The skill of global-scale sea surface temperature forecasts using a statist
ically based linear forecasting technique is investigated. Canonical variat
es are used to make monthly sea surface temperature anomaly forecasts using
evolutionary and steady-state features of antecedent sea surface temperatu
res as predictors. Levels of forecast skill are investigated over several m
onths' lead time by comparing the model performance with a simple forecast
strategy involving the persistence of sea surface temperature anomalies. Fo
recast skill is investigated over an independent test period of 18 yr (1982
/83-1999/2000), for which the model training period was updated after every
3 yr. Forecasts for the equatorial Pacific Ocean are a significant improve
ment over a strategy of random guessing, and outscore forecasts of persiste
d anomalies beyond lead times of about one season during the development st
ages of the El Nino-Southern Oscillation phenomenon, but only outscore fore
casts of persisted anomalies beyond 6 months' lead time during its most int
ense phase. Model predictions of the tropical Indian Ocean outscore persist
ence during the second half of the boreal winter, that is, from about Decem
ber or January, with maximum skill during the March-May spring season, but
poor skill during the autumn months from September to November. Some loss i
n predictability of the equatorial Pacific and Indian Oceans is evident dur
ing the early and mid-1990s, but forecasts appear to have improved in the l
ast few years. The tropical Atlantic Ocean forecast skill has generally bee
n poor. There is little evidence of forecast skill over the midlatitudes in
any of the oceans. However, during the spring months significant skill has
been found over the Indian Ocean as far south as 20 degreesS and over the
southern North Atlantic as far north as 30 degreesN, both of which outscore
persistence beyond a lead time of less than about one season.