A pattern recognition algorithm is developed to provide potential improvements over existing earthquake prediction practices. The parameters employed in the analysis include degree of spatial nonrandomness in two distance ranges, spatial correlation dimension, spatial repetitiveness of earthquakes with a similar size, average depth of events, time interval for the occurrence of a constant number of events, and ratio of the numbers of events in two magnitude intervals. The parameter temporal variations are compared quantitatively with the time series of large events using a technique of association in time. The significance of the association frequencies is evaluated by comparison with chance associations estimated from corresponding simulated random time series. The developed techniques differ from existing approaches in the following aspects. The parameters here emphasize the spatial distribution of earthquakes. Possible correlations among the parameters are evaluated to determine the final set of parameters to be monitored. Threshold values for the assumed anomalies are chosen with consideration of properties of the available earthquake catalogs, such as the number of large events to be retrospectively predicted. Equal weight is given to both locally high and locally low parameter values. Care is taken to distinguish between anomalies preceding large events and those following previous events. It is shown that the relationship between precursory local extrema and precursory trends is nonunique, with precursory local extrema of the same type frequently associated with opposite observable precursory trends. The application of the seismicity parameters and pattern recognition techniques is demonstrated using synthetic earthquake catalogs generated by models of segmented fault systems in a three-dimensional elastic solid <Ben-Zion, 1996>.¿ 1997 American Geophysical Union