By combining PP technique with auto-regression model （AR）

this paper proposes optimum projection pursuit auto-regression model（PPAR）

and realize two-factor prediction of occurring magnitude and time in selected regions

i.e. building up PPAR model of two-dimensional time sequence. This study chooses 14 research regions with different spatial range from Xinjiang Tianshan Mountain region Larger regions are forit chosen

then smaller regions for building corresponding PPAR models. For every research region

building PPAR model respectively with such sequences as undeleting aftershocks

deleting aftershocks and deleting both aftershocks and foreshocks are deleted. In general

PPAR model built by sequences with deleting both aftershocks and foreshocks is reasonable. Comparing PPAR moedls of all regions with different magnitude thresholds

it is shown that prediction efficiency of both magnitude and time sequence with magnitude thresholds M0 = 4. 5 are reasonable

and the qualified rate of 3 of 5 regions （smaller range

weaker seismicity）are above 75 per cent. The efficiency larger range

stronger seismicity) with magnitude thresholds M0 = 5. 0 are high with qualified rate of 5 of 6 regions above 62. 5 per cent. The efficiency （larger range

stronger seismicity） with magnitude thresholds M0 = 5. 5 are very high with qualified rate of 1 of 2 regions above 80. 0 per cent. Results from different research regions show that qualified rates of checking from PPAR models of larger regions with high frequency and strength of earthquake occurrence are high

therefore PPAR models are applicable and efficient. The confidence of the model for smaller regions with low frequency decreases due to limitedsamples

but prediction for magnitude and time sequence is still of practical value.