Selecting and De-duplicating Factors for a Random Forest Stock Model

Summary

This article outlines an approach to improving a random forest stock-selection model through factor screening and parameter experimentation. It first limits the stock universe by excluding firms under special treatment and delisted stocks. Candidate predictors span valuation, sentiment, technical, and momentum categories. To assess factors, the author proposes examining information coefficient averages and ratios, rank autocorrelation, grouped excess-return patterns, and turnover. A 30-day holding horizon and an absolute information-coefficient mean above 3% with statistical significance are given as baseline selection criteria.

The process also aims to reduce redundancy: factors with similar economic meaning may be combined using historically weighted returns, and highly correlated pairs are tested with Pearson correlation so the less significant factor can be removed. The author says an adjusted model produced better backtest returns and points to original and experimental models and several parameter combinations. However, the document includes no actual result figures, data-period details, validation design, or safeguards against overfitting, so the claimed improvement cannot be independently assessed from the note.

Key ideas

• The proposed stock universe excludes special-treatment and delisted stocks.
• Candidate factors cover valuation, sentiment, technical, and momentum information.
• Factor screening uses information-coefficient statistics, rank persistence, grouped excess-return behavior, and turnover.
• The stated baseline includes a significant absolute information-coefficient mean above 3% over a 30-day holding period.
• The method combines economically similar factors and removes less significant factors from highly correlated pairs.
• The article reports improved backtest results but provides no figures or validation details to assess that claim.

Tags

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