Capability for short-range forecasting of the Kuroshio path variabilities south of Japan is studied by assimilating the absolute sea surface dynamic height data, which are derived from in situ hydrography and TOPEX/Poseidon (T/P) altimetry, into a high-resolution 1-1/2 layer primitive equation model every 15 days for 5 years from 1993 to 1997. The initialization scheme used is the variational method, which contains the nonlinear dynamics of the model as a weak constraint and is suitable for initialization in the western boundary current regions where initialization by simple optimal interpolation is inappropriate. The time series of the analysis field well represents the Kuroshio path variations during the study period. For example, the root-mean-square (RMS) differences from the observations falls within 0.3¿ for the location in latitude of the current axis off Enshu-nada and 11.4 cm for the sea level at Hachijo-jima Island while their observed RMS variabilities are 0.5¿ and 30.7 cm, respectively. By using the resulting fields as the initial conditions, 116 cases of 90-day forecast experiments for the Kuroshio path variations are then performed to assess the statistical properties of our assimilation system. Though significant bias is found in the forecast field southeast of Kyushu, the accuracy of our forecasted field south of Japan is much higher than that of persistence experiments, and the typical evolution of the Kuroshio into a large-amplitude path off Enshu-nada is successfully reproduced after the initialization. In particular, the eastward progression speed and the horizontal scale of the meander are quite similar to those observed. These results and the comparison with the RMS variability of the analysis field enable us to infer that roughly 60-day forecasts of the Kuroshio path variabilities south of Japan are possible. This indicates that our assimilation system is suitable for the operational short-range forecasting of the upper Kuroshio.