FiveThirtyEight’s NFL Game Spreads Worse Than Bovada’s for 2020 Season

QB Elo and Traditional Elo both disappoint.

Objective

The goal of this analysis is simple: determine whether FiveThirtyEight’s game spreads for the 2020 NFL season were any better than those listed by the popular sports betting site Bovada.

I know there’s one week left, but with many teams playing for nothing, resting starters, etc., I prefer to run the analysis on weeks 1–16.

In my view, there are three possibilities:

1. If FiveThirtyEight’s spreads are quite a bit more predictive than Bovada’s, they might be useful for betting purposes.
2. If they’re at all more predictive than Bovada’s, they might be useful for other purposes, e.g. generally gaining insight into upcoming games.
3. If they’re worse than Bovada’s, they’re worthless.

Data Source and Methods

FiveThirtyEight publishes game spreads for two different methods, called “QB Elo” and “Traditional Elo.” Both are available here for every game of the 2020 season. Historical Bovada spreads are available on Sportsbook Review.

To compare the predictive value of FiveThirtyEight’s spreads vs. Bovada’s, I use the following metrics:

• Mean error, median error, and mean squared error, where “error” is defined as the actual margin minus the predicted spread
• Pearson correlation and Spearman correlation (more robust to outliers) between spread and actual margin
• Percent of games where one spread is closer than another to the actual margin
• Percent of games where the correct winner is picked

Bovada’s spreads also have corresponding odds, which are usually balanced to give about equal weight to either side of the bet, but sometimes not, e.g. Rams (-15.5) vs. Bears at +245 on 10/26/20. I exclude that game and 8 others with unusually high odds of -130 or -125, leaving 231 games in my dataset.

Results

Correlations among spreads

Correlations among the three spreads are shown in Table 1. They’re high, in the 0.9–0.94 range.

Table 1. Pearson/Spearman correlations for spreads.

Predictive performance

Performance metrics are compared in Table 2, with FiveThirtyEight cells color-coded green if better and red if worse than Bovada.

Table 2. Performance metrics for three spreads.

Both FiveThirtyEight spreads lagged Bovada in all categories, with one exception: QB Elo predicted the right winner at a slightly higher rate.

Results when spreads are dissimilar

To give FiveThirtyEight one last chance at demonstrating some utility, let’s look at games where their spreads are quite different from Bovada’s.

Table 3 lists the 20 games where QB Elo and Bovada differed by at least 4 points. FiveThirtyEight’s spread proved closer to the actual margin in 9/20 games. The correlation between QB Elo spread and actual margin was weaker than the correlation between Bovada spread and actual margin (0.502 vs. 0.597).

Results were similarly unimpressive for Traditional Elo (not shown). In the 23 games where it differed from Bovada by at least 5.5 points, Traditional Elo was closer 10 times, and its correlation was again lower than Bovada’s (0.464 vs. 0.589).

Table 3. Summary of games where FiveThirtyEight’s QB Elo spread and Bovada spread differed by at least 4.

Conclusions

Neither of FiveThirtyEight’s algorithms for predicting NFL game margins appears to be as good as Bovada’s spreads. In my opinion, that makes them worthless for betting purposes and in general.