Getting the Target Right in Return Prediction
A new study finds that how machine learning models define their prediction target is a key factor in forecasting stock returns. Analyzing 35 global markets over 30 years, researchers show that switching from raw returns to standardized or rank-based targets can nearly triple prediction accuracy and double portfolio gains. Rank-based targets work best but may miss important details in volatile or small markets. The study challenges common focus on feature engineering and calls for flexible approaches.
What it examines
This study examines how the way stock returns are defined as prediction targets affects machine learning performance in forecasting returns. Using data from 35 international markets over 30 years, the authors compare different transformations of stock characteristics and target returns to improve prediction accuracy.
What it concludes
The research finds that transforming target returns, especially using rank-based methods, greatly boosts prediction and portfolio performance. However, these methods may lose important information in certain cases. The findings can help investors and researchers design better models for stock selection and risk management across global markets.
Evidence objects
A global study spanning 35 stock markets and three decades finds that redefining the prediction target in machine learning can nearly triple forecast accuracy and double investment portfolio returns.
key_findings bullet 1 · key_findings · validation V0
Transforming targetsespecially using rank-based returnshas a far greater impact on performance than traditional feature engineering, but may overlook crucial differences in returns during volatile periods or among small companies.
key_findings bullet 2 · key_findings · validation V0
The research challenges conventional wisdom, showing that the best target definition varies by market and time, and raises questions about real-time selection and resilience in extreme market conditions.
key_findings bullet 3 · key_findings · validation V0
This paper uniquely highlights the overlooked impact of prediction target transformationsuch as raw, standardized, or rank-based returnson machine learning accuracy and portfolio performance. Its systematic, cross-market empirical analysis challenges the fields focus on feature engineering, offering compelling evidence that target engineering is crucial and underexplored in quantitative finance and trading.
key_findings bullet 4 · key_findings · validation V0
Raw abstract and provenance
Abstract. We show that a largely overlooked design choice-how returns are defined as prediction targets-drives machine learning performance in stock returns
Source row: 990 · abstract type: snippet