We used the approach of conditional nonlinear optimal perturbation (CNOP) to investigate optimal precursors for El Nino--Southern Oscillation (ENSO) events with a theoretical coupled ocean-atmosphere model. The CNOPs of the annual cycle of the coupled system were computed for different time periods, and the derived CNOPs were compared with the linear singular vectors (LSVs). The results show the existence of the CNOPs of annual cycle and local CNOPs. These CNOPs have the robust optimal patterns, which have opposite polarities in sea surface temperature and thermocline depth anomalies in the eastern equatorial Pacific. We demonstrate that the CNOP (local CNOP), rather than LSVs, has the highest likelihood to develop into an El Nino (La Nina) event; thus the CNOPs (local CNOPs) can be regarded as the optimal precursors for El Nino (La Nina) events. These optimal precursors agrees qualitatively well with the observations of period of 1980--2002. On the basis of the nonlinear oscillation described by the model, the physical mechanism of the optimal precursors for ENSO is discussed.