A company is building a classifier to rate monthly sales performance (1–5) for the past 20 years. The dataset includes month, sales region, regional aggregate sales, and number of stores. Two months each year show consistently high aggregate sales (seasonal spikes). After one-hot encoding categorical features and training, model accuracy on the validation set is worse than expected. What step will most likely improve validation accuracy?

A stratified train/validation split ensures that the seasonal spikes (high sales months) and regional variations are proportionally represented in both the training and validation sets. This prevents the validation set from being unrepresentative of the overall data distribution, leading to a more reliable accuracy metric and better model generalization. Removing outliers might discard valuable information, especially if the "outliers" are genuine seasonal spikes. Normalizing aggregate sales globally or per region might help with model convergence or feature scaling but doesn't address the fundamental issue of an unrepresentative validation set due to skewed data distribution.