A multivariate model testing procedure is used to investigate the thermocline depth anomalies in the tropical Pacific during 1985--1992 simulated by a two-layer model forced by five different surface wind stress fields. Two of these are derived from the pseudo wind stress produced from surface marine reports at the Florida State University (FSU), using a constant and a stability dependent drag coefficient. The others come from the response of three atmospheric general circulation models (ARPEGE, ECHAM3, and Hadley Centre) to the observed sea surface temperature. The simulations are compared to a new bimonthly analysis of the 20 ¿C isotherm depth anomalies based on ship and moored buoy measurements. The best model-data fit is gained when the forcing uncertainties are taken into account, considering the multiple simulations as representative of the latter. But they are still insufficient to explain the model-reality discrepancies. However, although the two-layer model has serious limitations in the eastern Pacific where the simulated oceanic response is minimal while the observed variability is large, the testing procedure shows that it reproduces the main features of the observed variability and may thus be used to understand how, in a simulation, a particular feature of a forcing field translates. It is found that the simulations based on the FSU product and the Hadley Centre wind stresses are comparable in realism, and are significantly more realistic than the ones based on the ARPEGE or ECHAM3 wind stresses, because these misrepresent the phase of the 1986--1987 El Ni¿o event. In addition, the ECHAM3 wind stress peaks off the equator, decreasing the quality of the two-layer model response. ¿ 1998 American Geophysical Union