A measure of an AVM’s precision, median absolute error enables testers to determine the ‘middle’ degree of variation between an AVM’s outputs and the corresponding benchmarks. To calculate this metric, analysts take the absolute value of the percentage error (i.e., the magnitude or size of the variance) instead of distinguishing between positive and negative error. For example, a median absolute error of the 8.75 percent suggests that half of the model’s predictions are within 8.75 percent of the sale price (whether above or below) and half are outside of that range. As with measures of AVM bias, the majority of users prefer median absolute error to mean absolute error because the former is less sensitive to a few anomalous outliers.
« Back to Glossary Index« Back to Glossary Index