Constructing Forward Return Labels with Quantile Clipping for Equity Models

Summary

The document describes preparing supervised-learning labels from Chinese daily stock bars. It calculates a forward price ratio using lead values, clips observations to the cross-sectional 1st and 99th percentile boundaries, and assigns the result as the target label. It also filters out missing values and observations where the next day's high equals its low, which the notes describe as excluding limit-locked cases. Although the surrounding explanation mentions dividing returns into 20 buckets, the displayed query assigns the clipped continuous return directly as the label; the binning expression is commented out.

This is a labeling recipe, not a predictive model or trading strategy, and it reports no model performance. The stated horizon description and the visible lead offsets do not clearly align, so the intended holding period should be checked before use. Quantile clipping can reduce the influence of extreme outcomes, but the snippet does not clarify whether percentile thresholds are computed within each date or across the full sample, a distinction that matters for leakage and label consistency.

Key ideas

• The query derives a forward target from future price fields in daily stock data.\nIt clips target values at the 1st and 99th percentile thresholds.\nRows with missing fields and a flat next-day high-low range are excluded.\nThe prose describes 20 return buckets, but the shown target remains continuous because binning is commented out.\nThe horizon and percentile calculation scope need clarification before model training.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.