<p>This note offers a reflection on the methodological trade-off between understanding and prediction in quantitative models used in marketing research. This tension is expressed in terms of bias and variance: while greater understanding requires reducing bias, good predictive ability requires minimizing both bias and variance. Model complexity tends to reduce bias but increase variance, resulting in a risk of overfitting. Conversely, simpler models reduce variance but can lead to underfitting. To balance this trade-off, the analyst can use tools such as cross-validation. These topics; overfitting, underfitting and cross-validation; play out differently depending on whether the analyst is using traditional frequentist or Bayesian modeling and so we discuss both approaches. We note that for both the frequentist and Bayesian paradigms, cross-validation mitigates model sensitivity to observed data and also promotes replicable results. Replication, understood as the ability to obtain consistent conclusions on new samples, has emerged as an essential criterion for assessing scientific reliability in both quantitative and qualitative research. We therefore hope that this work can contribute to the transparent and cumulative construction of knowledge in marketing.</p>