# Good NFL bets go bad in Week 9, says Harvard Analytics

After eight weeks, we are halfway through the NFL season and we seem to have a slightly better idea of which teams are good and which are bad.

We came into the season with preconceived ideas, and while some teams have lived up those (Patriots, Broncos) others have either fallen way short (Cardinals, Panthers) or vastly overachieved (Vikings, Dallas).

So, as we approach the halfway point, it is worth considering how a team’s performance in the first half of the season translates to their performance in the second half, especially when it comes to performance against the spread.

There are three separate measures of performance that it is worth looking at. First, and most simply, is winning percentage. We would expect the relationship between winning percentage before and after each week to be fairly high, as teams will probably win at the same rate before and after.

Next, we can look at the winning percentage ATS. We would expect there to be almost no relationship here if lines are perfectly accurate, as perfectly accurate lines would mean that each team has a 50 percent chance of winning ATS so any deviations from that would be solely due to chance.

Along those lines, a third measure we can look at is the rate at which teams go Over the betting total, again hypothesizing that there would be no relationship as with perfectly accurate totals any deviation from 50 percent Over would be solely due to chance.

As a measure of the strength of the relationship, we will regress the various metrics after each week against the metric before each week (i.e. after Week 8 we will regress winning percentage in the first eight games against winning percentage in the second eight games).

If performance prior to the week is a significant predictor, then that would suggest that there is a strong relationship between the two. To determine significance we will look at the p-value associated with the coefficient of the regression.

If you recall, a 0.05 p-value is the most commonly accepted threshold for statistical significance, where the lower the p-value the more significant the finding. After looking for significance, we can then look at the coefficient itself to see in what direction the relationship goes.

For example, let us do this for winning percentage straight up. For Weeks 1 through 16, we ran a regression on winning percentage after that week on winning percentage prior to that week. We then plotted the p-values by week below.

As we can see, the p-value is always incredibly low, well below 0.05. The only two weeks where it is basically non-zero are the first and last week, following the intuition laid out above that one week may not be enough information as to a team’s true winning percentage, whether that is the dependent (one week left) or independent (after one week) variable in the regression.

And checking the magnitude of the coefficient, we can see that it is always positive, as we would expect. This means that if Team A wins at a higher rate than Team B up to Week X, we can generally expect that they will continue to win at a higher rate after Week X as well. Not a surprising finding, but one we included as it is hopefully fairly simple and intuitive.

Now, let us do the same with winning percentage ATS. We have included a line at 0.05, so any points below that line would suggest significance for that week.

Somewhat surprisingly, we can see the p-value dip below 0.05 towards the middle. That suggest that how a team does ATS up to Week X (if X is a week towards the middle of the season) actually provides some information on they will do after that week. But what is the relationship like?

Is it positive (i.e. a team with a good ATS record is likely to continue to do well ATS in future weeks) or negative (a team with a good ATS record will actually do worse ATS)? In order to determine that, we have plotted the coefficients of the regression by week below.

As you can see, the coefficients for the significant weeks are all less than zero. This suggest that teams that perform well ATS up to a certain week (around the middle of the season) are likely to do worse ATS after that. The coefficient is the largest in magnitude (and still significant) around Weeks 8 and 9.

Now let us look at Over percentage, again plotting a line at 0.05.

Even more surprising than before, this relationship is significant for basically the whole season. Let us again look at the coefficients.

Again, the coefficients are all negative, suggesting that teams that go Over the betting total more than Under up to a certain week are less likely to go Over after that.

Main Takeaways

There have been a lot of graphs and some more technical jargon in this post, so to briefly summarize relevant points:

* Around halfway through the season (Weeks 8 and 9) exists the most significant, negative, relationship between ATS performance before and after that week.

* That means you should bet against teams that have done well up to that point

* Basically at any point in the season, you should take the Under on teams that have been going Over more often than not

Although this negative relationship may be surprising to some, it’s actually very much in line with many of our other posts. Lines are often very accurate, so when a team does well ATS it is often (but not always) just a product of randomness.

However, people in general are very bad at recognizing randomness and therefore read too much into an ATS streak. Because lines have to adjust to what the bettors are thinking, they then overvalue teams that have done well ATS. Same logic goes for the Over/Under.