Abstract: The performance of the operational global model system (GRAPES_GFS) for day 1 (D1) to day 5 (D5) precipitation forecasts is evaluated in terms of precipitation amount, frequency and diurnal cycle etc. against observational product of the Global Precipitation Measurement (GPM) data over four months (Jan, April, Jul and Oct) of 2017. Special attention is focused on areas of the Warm Pool (WP) in the Western Pacific and the Storm Track in the Northern Hemisphere (STNH). Results show that: (1) the D1—D5 forecasts reasonably capture global distribution pattern of precipitation. In particular, the model accurately reproduces the typical observed maximum feature of global zonal mean precipitation (amount and frequency) between 20°S—20°N and in latitudes of 40°—50°, and the simulated daily variation and diurnal cycle of precipitation well resemble observations over the WP and the STNH; (2) in the low-latitudes, the "double peaks" of positive forecast bias of daily precipitation amount and frequency of heavy rainfall days (＞25 mm/d) occur in the same locations of observed precipitation maximums, while the positive forecast deviation of wet-day (≥0.1 mm/d) frequency is quite small. In the mid-latitudes, the forecast of daily precipitation amount is nearly unbiased, but positive deviation of wet-day frequency and negative bias of heavy rainfall day frequency are obvious in latitudes of 40°—60°. There is almost no change in the bias distribution pattern, and the deviation values vary a little in different seasons and increase from D1 to D5, while RMSE is several times of AME. The above results indicate that the model may have systematic errors in precipitation forecast and the model performance in daily forecast varies widely; (3) in terms of diurnal cycle, the positive bias of prediction amount is due to over-prediction in precipitation intensity, and the slight negative deviation of precipitation frequency is attributed to under-prediction of the rainfall coverage in the WP. However, the positive bias of precipitation frequency is resulted from both factors in STNH, the over-prediction of rainfall coverage and the failed forecast of weak precipitation events; (4) the obvious differences in precipitation (amount and frequency) deviation between low- and mid-latitudes are related to the incongruity of the proportion of grid and sub-grid scale precipitation in the model. The clues of model improvement point to the trigger function and process of deep convection precipitation in model cumulus parameterization scheme, and the coordination between the cumulus parameterization scheme and the cloud microphysics scheme.