear trend for Minn. to hold their opponents to less points than you'd expect when playing at home, as they did with Washington in their last contest last Thursday (marked on graph). It is very rare for both a team's defense and offense to perform better at home (it's usually one or the other) and can fluctuate with injuries to key players. You might think that domed stadiums are lucky for the home team given my use of Minn. and New Orleans as examples, but I don't see much of a home effect with either the offense or defense for Detroit or Atlanta.
Given above are the performance values for the New England Patriots, who are unquestionably the best team in the league right now (I hate to say it, but it's true). You can see that their defense (top portion of the graph) typically holds their opponents to less points than expected, and these values are about the same whether they play away or at home. But it's the offensive performance at home that truly stands out (lower right side of graph). You can see all median performance values for all days well into the positive category for home games, especially those played on Saturday during the playoffs. Right now the New England offense under Tom Brady (another QB that has played many, many games at the same home stadium) scores on average 7.7 more points than you would expect at home (p=0.02, t-test) based on their season average and competition. No wonder they always go to the Super Bowl when they win their division and secure home field advantage because it's hard to score more points than they will at home.
It's pretty obvious that there is a home/away effect regarding performance, but I wanted to see how this effect plays out over the season. Given above in Figure 6 is the average performance for all teams each week of the season for the past 3.5 years for both home (red) and away (blue) games. The shaded, colored area around the lines is the confidence of fit (bootstrap) and reflects the amount of variability surrounding this average; the wider the distance, the more varied the performance levels at that point in the season. You can see that both lines start out the same at the beginning of the season, then diverge until about week 7 (about the time most teams are coming off of or entering their bye week), then seem to repeat the divergence pattern until both home and away averages decrease precipitously starting at around week 10 and week 9 respectively. It is interesting that these two effects kind of level off towards the end of the season, but then the home team seems to perform much better once the playoffs start by week 18.
Given above in Figure 7 is the same performance information over the season, just broken down by what day the game was played. You can see the huge variability in the performance of games played on Thursday, and likely reflects the wide advantage the home team gets on that day due to the disadvantage of the traveling team playing with only 4 days to recover from their last game. You can also see a huge spike in scoring on Monday nights on week 10 and 11 that I can't really explain, and a very tight distribution of performance for the usual Sunday games that hovers around a mean of zero (no advantage).
So what does this all mean for predicting NFL games? I would say that even though San Francisco is favored by 10 points over Arizona in tonight's Thursday night game, I would say that Arizona has a good chance of covering the spread given that they're the home team and it's past the half way mark in the season. I would also say that the next two Monday night games for week 9 and 10 should be high scoring affairs if the past 3 years have anything to say about it (odds that they go over the predicted total score increase), and that I would be hesitant to bet against the home team during the playoffs, especially if the home team is New England. And Denver still sucks, but so does Oakland, so I don't feel so bad. :)
-Michael Edwards, Bioinfo Solutions LLC
P.S. I'm sure you'll find all kinds of grammatical errors in the above post, but I always considered myself a scientist first and not a novelist. I'd rather work with data than proofread copy, so let me know about any mistakes (firstname.lastname@example.org) and I'll fix them later.
*I've played with several different ways to compute the average performance values for a team's defense and offense over the years. The question becomes how to calculate the performance of a team at the beginning of the season without any data. I originally used the overall averages from the previous season as a starting point, but the distribution was all over the place in the beginning of the season due to so few data points. I then started to incorporate the scores from midway through the previous season (week 8) to add to the current one, which seems to stabilize the scores at the beginning of the season. My median value of performance for Sunday games is exactly 0 using the current method, so I think it's now evenly balanced.
**Microsoft Excel for Mac ver.16.2 and JMP Pro ver.14 were used to collect, analyze and display NFL data.