The strategy for forecasting the BP neural network was researched on the basis of the training-studying samples that were obtained in the orthogonal test of L₉（3⁴). The self-contained orthogonal sample was the basic training and studying cell. When others samples were added into the self-contained orthogonal samples or the self-contained orthogonal samples were cut down, the forecasting results were completely irresponsible. On the same test condition and orthogonal test type, the self-contained orthogonal sample with large information content could forecast that with small information content at high precision. A new test-design approach was put forward. Namely, the self-contained orthogonal sample was obtained through the orthogonal test, and then, the values of all other samples whose factors were the same as that of the self-contained orthogonal sample could be forecast in the BP neural network and its precision was considerable high. Therefore, the time and labors were enormously saved.