Winner Stock Prediction as Decision-Aligned Multiclass Classification

We study short-horizon equity selection through a decision-aligned winner prediction task: given recent market history for a fixed universe, predict which asset will achieve the highest next-day return. This reframes return forecasting as a multiclass classification problem and provides a natural analogue to next-token prediction in sequence modeling. Using a decade of daily data for seven large-cap NASDAQ equities, we compare forecast-then-argmax pipelines with direct discriminative models, including classical classifiers and Transformer encoders. We find that models achieving the highest top-1 accuracy often do so by collapsing predictions onto a small subset of dominant assets, while sequence models trade accuracy for substantially broader class coverage. These results expose an accuracy–concentration tradeoff in relative-return forecasting and highlight the need for class-sensitive and diversity-aware evaluation beyond headline accuracy in Financial AI benchmarks.