One of the major difficulties in predicting the retention of chemically-diverse compounds in gradient elution is that their retention dependence on solvent composition can differ dramatically. For example, the following figure [insert figure here] shows the isocratic retention behavior of three different compounds at a range of solvent compositions. You clearly can't assume that the retention vs. solvent composition relationships all have the same shape - you can't even assume that they're linear.
However, many other approaches at retention prediction like retention indexing and peak alignment schemes like non-linear retention time warping assume that the retention vs. solvent composition relationships all have the same general shape. Because of this very rough assumption, these techniques cannot accurately predict retention when the gradient program or flow rate is changed. In fact, even the relatively small differences in the gradient and flow rate profiles produced by different HPLC pumps are enough to significantly lower retention prediction accuracy with these schemes.
To give an idea of how big these errors can be, consider that the isocratic retention index of aspirin was measured at 10% methanol to be 302. However, when the mobile phase composition changed to 30%, the retention index of aspirin dropped to 8!
We avoid this problem by making no assumption about the retention vs. solvent composition relationship of each solute. We measure the isocratic retention of each solute at a range of solvent compositions and store that information in a retention database that is used to calculate gradient retention times.