How efficient has the NFL market been? Its hard to use the term efficiency when the mode of efficiency is scarcely measuring veracity. Efficiency here being designated as information that is verified and available, constraining the line to a more precise number. In other words, time and data calibrates expectation. The only thing that resonates with an NFL market is how much commission is being made. Since lines are appropriated solely on its ability for a sportsbook to earn commission, the optimal strategy is to find a number that on average would procure the most commission. Using that concept, efficiency and NFL markets, or any sports betting market, are mutually incompatible. If a slight difference between the closer and the actual result were to be desirable, from the view point of a sportsbook, then we would expect as the week progresses, new information to breed a more accurate win probability expectation. But an expectation of win probability is immaterial to a linesmaker, where the only way to exploit the system is to, on average, procure 50/50 action or better. Obviously the opener will be posted in close proximity to an extrapolated central tendency for that game. The market won’t support a line disproportionate from standard expectations, such as a 10-0 time listed as a 10 pt home underdog, again for obvious reasons.
Many essential features in sports betting markets are over analyzed. And it is my belief that, for the most part, line movement is purely random fluctuation taking up market space, thus closers being explicitly a stoppage of variation.
Let’s look at the veracity of the NFL closing spread from Vegas covering the 32 seasons since 1978 (1978-2009). The first graph shows the correlation between margin of victory and the closing number. Actually its the unit-less correlation squared, indicating the coefficient of determination, R2, and is simply the proportion of variance in margin of victory (or defeat) that can be explained by the spread, or simply the strength of the relationship.
I drew a trend line to show the fallacy in relative trend lines. First, one should realize the only pattern conveyed by the graph is that of random fluctuation, and I’ll explain why there is no tendency to move “somewhere”, in this case the trend line showing a veneer of progression. The initial data point has roughly an R2 equal to .15. Not a very strong relationship, and because of this, one would think the next year would undoubtedly be inclined to see an increase. It turns out this is not the case, but the subsequent years an increase would yet be more than likely. Why would this be? Well in any system with a statistical barrier, as more data comes in there will be a higher potential for data points to be positioned further from the lower wall, while at the same time the magnitude of fluctuation will decrease. The coefficient of determination basically has one way to move, because based on simple mathematical principles, a real number squared is > 0. The lowest conceivable coefficient of determination is zero. The subsequent numbers are using up random statistical space, and the only trend that has occurred, the only thing that can be taken from the graph is the amount of variance above the lower wall of zero. In any system the range will increase through time, certainly in a system where veracity and efficiency is not the main objective.
To prove my point on the matter, here is the same graph using the closing line and game result data from 1984-2006.
A sagacious viewer can deduce the nature of the graph.
Here is another graph showing the root mean square error (RMSE), simply for ocular sophistication. RMSE is an aggregation of the differences between spread and game result, serving as a way to further measure precision, though adducing implications that slightly vary from calculating R2.
The same methodology can be used to explain this graph as well. Statistics often are just numbers taking up random market space.