Typical neuroimaging studies analyze associations between physiological or behavioral traits and brain structure or function. Some rely on predicting these scores from neuroimaging data. To explain association between brain features and multiple traits, reduced-rank regression (RRR) models are often used, such as canonical correlation analysis (CCA) and partial least squares (PLS). These methods estimate latent variables, or canonical modes, that maximize the covariations between neuroimaging features and behavioral scores. Here, we investigate theoretically and empirically the extent to which reduced-rank models predict out-of-sample clinical scores from functional connectivity. Experiments on a schizophrenia dataset show that i) significant correlations between canonical modes do not necessarily mean accurate generalization on unseen data, and ii) better accuracy is achieved when taking into account regularized covariance between scores.