Based on the T213 global ensemble prediction system (EPS), a new ensemble forecast product——extreme forecast index (EFI), which combines the cumulative distribution functions (CDF) derived from climate and the EPS forecast, is designed. EFI is an index which can measure the continuous differences between the climate CDF and the EPS forecast CDF. When the EFI gets close to 1 or -1, extreme weather is likely to occur. Because EFI is calculated by both climate CDF and the EPS forecast CDF, then there is negative deviation in T213 EPS forecast. In order to ensure the result of the EFI is unaffected by the model error, we choose the T213 EPS model data to form the climate CDF. Next take the 2 m temperature for example, we use EFI3 to forecast the extreme low temperature in January 2008. The results show when the threshold of EFI3 that gives extreme low temperature signal is set to -0.3, most areas of extreme low temperature can be predicted on EFI3 map in 3 to 5 days in advance. We should note that the areas of extreme low temperature do not agree very well with where the EFI3 signals the extreme low temperature. It may be there are some deficiencies in the present formula of EFI3 and the EFI3 is less sensitive to extreme value of the EPS forecast. Later, a relative operating characteristic (ROC) curve is used to verify EFI3. In all three leading times shows that the averaged ROC curves valid for 2 m temperature EFI3 during 25—29 January 2008 are above the diagonal, the areas of the averaged ROC of three leading times are 0.680, 0.657, and 0.542, respectively. This means that the EFI3 is skillful in forecasting the extreme low temperature, with the leading time growing, the EFI3 is less skillful. Finally, another model climate cumulative distribution function is used to generate the EFI3 and a comparison of skill derived from two kinds of EFI3 is presented. The results show that the skill of EFI3 generated by the former model climate is slightly higher than that of EFI3 generated by the latter model climate. The reason may be that the initial field of the former model climate data is derived from the same as similation scheme while the initial field of the latter model climate data is derived from two kinds of schemes. Therefore the more consistent the model data are, the more equitable the cumulative probability distribution of model climate is. We need not to collect much model data if they be not consistent.