A measure of AVM bias, or the extent to which an AVM tends to overvalue or undervalue subject properties. Analysts determine the median variance by sorting the percentage error for each valuation from lowest to highest. For samples with an odd number of results, the ‘middle’ percentage error represents the median error. In instances where there is an even number of results, the median is the average of the middle two percentage errors. A median error of 2.5 percent suggests that 50 percent of the AVM’s valuation errors will be greater than 2.5 percent and 50 percent of the model’s value predictions will be at least 2.5 percent lower than the benchmark value. In general, the median error is preferable to the mean because a small number of unusually low or high value estimates will not distort the results.