A better prediction target did not rescue our model.

Public research brief

Before replacing the model, consider what you are asking it to predict.

Cakici and Zaremba report that transforming return targets can materially improve machine-learning portfolio results across an international equity panel. Ranks often help, but they also discard information about the size of returns. The paper's abstract emphasizes that the best transformation depends on the market and return distribution.

We tested the rank-target idea in an existing earnings research model. The attraction was a clean comparison: change the target encoding while preserving the model and evaluation rules. The result was a useful negative.

AI assists curation and drafting. The research status distinguishes paper reviews from house tests. Read our research approach.