Our replicated real-time forecast scheme uses an initial model training period from 1950 (or later) through to 1987, leaving the 15 years 1988-2002 for model forecast assessment [eg Saunders and Qian, 2002]. The training period increases one year at a time as each forecast is made. This updating exactly replicates the operation of a real-time forecast. The prediction skill is computed by comparing the 15 independent forecasts with the 15 observed values. While this skill is the true forecast skill for the 15-year period 1988-2002, it may not be representative of the true skill over longer periods.
Saunders, M. A. and B. Qian, Seasonal Predictability of the winter NAO from North Atlantic sea surface temperatures, Geophys. Res. Lett., 29(22), 2049, doi:10.1029/2002GL014952, 2002.