The aim of loop prediction in protein homology modeling is to connect the main chain ends of two successive regions, conserved in template and target structures by protein fragments that are as similar to the target as possible. For the development of a new loop prediction method, examples of insertions and deletions were searched automatically in data sets of structurally aligned protein pairs. Three different criteria were applied for the determination of the positions where the main chain conformations of the proteins begin to differ, i.e., the anchoring groups of the insertions and deletions, giving three test data sets. The target structures in these data sets were predicted by inserting fragments from different fragment data banks between the anchoring groups of the templates. The proposals of matching fragments were sorted with decreasing correspondence in the geometry of the anchoring groups. For assessment of the prediction quality, the template loops were substituted by the proposed ones, and their root mean square deviations to the target structures were determined. In addition, the best 20 fragments in the whole loop data bank used-those with the lowest deviations from the target structures after insertion into the templates-were determined and compared with the proposals. The analysis of the results shows limitations of knowledge-based loop prediction. It is demonstrated that the selection of the anchoring groups is the most important step in the whole procedure.