Algorithm? Fancy word for "assume whatever trend was in the last two polls will continue until election day."
Well, the way he made the calculated prediction was essentially to get the best fit for the data points to two straight lines (one for Kerry and one for Bush) and then make the prediction be whatever the percentages are when the x-value on the line is November 2.
Unfortunately, there were only two whole polls from Utah, which made this procedure rather subject to the margin of error, and he edited the algorithm to simply use the last poll received if there were three polls or less from that state.