Question for Ganchrow

[cincy_1]

I was using your calculator for the half point and I noticed that there was no box to input the game total for NFL games ... which implies that the value of the half point is the same for a game lined at +3 (with a total of 34) and another game at +3 (with a total of 52).

In MLB, this is not the case ... the +1.5 line is cheaper when the total is larger .. using the same logic, the half-point should be more valuable when the total is smaller, no? ... especially when one total is 150% of the other total.

[Ganchrow]

I was using your calculator for the half point and I noticed that there was no box to input the game total for NFL games ... which implies that the value of the half point is the same for a game lined at +3 (with a total of 34) and another game at +3 (with a total of 52).

In MLB, this is not the case ... the +1.5 line is cheaper when the total is larger .. using the same logic, the half-point should be more valuable when the total is smaller, no? ... especially when one total is 150% of the other total.

You are completely correct. (This is one of the reasons why there is not a spread option for MLB. That and the very large dfference in point values between home and away teams.)

The reality is that it's somewhat difficult to explicitly take this into account through pure data sampling. There are few enough data points in the NFL as it, and further partitioning the data is likely to ultimately lead to even less predictive accuracy (especially for the less trafficked numbers).

To give you example, the push prob for a 3 using the entire data set is about 9.79%, with a standard error of 0.19%. When only considering totals of more than 39.5, that percentage falls to 9.74% with a standard error of 0.27%. When only considering totals of 39.5 or less, that percentage jumps to 9.83% with a standard error 0.26%. So in other words the 2σ confidence interval for the data set is entire entirely contained within each the of the two partitioned confidence intervals.

In practice, the best way to take this into account would be to combine data sampling with a score distribution model. Users should certainly include their own push estimates in the calculator if ever they differ substantially from mine, and should certainly consider tweaking them in response to match-up specific circumstances.

The default push percentages are fair to very good estimates of the true underlying probabilities, but they aren't great estimates -- especially in special cases such as the one you've pointed out.