This paper reports the effort of deducing the initial strategy distributions of agents in mix-game model which is used to predict a real financial time series generated from a target financial market. Using mix-game to predict Shanghai Index, we find the time series of prediction accurate rates is sensitive to the initial strategy distributions of agents in group 2 who play minority game, but less sensitive to the initial strategy distributions of agents in group 1 who play majority game. And agents in group 2 tend to cluster in full strategy space (FSS) if the real financial time series has obvious tendency (upward or downward), otherwise they tend to scatter in FSS. We also find that the initial strategy distributions and the number of agents in group 1 influence the level of prediction accurate rates. Finally, this paper gives suggestion about further research.