As part of this web site, I calculate E-Ratings for many sports using the basic ideas of Mr. Potemkin. Both the Buyer and Seller models are used. Instead of subtracting (Rating = Seller - Buyer), I prefer to use a geometric average of Seller and 1/Buyer (Rating = SQRT(Seller/Buyer)). The reason for this is that the E-Ratings are based on what is essentially an exponential model. Additional adjustments are made to the algorithm in order to estimate the home advantage constant.
The essence of the resulting ratings is that the ratio of points scored to points allowed can be determined. However, this is an awkward piece of information. Instead we would like to be able to estimate the point spread and the probability that one team would beat another.
I have devised an admittedly inelegant way to accomplish this which won't be described here in detail. Since the E-Ratings are basicly exponential, I first convert them to a linear scale by taking logarithms. Then they are scaled to "fit" the actual game scores played so far.
Kenneth Massey, January 1999.