Using a recently developed stability estimator, stability is leveraged at the cost of discriminatory power, in order to improve feature selection for small 2 x 2 metabolomics crossover designs. This is done using Pareto Front Cross-Validation (PFCV) adapted with an automated hyperparameter selection criteria. PFCV is evaluated for Partial Least Squares Discriminant Analysis’s (PLSDA) Variable Importance Projections, Significant Multivariate Correlations, Nearest Shrunken Centroids and the Soft-Threshold PLSDA using a simulation study and real metabolomics data. Variable importance projections with PFCV provided the best overall feature selection and is recommended for subject sizes as small as 6. However, for larger subject sizes, this recommendation was shown to potentially vary depending on the goals of the practitioner. Overall, the use of PFCV for model selection in small 2 x 2 metabolomics crossover designs is advocated in future research.