Investigating Forward-Return Horizons in Rank IC Evaluation
Summary
The document raises a measurement question about Rank IC, a rank-based measure of association between model predictions and subsequent returns. The author reports that Rank IC computed on a custom training set differs substantially from the value produced by a platform evaluation function. Using cumulative returns over the following twenty days gives a result closer to the platform's figure, leading to a question about which return horizon the evaluation uses.
This is a troubleshooting observation, not a confirmed explanation or a complete method. The document does not establish that the platform evaluates against a twenty-day return, nor does it resolve whether the discrepancy comes from label definitions, date and asset alignment, or other evaluation settings. It suggests checking the target return horizon used for training against the horizon used for evaluation, while validating data alignment and metric conventions before drawing conclusions. No broader empirical evidence or performance results are provided.
Key ideas
- Rank IC results can differ between a custom dataset calculation and a platform evaluation.
- A twenty-day cumulative forward return produced a closer match in the reported case.
- The observation does not establish which return horizon the platform actually uses.
- Researchers should check target definitions, evaluation settings, and data alignment.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.