Training a Stock Model on Recent Limit-Up Equities

Summary

The document explains how to restrict a Chinese stock machine-learning dataset to shares that closed at their daily price limit, or to shares that hit the limit on the previous trading day. It describes filtering factor data by limit-status value and using a lagged status for the prior-day variant. The stated motivation is to reduce noise and computational effort by training only on a selected subset of stocks.

For one stated sample period, the filtered dataset contained just over 15,000 rows. The author recommends adjusting a StockRanker model for the smaller sample, including raising its minimum leaf sample size and reducing the number of trees. In the reported comparison for that period, the limited-universe training result was 17.94%, below the full-market result of 97.95%. The document warns that the reduced sample can make results highly variable; the figures are period-specific and do not establish general performance.

Key ideas

• Filtering on the limit-status field can restrict training examples to stocks that closed at the daily price limit.
• A lagged limit-status condition selects stocks that reached the limit on the previous trading day.
• The author reports just over 15,000 filtered observations in the example period and suggests adapting tree-model settings to the smaller sample.
• In that particular comparison, limit-up-only training performed below training on the full market.
• The smaller dataset may lead to highly variable results, so the reported comparison should not be generalized.

Tags

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