Comparing XGBoost and Linear Regression for Chinese Micro-Cap Stock Selection

Summary

This assignment describes two Chinese equity ranking approaches focused on micro-cap stocks. Both screen for positive trailing earnings, the smallest market-cap segment, and exclude special-treatment, suspended, and recently listed shares. The XGBoost model combines valuation, short-horizon momentum and reversal, liquidity, volatility, profit-quality, size, price, and RSI features. The author leaves size exposure intentional, standardizes selected inputs and the five-day return label, and contrasts short training windows that adapt quickly with longer windows that are more stable. A linear regression alternative uses a smaller set of monotonicity-oriented factors, rank-transforms them, and selects among three linear models.

The author reports that XGBoost performed better than the linear approach, while noting drawdowns above 30% and substantial risk of overfitting. Backtest comparisons are described, but the configurations and factor sets differ, limiting a clean comparison. The strategy may also have benefited from recent small-cap gains in Beijing-listed shares; the author recommends stronger neutralization and risk controls. The document offers a practitioner’s account, not a controlled or independently validated study.

Key ideas

• The XGBoost screen targets profitable micro-caps and combines valuation, momentum, liquidity, volatility, quality, size, price, and RSI features.
• The author favors shorter training windows for faster adaptation, while acknowledging that they require more frequent retraining.
• The linear model uses factors expected to behave more monotonically and applies cross-sectional ranking to align their scales.
• The reported XGBoost results exceed the linear model’s, but the comparison is limited by differences in factors and backtest settings.
• Large drawdowns, possible overfitting, and exposure to Beijing-listed small caps are identified as risks.

Tags

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