This study develops two new statistical methods to estimate the potential seasonal predictability of global surface temperature. The first one is based on the concept of Analysis of Covariance (ANOCOVA). It has the advantage of not only taking into account autocorrelation structure in the daily time series but also accounting for the uncertainty of the estimated parameters in the significance test. This method tests whether interannual variability of seasonal means exceeds that due to weather noise under the null hypothesis that seasonal means are identical every year. The second method is based on the bootstrap technique that makes few assumptions about physical process, model structure and underlying distribution. The essence of the bootstrap is to randomly resample the daily time series to build up an empirical distribution of the variance of seasonal means under the null hypothesis that seasonal mean is independent of year. This study then applies these two new methods to global surface temperature data from the NCEP/NCAR reanalysis to calculate potential seasonal predictability. These results are then compared with estimates from methods previously proposed by Shukla-Gutzler (SG) and Madden (MN). The results from different methods consistently show a high fraction of predictable variance in the tropics and extratropical oceans, low predictability over the extratropical land regions, more potential predictability over the ocean than land, and a stronger seasonal variation in potential predictability over land than ocean. The MN method tends to identify less potential predictability than the other methods, especially over the extratropical land areas. The SG and ANOCOVA methods tend to yield similar patterns of potential predictability. The bootstrap tends to detect more potential predictability than the other methods, with more than 91% of the globe identified as potentially predictable.