Canonical correlation analysis (CCA) is a linear statistical technique that identifies coupled patterns between predictor and predictand fields. CCA is a standard tool in seasonal climate forecasting and it is widely used to generate operational forecasts of temperature, precipitation and sea surface temperature. In this study, CCA is used to estimate the levels and sources of seasonal forecast skill for September sea-ice concentration over the 1980–2006 time period. Three years, 2007, 2008 and 2009, are held in abeyance for an informal independent verification of the empirical models. The potential to generate forecasts at lead times between 1 month and 1 year is explored by comparing the predictive skill of various climate variables to persistence and climatology. The pool of climate variables tested as potential predictors include sea surface temperature, surface air temperature, sea-level pressure, 500 mb geopotential height, surface winds and multi-year sea ice. Models with the highest forecast skill will be used to generate a forecast for the 2010 September minimum.