In this study, we introduce a novel variable transformation scheme to improve short-lead forecast skills of statistical models. This transformation is called persistence neutralization transformation because it can effectively eliminate persistence characteristic of a time series by absorbing the damped persistence part of the original series into the time mean state of the transformed series. Therefore, one can expect that any statistical model using such a transformation to preprocess the predictand will always yield forecast skills comparable to or better than those of persistent forecasts. Results from retroactive forecast experiments of global tropical sea surface temperature between 1980 and 2008 clearly showed that the persistence neutralization transformation indeed provides an effective way to improve short-lead forecast skills. Furthermore, this transformation allows that lead-lag relations among variables can be more faithfully reflected in forecast capability of statistical models without the distortion by persistence characteristics of individual time series. Therefore, it can also be used to search for predictors with better forecast capability for each lead time to further improve statistical models' forecast skills.